Print this page
9059 Simplify SMAP relocations with krtld
Portions contributed by: John Levon <john.levon@joyent.com>
| Split |
Close |
| Expand all |
| Collapse all |
--- old/usr/src/uts/i86pc/os/startup.c
+++ new/usr/src/uts/i86pc/os/startup.c
1 1 /*
2 2 * CDDL HEADER START
3 3 *
4 4 * The contents of this file are subject to the terms of the
5 5 * Common Development and Distribution License (the "License").
6 6 * You may not use this file except in compliance with the License.
7 7 *
8 8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9 9 * or http://www.opensolaris.org/os/licensing.
10 10 * See the License for the specific language governing permissions
11 11 * and limitations under the License.
12 12 *
13 13 * When distributing Covered Code, include this CDDL HEADER in each
14 14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15 15 * If applicable, add the following below this CDDL HEADER, with the
16 16 * fields enclosed by brackets "[]" replaced with your own identifying
17 17 * information: Portions Copyright [yyyy] [name of copyright owner]
18 18 *
19 19 * CDDL HEADER END
20 20 */
21 21
22 22 /*
23 23 * Copyright (c) 1993, 2010, Oracle and/or its affiliates. All rights reserved.
24 24 * Copyright 2012 DEY Storage Systems, Inc. All rights reserved.
25 25 * Copyright 2017 Nexenta Systems, Inc.
26 26 * Copyright (c) 2018 Joyent, Inc.
27 27 * Copyright (c) 2015 by Delphix. All rights reserved.
28 28 */
29 29 /*
30 30 * Copyright (c) 2010, Intel Corporation.
31 31 * All rights reserved.
32 32 */
33 33
34 34 #include <sys/types.h>
35 35 #include <sys/t_lock.h>
36 36 #include <sys/param.h>
37 37 #include <sys/sysmacros.h>
38 38 #include <sys/signal.h>
39 39 #include <sys/systm.h>
40 40 #include <sys/user.h>
41 41 #include <sys/mman.h>
42 42 #include <sys/vm.h>
43 43 #include <sys/conf.h>
44 44 #include <sys/avintr.h>
45 45 #include <sys/autoconf.h>
46 46 #include <sys/disp.h>
47 47 #include <sys/class.h>
48 48 #include <sys/bitmap.h>
49 49
50 50 #include <sys/privregs.h>
51 51
52 52 #include <sys/proc.h>
53 53 #include <sys/buf.h>
54 54 #include <sys/kmem.h>
55 55 #include <sys/mem.h>
56 56 #include <sys/kstat.h>
57 57
58 58 #include <sys/reboot.h>
59 59
60 60 #include <sys/cred.h>
61 61 #include <sys/vnode.h>
62 62 #include <sys/file.h>
63 63
64 64 #include <sys/procfs.h>
65 65
66 66 #include <sys/vfs.h>
67 67 #include <sys/cmn_err.h>
68 68 #include <sys/utsname.h>
69 69 #include <sys/debug.h>
70 70 #include <sys/kdi.h>
71 71
72 72 #include <sys/dumphdr.h>
73 73 #include <sys/bootconf.h>
74 74 #include <sys/memlist_plat.h>
75 75 #include <sys/varargs.h>
76 76 #include <sys/promif.h>
77 77 #include <sys/modctl.h>
78 78
79 79 #include <sys/sunddi.h>
80 80 #include <sys/sunndi.h>
81 81 #include <sys/ndi_impldefs.h>
82 82 #include <sys/ddidmareq.h>
83 83 #include <sys/psw.h>
84 84 #include <sys/regset.h>
85 85 #include <sys/clock.h>
86 86 #include <sys/pte.h>
87 87 #include <sys/tss.h>
88 88 #include <sys/stack.h>
89 89 #include <sys/trap.h>
90 90 #include <sys/fp.h>
91 91 #include <vm/kboot_mmu.h>
92 92 #include <vm/anon.h>
93 93 #include <vm/as.h>
94 94 #include <vm/page.h>
95 95 #include <vm/seg.h>
96 96 #include <vm/seg_dev.h>
97 97 #include <vm/seg_kmem.h>
98 98 #include <vm/seg_kpm.h>
99 99 #include <vm/seg_map.h>
100 100 #include <vm/seg_vn.h>
101 101 #include <vm/seg_kp.h>
102 102 #include <sys/memnode.h>
103 103 #include <vm/vm_dep.h>
104 104 #include <sys/thread.h>
105 105 #include <sys/sysconf.h>
106 106 #include <sys/vm_machparam.h>
107 107 #include <sys/archsystm.h>
108 108 #include <sys/machsystm.h>
109 109 #include <vm/hat.h>
110 110 #include <vm/hat_i86.h>
111 111 #include <sys/pmem.h>
112 112 #include <sys/smp_impldefs.h>
113 113 #include <sys/x86_archext.h>
114 114 #include <sys/cpuvar.h>
115 115 #include <sys/segments.h>
116 116 #include <sys/clconf.h>
117 117 #include <sys/kobj.h>
118 118 #include <sys/kobj_lex.h>
119 119 #include <sys/cpc_impl.h>
120 120 #include <sys/cpu_module.h>
121 121 #include <sys/smbios.h>
122 122 #include <sys/debug_info.h>
123 123 #include <sys/bootinfo.h>
124 124 #include <sys/ddi_periodic.h>
125 125 #include <sys/systeminfo.h>
126 126 #include <sys/multiboot.h>
127 127 #include <sys/ramdisk.h>
128 128
129 129 #ifdef __xpv
130 130
131 131 #include <sys/hypervisor.h>
132 132 #include <sys/xen_mmu.h>
133 133 #include <sys/evtchn_impl.h>
134 134 #include <sys/gnttab.h>
135 135 #include <sys/xpv_panic.h>
136 136 #include <xen/sys/xenbus_comms.h>
137 137 #include <xen/public/physdev.h>
138 138
139 139 extern void xen_late_startup(void);
140 140
141 141 struct xen_evt_data cpu0_evt_data;
142 142
143 143 #else /* __xpv */
144 144 #include <sys/memlist_impl.h>
145 145
146 146 extern void mem_config_init(void);
147 147 #endif /* __xpv */
148 148
149 149 extern void progressbar_init(void);
150 150 extern void brand_init(void);
151 151 extern void pcf_init(void);
152 152 extern void pg_init(void);
153 153 extern void ssp_init(void);
154 154
155 155 extern int size_pse_array(pgcnt_t, int);
156 156
157 157 #if defined(_SOFT_HOSTID)
158 158
159 159 #include <sys/rtc.h>
160 160
161 161 static int32_t set_soft_hostid(void);
162 162 static char hostid_file[] = "/etc/hostid";
163 163
164 164 #endif
165 165
166 166 void *gfx_devinfo_list;
167 167
168 168 #if defined(__amd64) && !defined(__xpv)
169 169 extern void immu_startup(void);
170 170 #endif
171 171
172 172 /*
173 173 * XXX make declaration below "static" when drivers no longer use this
174 174 * interface.
175 175 */
176 176 extern caddr_t p0_va; /* Virtual address for accessing physical page 0 */
177 177
178 178 /*
179 179 * segkp
180 180 */
181 181 extern int segkp_fromheap;
182 182
183 183 static void kvm_init(void);
184 184 static void startup_init(void);
185 185 static void startup_memlist(void);
186 186 static void startup_kmem(void);
187 187 static void startup_modules(void);
188 188 static void startup_vm(void);
189 189 static void startup_end(void);
190 190 static void layout_kernel_va(void);
191 191
192 192 /*
193 193 * Declare these as initialized data so we can patch them.
194 194 */
195 195 #ifdef __i386
196 196
197 197 /*
198 198 * Due to virtual address space limitations running in 32 bit mode, restrict
199 199 * the amount of physical memory configured to a max of PHYSMEM pages (16g).
200 200 *
201 201 * If the physical max memory size of 64g were allowed to be configured, the
202 202 * size of user virtual address space will be less than 1g. A limited user
203 203 * address space greatly reduces the range of applications that can run.
204 204 *
205 205 * If more physical memory than PHYSMEM is required, users should preferably
206 206 * run in 64 bit mode which has far looser virtual address space limitations.
207 207 *
208 208 * If 64 bit mode is not available (as in IA32) and/or more physical memory
209 209 * than PHYSMEM is required in 32 bit mode, physmem can be set to the desired
210 210 * value or to 0 (to configure all available memory) via eeprom(1M). kernelbase
211 211 * should also be carefully tuned to balance out the need of the user
212 212 * application while minimizing the risk of kernel heap exhaustion due to
213 213 * kernelbase being set too high.
214 214 */
215 215 #define PHYSMEM 0x400000
216 216
217 217 #else /* __amd64 */
218 218
219 219 /*
220 220 * For now we can handle memory with physical addresses up to about
221 221 * 64 Terabytes. This keeps the kernel above the VA hole, leaving roughly
222 222 * half the VA space for seg_kpm. When systems get bigger than 64TB this
223 223 * code will need revisiting. There is an implicit assumption that there
224 224 * are no *huge* holes in the physical address space too.
225 225 */
226 226 #define TERABYTE (1ul << 40)
227 227 #define PHYSMEM_MAX64 mmu_btop(64 * TERABYTE)
228 228 #define PHYSMEM PHYSMEM_MAX64
229 229 #define AMD64_VA_HOLE_END 0xFFFF800000000000ul
230 230
231 231 #endif /* __amd64 */
232 232
233 233 pgcnt_t physmem = PHYSMEM;
234 234 pgcnt_t obp_pages; /* Memory used by PROM for its text and data */
235 235
236 236 char *kobj_file_buf;
237 237 int kobj_file_bufsize; /* set in /etc/system */
238 238
239 239 /* Global variables for MP support. Used in mp_startup */
240 240 caddr_t rm_platter_va = 0;
241 241 uint32_t rm_platter_pa;
242 242
243 243 int auto_lpg_disable = 1;
244 244
245 245 /*
246 246 * Some CPUs have holes in the middle of the 64-bit virtual address range.
247 247 */
248 248 uintptr_t hole_start, hole_end;
249 249
250 250 /*
251 251 * kpm mapping window
252 252 */
253 253 caddr_t kpm_vbase;
254 254 size_t kpm_size;
255 255 static int kpm_desired;
256 256 #ifdef __amd64
257 257 static uintptr_t segkpm_base = (uintptr_t)SEGKPM_BASE;
258 258 #endif
259 259
260 260 /*
261 261 * Configuration parameters set at boot time.
262 262 */
263 263
264 264 caddr_t econtig; /* end of first block of contiguous kernel */
265 265
266 266 struct bootops *bootops = 0; /* passed in from boot */
267 267 struct bootops **bootopsp;
268 268 struct boot_syscalls *sysp; /* passed in from boot */
269 269
270 270 char bootblock_fstype[16];
271 271
272 272 char kern_bootargs[OBP_MAXPATHLEN];
273 273 char kern_bootfile[OBP_MAXPATHLEN];
274 274
275 275 /*
276 276 * ZFS zio segment. This allows us to exclude large portions of ZFS data that
277 277 * gets cached in kmem caches on the heap. If this is set to zero, we allocate
278 278 * zio buffers from their own segment, otherwise they are allocated from the
279 279 * heap. The optimization of allocating zio buffers from their own segment is
280 280 * only valid on 64-bit kernels.
281 281 */
282 282 #if defined(__amd64)
283 283 int segzio_fromheap = 0;
284 284 #else
285 285 int segzio_fromheap = 1;
286 286 #endif
287 287
288 288 /*
289 289 * Give folks an escape hatch for disabling SMAP via kmdb. Doesn't work
290 290 * post-boot.
291 291 */
292 292 int disable_smap = 0;
293 293
294 294 /*
295 295 * new memory fragmentations are possible in startup() due to BOP_ALLOCs. this
296 296 * depends on number of BOP_ALLOC calls made and requested size, memory size
297 297 * combination and whether boot.bin memory needs to be freed.
298 298 */
299 299 #define POSS_NEW_FRAGMENTS 12
300 300
301 301 /*
302 302 * VM data structures
303 303 */
304 304 long page_hashsz; /* Size of page hash table (power of two) */
305 305 unsigned int page_hashsz_shift; /* log2(page_hashsz) */
306 306 struct page *pp_base; /* Base of initial system page struct array */
307 307 struct page **page_hash; /* Page hash table */
308 308 pad_mutex_t *pse_mutex; /* Locks protecting pp->p_selock */
309 309 size_t pse_table_size; /* Number of mutexes in pse_mutex[] */
310 310 int pse_shift; /* log2(pse_table_size) */
311 311 struct seg ktextseg; /* Segment used for kernel executable image */
312 312 struct seg kvalloc; /* Segment used for "valloc" mapping */
313 313 struct seg kpseg; /* Segment used for pageable kernel virt mem */
314 314 struct seg kmapseg; /* Segment used for generic kernel mappings */
315 315 struct seg kdebugseg; /* Segment used for the kernel debugger */
316 316
317 317 struct seg *segkmap = &kmapseg; /* Kernel generic mapping segment */
318 318 static struct seg *segmap = &kmapseg; /* easier to use name for in here */
319 319
320 320 struct seg *segkp = &kpseg; /* Pageable kernel virtual memory segment */
321 321
322 322 #if defined(__amd64)
323 323 struct seg kvseg_core; /* Segment used for the core heap */
324 324 struct seg kpmseg; /* Segment used for physical mapping */
325 325 struct seg *segkpm = &kpmseg; /* 64bit kernel physical mapping segment */
326 326 #else
327 327 struct seg *segkpm = NULL; /* Unused on IA32 */
328 328 #endif
329 329
330 330 caddr_t segkp_base; /* Base address of segkp */
331 331 caddr_t segzio_base; /* Base address of segzio */
332 332 #if defined(__amd64)
333 333 pgcnt_t segkpsize = btop(SEGKPDEFSIZE); /* size of segkp segment in pages */
334 334 #else
335 335 pgcnt_t segkpsize = 0;
336 336 #endif
337 337 pgcnt_t segziosize = 0; /* size of zio segment in pages */
338 338
339 339 /*
340 340 * A static DR page_t VA map is reserved that can map the page structures
341 341 * for a domain's entire RA space. The pages that back this space are
342 342 * dynamically allocated and need not be physically contiguous. The DR
343 343 * map size is derived from KPM size.
344 344 * This mechanism isn't used by x86 yet, so just stubs here.
345 345 */
346 346 int ppvm_enable = 0; /* Static virtual map for page structs */
347 347 page_t *ppvm_base = NULL; /* Base of page struct map */
348 348 pgcnt_t ppvm_size = 0; /* Size of page struct map */
349 349
350 350 /*
351 351 * VA range available to the debugger
352 352 */
353 353 const caddr_t kdi_segdebugbase = (const caddr_t)SEGDEBUGBASE;
354 354 const size_t kdi_segdebugsize = SEGDEBUGSIZE;
355 355
356 356 struct memseg *memseg_base;
357 357 struct vnode unused_pages_vp;
358 358
359 359 #define FOURGB 0x100000000LL
360 360
361 361 struct memlist *memlist;
362 362
363 363 caddr_t s_text; /* start of kernel text segment */
364 364 caddr_t e_text; /* end of kernel text segment */
365 365 caddr_t s_data; /* start of kernel data segment */
366 366 caddr_t e_data; /* end of kernel data segment */
367 367 caddr_t modtext; /* start of loadable module text reserved */
368 368 caddr_t e_modtext; /* end of loadable module text reserved */
369 369 caddr_t moddata; /* start of loadable module data reserved */
370 370 caddr_t e_moddata; /* end of loadable module data reserved */
371 371
372 372 struct memlist *phys_install; /* Total installed physical memory */
373 373 struct memlist *phys_avail; /* Total available physical memory */
374 374 struct memlist *bios_rsvd; /* Bios reserved memory */
375 375
376 376 /*
377 377 * kphysm_init returns the number of pages that were processed
378 378 */
379 379 static pgcnt_t kphysm_init(page_t *, pgcnt_t);
380 380
381 381 #define IO_PROP_SIZE 64 /* device property size */
382 382
383 383 /*
384 384 * a couple useful roundup macros
385 385 */
386 386 #define ROUND_UP_PAGE(x) \
387 387 ((uintptr_t)P2ROUNDUP((uintptr_t)(x), (uintptr_t)MMU_PAGESIZE))
388 388 #define ROUND_UP_LPAGE(x) \
389 389 ((uintptr_t)P2ROUNDUP((uintptr_t)(x), mmu.level_size[1]))
390 390 #define ROUND_UP_4MEG(x) \
391 391 ((uintptr_t)P2ROUNDUP((uintptr_t)(x), (uintptr_t)FOUR_MEG))
392 392 #define ROUND_UP_TOPLEVEL(x) \
393 393 ((uintptr_t)P2ROUNDUP((uintptr_t)(x), mmu.level_size[mmu.max_level]))
394 394
395 395 /*
396 396 * 32-bit Kernel's Virtual memory layout.
397 397 * +-----------------------+
398 398 * | |
399 399 * 0xFFC00000 -|-----------------------|- ARGSBASE
400 400 * | debugger |
401 401 * 0xFF800000 -|-----------------------|- SEGDEBUGBASE
402 402 * | Kernel Data |
403 403 * 0xFEC00000 -|-----------------------|
404 404 * | Kernel Text |
405 405 * 0xFE800000 -|-----------------------|- KERNEL_TEXT (0xFB400000 on Xen)
406 406 * |--- GDT ---|- GDT page (GDT_VA)
407 407 * |--- debug info ---|- debug info (DEBUG_INFO_VA)
408 408 * | |
409 409 * | page_t structures |
410 410 * | memsegs, memlists, |
411 411 * | page hash, etc. |
412 412 * --- -|-----------------------|- ekernelheap, valloc_base (floating)
413 413 * | | (segkp is just an arena in the heap)
414 414 * | |
415 415 * | kvseg |
416 416 * | |
417 417 * | |
418 418 * --- -|-----------------------|- kernelheap (floating)
419 419 * | Segkmap |
420 420 * 0xC3002000 -|-----------------------|- segmap_start (floating)
421 421 * | Red Zone |
422 422 * 0xC3000000 -|-----------------------|- kernelbase / userlimit (floating)
423 423 * | | ||
424 424 * | Shared objects | \/
425 425 * | |
426 426 * : :
427 427 * | user data |
428 428 * |-----------------------|
429 429 * | user text |
430 430 * 0x08048000 -|-----------------------|
431 431 * | user stack |
432 432 * : :
433 433 * | invalid |
434 434 * 0x00000000 +-----------------------+
435 435 *
436 436 *
437 437 * 64-bit Kernel's Virtual memory layout. (assuming 64 bit app)
438 438 * +-----------------------+
439 439 * | |
440 440 * 0xFFFFFFFF.FFC00000 |-----------------------|- ARGSBASE
441 441 * | debugger (?) |
442 442 * 0xFFFFFFFF.FF800000 |-----------------------|- SEGDEBUGBASE
443 443 * | unused |
444 444 * +-----------------------+
445 445 * | Kernel Data |
446 446 * 0xFFFFFFFF.FBC00000 |-----------------------|
447 447 * | Kernel Text |
448 448 * 0xFFFFFFFF.FB800000 |-----------------------|- KERNEL_TEXT
449 449 * |--- debug info ---|- debug info (DEBUG_INFO_VA)
450 450 * |--- GDT ---|- GDT page (GDT_VA)
451 451 * |--- IDT ---|- IDT page (IDT_VA)
452 452 * |--- LDT ---|- LDT pages (LDT_VA)
453 453 * | |
454 454 * | Core heap | (used for loadable modules)
455 455 * 0xFFFFFFFF.C0000000 |-----------------------|- core_base / ekernelheap
456 456 * | Kernel |
457 457 * | heap |
458 458 * 0xFFFFFXXX.XXX00000 |-----------------------|- kernelheap (floating)
459 459 * | segmap |
460 460 * 0xFFFFFXXX.XXX00000 |-----------------------|- segmap_start (floating)
461 461 * | device mappings |
462 462 * 0xFFFFFXXX.XXX00000 |-----------------------|- toxic_addr (floating)
463 463 * | segzio |
464 464 * 0xFFFFFXXX.XXX00000 |-----------------------|- segzio_base (floating)
465 465 * | segkp |
466 466 * --- |-----------------------|- segkp_base (floating)
467 467 * | page_t structures | valloc_base + valloc_sz
468 468 * | memsegs, memlists, |
469 469 * | page hash, etc. |
470 470 * 0xFFFFFF00.00000000 |-----------------------|- valloc_base (lower if >256GB)
471 471 * | segkpm |
472 472 * 0xFFFFFE00.00000000 |-----------------------|
473 473 * | Red Zone |
474 474 * 0xFFFFFD80.00000000 |-----------------------|- KERNELBASE (lower if >256GB)
475 475 * | User stack |- User space memory
476 476 * | |
477 477 * | shared objects, etc | (grows downwards)
478 478 * : :
479 479 * | |
480 480 * 0xFFFF8000.00000000 |-----------------------|
481 481 * | |
482 482 * | VA Hole / unused |
483 483 * | |
484 484 * 0x00008000.00000000 |-----------------------|
485 485 * | |
486 486 * | |
487 487 * : :
488 488 * | user heap | (grows upwards)
489 489 * | |
490 490 * | user data |
491 491 * |-----------------------|
492 492 * | user text |
493 493 * 0x00000000.04000000 |-----------------------|
494 494 * | invalid |
495 495 * 0x00000000.00000000 +-----------------------+
496 496 *
497 497 * A 32 bit app on the 64 bit kernel sees the same layout as on the 32 bit
498 498 * kernel, except that userlimit is raised to 0xfe000000
499 499 *
500 500 * Floating values:
501 501 *
502 502 * valloc_base: start of the kernel's memory management/tracking data
503 503 * structures. This region contains page_t structures for
504 504 * physical memory, memsegs, memlists, and the page hash.
505 505 *
506 506 * core_base: start of the kernel's "core" heap area on 64-bit systems.
507 507 * This area is intended to be used for global data as well as for module
508 508 * text/data that does not fit into the nucleus pages. The core heap is
509 509 * restricted to a 2GB range, allowing every address within it to be
510 510 * accessed using rip-relative addressing
511 511 *
512 512 * ekernelheap: end of kernelheap and start of segmap.
513 513 *
514 514 * kernelheap: start of kernel heap. On 32-bit systems, this starts right
515 515 * above a red zone that separates the user's address space from the
516 516 * kernel's. On 64-bit systems, it sits above segkp and segkpm.
517 517 *
518 518 * segmap_start: start of segmap. The length of segmap can be modified
519 519 * through eeprom. The default length is 16MB on 32-bit systems and 64MB
520 520 * on 64-bit systems.
521 521 *
522 522 * kernelbase: On a 32-bit kernel the default value of 0xd4000000 will be
523 523 * decreased by 2X the size required for page_t. This allows the kernel
524 524 * heap to grow in size with physical memory. With sizeof(page_t) == 80
525 525 * bytes, the following shows the values of kernelbase and kernel heap
526 526 * sizes for different memory configurations (assuming default segmap and
527 527 * segkp sizes).
528 528 *
529 529 * mem size for kernelbase kernel heap
530 530 * size page_t's size
531 531 * ---- --------- ---------- -----------
532 532 * 1gb 0x01400000 0xd1800000 684MB
533 533 * 2gb 0x02800000 0xcf000000 704MB
534 534 * 4gb 0x05000000 0xca000000 744MB
535 535 * 6gb 0x07800000 0xc5000000 784MB
536 536 * 8gb 0x0a000000 0xc0000000 824MB
537 537 * 16gb 0x14000000 0xac000000 984MB
538 538 * 32gb 0x28000000 0x84000000 1304MB
539 539 * 64gb 0x50000000 0x34000000 1944MB (*)
540 540 *
541 541 * kernelbase is less than the abi minimum of 0xc0000000 for memory
542 542 * configurations above 8gb.
543 543 *
544 544 * (*) support for memory configurations above 32gb will require manual tuning
545 545 * of kernelbase to balance out the need of user applications.
546 546 */
547 547
548 548 /* real-time-clock initialization parameters */
549 549 extern time_t process_rtc_config_file(void);
550 550
551 551 uintptr_t kernelbase;
552 552 uintptr_t postbootkernelbase; /* not set till boot loader is gone */
553 553 uintptr_t eprom_kernelbase;
554 554 size_t segmapsize;
555 555 uintptr_t segmap_start;
556 556 int segmapfreelists;
557 557 pgcnt_t npages;
558 558 pgcnt_t orig_npages;
559 559 size_t core_size; /* size of "core" heap */
560 560 uintptr_t core_base; /* base address of "core" heap */
561 561
562 562 /*
563 563 * List of bootstrap pages. We mark these as allocated in startup.
564 564 * release_bootstrap() will free them when we're completely done with
565 565 * the bootstrap.
566 566 */
567 567 static page_t *bootpages;
568 568
569 569 /*
570 570 * boot time pages that have a vnode from the ramdisk will keep that forever.
571 571 */
572 572 static page_t *rd_pages;
573 573
574 574 /*
575 575 * Lower 64K
576 576 */
577 577 static page_t *lower_pages = NULL;
578 578 static int lower_pages_count = 0;
579 579
580 580 struct system_hardware system_hardware;
581 581
582 582 /*
583 583 * Enable some debugging messages concerning memory usage...
584 584 */
585 585 static void
586 586 print_memlist(char *title, struct memlist *mp)
587 587 {
588 588 prom_printf("MEMLIST: %s:\n", title);
589 589 while (mp != NULL) {
590 590 prom_printf("\tAddress 0x%" PRIx64 ", size 0x%" PRIx64 "\n",
591 591 mp->ml_address, mp->ml_size);
592 592 mp = mp->ml_next;
593 593 }
594 594 }
595 595
596 596 /*
597 597 * XX64 need a comment here.. are these just default values, surely
598 598 * we read the "cpuid" type information to figure this out.
599 599 */
600 600 int l2cache_sz = 0x80000;
601 601 int l2cache_linesz = 0x40;
602 602 int l2cache_assoc = 1;
603 603
604 604 static size_t textrepl_min_gb = 10;
605 605
606 606 /*
607 607 * on 64 bit we use a predifined VA range for mapping devices in the kernel
608 608 * on 32 bit the mappings are intermixed in the heap, so we use a bit map
609 609 */
610 610 #ifdef __amd64
611 611
612 612 vmem_t *device_arena;
613 613 uintptr_t toxic_addr = (uintptr_t)NULL;
614 614 size_t toxic_size = 1024 * 1024 * 1024; /* Sparc uses 1 gig too */
615 615
616 616 #else /* __i386 */
617 617
618 618 ulong_t *toxic_bit_map; /* one bit for each 4k of VA in heap_arena */
619 619 size_t toxic_bit_map_len = 0; /* in bits */
620 620
621 621 #endif /* __i386 */
622 622
623 623 /*
624 624 * Simple boot time debug facilities
625 625 */
626 626 static char *prm_dbg_str[] = {
627 627 "%s:%d: '%s' is 0x%x\n",
628 628 "%s:%d: '%s' is 0x%llx\n"
629 629 };
630 630
631 631 int prom_debug;
632 632
633 633 #define PRM_DEBUG(q) if (prom_debug) \
634 634 prom_printf(prm_dbg_str[sizeof (q) >> 3], "startup.c", __LINE__, #q, q);
635 635 #define PRM_POINT(q) if (prom_debug) \
636 636 prom_printf("%s:%d: %s\n", "startup.c", __LINE__, q);
637 637
638 638 /*
639 639 * This structure is used to keep track of the intial allocations
640 640 * done in startup_memlist(). The value of NUM_ALLOCATIONS needs to
641 641 * be >= the number of ADD_TO_ALLOCATIONS() executed in the code.
642 642 */
643 643 #define NUM_ALLOCATIONS 8
644 644 int num_allocations = 0;
645 645 struct {
646 646 void **al_ptr;
647 647 size_t al_size;
648 648 } allocations[NUM_ALLOCATIONS];
649 649 size_t valloc_sz = 0;
650 650 uintptr_t valloc_base;
651 651
652 652 #define ADD_TO_ALLOCATIONS(ptr, size) { \
653 653 size = ROUND_UP_PAGE(size); \
654 654 if (num_allocations == NUM_ALLOCATIONS) \
655 655 panic("too many ADD_TO_ALLOCATIONS()"); \
656 656 allocations[num_allocations].al_ptr = (void**)&ptr; \
657 657 allocations[num_allocations].al_size = size; \
658 658 valloc_sz += size; \
659 659 ++num_allocations; \
660 660 }
661 661
662 662 /*
663 663 * Allocate all the initial memory needed by the page allocator.
664 664 */
665 665 static void
666 666 perform_allocations(void)
667 667 {
668 668 caddr_t mem;
669 669 int i;
670 670 int valloc_align;
671 671
672 672 PRM_DEBUG(valloc_base);
673 673 PRM_DEBUG(valloc_sz);
674 674 valloc_align = mmu.level_size[mmu.max_page_level > 0];
675 675 mem = BOP_ALLOC(bootops, (caddr_t)valloc_base, valloc_sz, valloc_align);
676 676 if (mem != (caddr_t)valloc_base)
677 677 panic("BOP_ALLOC() failed");
678 678 bzero(mem, valloc_sz);
679 679 for (i = 0; i < num_allocations; ++i) {
680 680 *allocations[i].al_ptr = (void *)mem;
681 681 mem += allocations[i].al_size;
682 682 }
683 683 }
684 684
685 685 /*
686 686 * Set up and enable SMAP now before we start other CPUs, but after the kernel's
687 687 * VM has been set up so we can use hot_patch_kernel_text().
688 688 *
689 689 * We can only patch 1, 2, or 4 bytes, but not three bytes. So instead, we
|
↓ open down ↓ |
689 lines elided |
↑ open up ↑ |
690 690 * replace the four byte word at the patch point. See uts/intel/ia32/ml/copy.s
691 691 * for more information on what's going on here.
692 692 */
693 693 static void
694 694 startup_smap(void)
695 695 {
696 696 int i;
697 697 uint32_t inst;
698 698 uint8_t *instp;
699 699 char sym[128];
700 + struct modctl *modp;
700 701
701 702 extern int _smap_enable_patch_count;
702 703 extern int _smap_disable_patch_count;
703 704
704 705 if (disable_smap != 0)
705 706 remove_x86_feature(x86_featureset, X86FSET_SMAP);
706 707
707 708 if (is_x86_feature(x86_featureset, X86FSET_SMAP) == B_FALSE)
708 709 return;
709 710
710 711 for (i = 0; i < _smap_enable_patch_count; i++) {
711 712 int sizep;
712 713
713 714 VERIFY3U(i, <, _smap_enable_patch_count);
714 715 VERIFY(snprintf(sym, sizeof (sym), "_smap_enable_patch_%d", i) <
715 716 sizeof (sym));
716 717 instp = (uint8_t *)(void *)kobj_getelfsym(sym, NULL, &sizep);
717 718 VERIFY(instp != 0);
718 719 inst = (instp[3] << 24) | (SMAP_CLAC_INSTR & 0x00ffffff);
719 720 hot_patch_kernel_text((caddr_t)instp, inst, 4);
720 721 }
721 722
722 723 for (i = 0; i < _smap_disable_patch_count; i++) {
|
↓ open down ↓ |
13 lines elided |
↑ open up ↑ |
723 724 int sizep;
724 725
725 726 VERIFY(snprintf(sym, sizeof (sym), "_smap_disable_patch_%d",
726 727 i) < sizeof (sym));
727 728 instp = (uint8_t *)(void *)kobj_getelfsym(sym, NULL, &sizep);
728 729 VERIFY(instp != 0);
729 730 inst = (instp[3] << 24) | (SMAP_STAC_INSTR & 0x00ffffff);
730 731 hot_patch_kernel_text((caddr_t)instp, inst, 4);
731 732 }
732 733
733 - hot_patch_kernel_text((caddr_t)smap_enable, SMAP_CLAC_INSTR, 4);
734 - hot_patch_kernel_text((caddr_t)smap_disable, SMAP_STAC_INSTR, 4);
734 + /*
735 + * Hotinline calls to smap_enable and smap_disable within
736 + * unix module. Hotinlines in other modules are done on
737 + * mod_load().
738 + */
739 + modp = mod_hold_by_name("unix");
740 + do_hotinlines(modp->mod_mp);
741 + mod_release_mod(modp);
742 +
735 743 setcr4(getcr4() | CR4_SMAP);
736 744 smap_enable();
737 745 }
738 746
739 747 /*
740 748 * Our world looks like this at startup time.
741 749 *
742 750 * In a 32-bit OS, boot loads the kernel text at 0xfe800000 and kernel data
743 751 * at 0xfec00000. On a 64-bit OS, kernel text and data are loaded at
744 752 * 0xffffffff.fe800000 and 0xffffffff.fec00000 respectively. Those
745 753 * addresses are fixed in the binary at link time.
746 754 *
747 755 * On the text page:
748 756 * unix/genunix/krtld/module text loads.
749 757 *
750 758 * On the data page:
751 759 * unix/genunix/krtld/module data loads.
752 760 *
753 761 * Machine-dependent startup code
754 762 */
755 763 void
756 764 startup(void)
757 765 {
758 766 #if !defined(__xpv)
759 767 extern void startup_pci_bios(void);
760 768 #endif
761 769 extern cpuset_t cpu_ready_set;
762 770
763 771 /*
764 772 * Make sure that nobody tries to use sekpm until we have
765 773 * initialized it properly.
766 774 */
767 775 #if defined(__amd64)
768 776 kpm_desired = 1;
769 777 #endif
770 778 kpm_enable = 0;
771 779 CPUSET_ONLY(cpu_ready_set, 0); /* cpu 0 is boot cpu */
772 780
773 781 #if defined(__xpv) /* XXPV fix me! */
774 782 {
775 783 extern int segvn_use_regions;
776 784 segvn_use_regions = 0;
777 785 }
778 786 #endif
779 787 ssp_init();
780 788 progressbar_init();
781 789 startup_init();
782 790 #if defined(__xpv)
783 791 startup_xen_version();
784 792 #endif
785 793 startup_memlist();
786 794 startup_kmem();
787 795 startup_vm();
788 796 #if !defined(__xpv)
789 797 /*
790 798 * Note we need to do this even on fast reboot in order to access
791 799 * the irq routing table (used for pci labels).
792 800 */
793 801 startup_pci_bios();
794 802 startup_smap();
795 803 #endif
796 804 #if defined(__xpv)
797 805 startup_xen_mca();
798 806 #endif
799 807 startup_modules();
800 808
801 809 startup_end();
802 810 }
803 811
804 812 static void
805 813 startup_init()
806 814 {
807 815 PRM_POINT("startup_init() starting...");
808 816
809 817 /*
810 818 * Complete the extraction of cpuid data
811 819 */
812 820 cpuid_pass2(CPU);
813 821
814 822 (void) check_boot_version(BOP_GETVERSION(bootops));
815 823
816 824 /*
817 825 * Check for prom_debug in boot environment
818 826 */
819 827 if (BOP_GETPROPLEN(bootops, "prom_debug") >= 0) {
820 828 ++prom_debug;
821 829 PRM_POINT("prom_debug found in boot enviroment");
822 830 }
823 831
824 832 /*
825 833 * Collect node, cpu and memory configuration information.
826 834 */
827 835 get_system_configuration();
828 836
829 837 /*
830 838 * Halt if this is an unsupported processor.
831 839 */
832 840 if (x86_type == X86_TYPE_486 || x86_type == X86_TYPE_CYRIX_486) {
833 841 printf("\n486 processor (\"%s\") detected.\n",
834 842 CPU->cpu_brandstr);
835 843 halt("This processor is not supported by this release "
836 844 "of Solaris.");
837 845 }
838 846
839 847 PRM_POINT("startup_init() done");
840 848 }
841 849
842 850 /*
843 851 * Callback for copy_memlist_filter() to filter nucleus, kadb/kmdb, (ie.
844 852 * everything mapped above KERNEL_TEXT) pages from phys_avail. Note it
845 853 * also filters out physical page zero. There is some reliance on the
846 854 * boot loader allocating only a few contiguous physical memory chunks.
847 855 */
848 856 static void
849 857 avail_filter(uint64_t *addr, uint64_t *size)
850 858 {
851 859 uintptr_t va;
852 860 uintptr_t next_va;
853 861 pfn_t pfn;
854 862 uint64_t pfn_addr;
855 863 uint64_t pfn_eaddr;
856 864 uint_t prot;
857 865 size_t len;
858 866 uint_t change;
859 867
860 868 if (prom_debug)
861 869 prom_printf("\tFilter: in: a=%" PRIx64 ", s=%" PRIx64 "\n",
862 870 *addr, *size);
863 871
864 872 /*
865 873 * page zero is required for BIOS.. never make it available
866 874 */
867 875 if (*addr == 0) {
868 876 *addr += MMU_PAGESIZE;
869 877 *size -= MMU_PAGESIZE;
870 878 }
871 879
872 880 /*
873 881 * First we trim from the front of the range. Since kbm_probe()
874 882 * walks ranges in virtual order, but addr/size are physical, we need
875 883 * to the list until no changes are seen. This deals with the case
876 884 * where page "p" is mapped at v, page "p + PAGESIZE" is mapped at w
877 885 * but w < v.
878 886 */
879 887 do {
880 888 change = 0;
881 889 for (va = KERNEL_TEXT;
882 890 *size > 0 && kbm_probe(&va, &len, &pfn, &prot) != 0;
883 891 va = next_va) {
884 892
885 893 next_va = va + len;
886 894 pfn_addr = pfn_to_pa(pfn);
887 895 pfn_eaddr = pfn_addr + len;
888 896
889 897 if (pfn_addr <= *addr && pfn_eaddr > *addr) {
890 898 change = 1;
891 899 while (*size > 0 && len > 0) {
892 900 *addr += MMU_PAGESIZE;
893 901 *size -= MMU_PAGESIZE;
894 902 len -= MMU_PAGESIZE;
895 903 }
896 904 }
897 905 }
898 906 if (change && prom_debug)
899 907 prom_printf("\t\ttrim: a=%" PRIx64 ", s=%" PRIx64 "\n",
900 908 *addr, *size);
901 909 } while (change);
902 910
903 911 /*
904 912 * Trim pages from the end of the range.
905 913 */
906 914 for (va = KERNEL_TEXT;
907 915 *size > 0 && kbm_probe(&va, &len, &pfn, &prot) != 0;
908 916 va = next_va) {
909 917
910 918 next_va = va + len;
911 919 pfn_addr = pfn_to_pa(pfn);
912 920
913 921 if (pfn_addr >= *addr && pfn_addr < *addr + *size)
914 922 *size = pfn_addr - *addr;
915 923 }
916 924
917 925 if (prom_debug)
918 926 prom_printf("\tFilter out: a=%" PRIx64 ", s=%" PRIx64 "\n",
919 927 *addr, *size);
920 928 }
921 929
922 930 static void
923 931 kpm_init()
924 932 {
925 933 struct segkpm_crargs b;
926 934
927 935 /*
928 936 * These variables were all designed for sfmmu in which segkpm is
929 937 * mapped using a single pagesize - either 8KB or 4MB. On x86, we
930 938 * might use 2+ page sizes on a single machine, so none of these
931 939 * variables have a single correct value. They are set up as if we
932 940 * always use a 4KB pagesize, which should do no harm. In the long
933 941 * run, we should get rid of KPM's assumption that only a single
934 942 * pagesize is used.
935 943 */
936 944 kpm_pgshft = MMU_PAGESHIFT;
937 945 kpm_pgsz = MMU_PAGESIZE;
938 946 kpm_pgoff = MMU_PAGEOFFSET;
939 947 kpmp2pshft = 0;
940 948 kpmpnpgs = 1;
941 949 ASSERT(((uintptr_t)kpm_vbase & (kpm_pgsz - 1)) == 0);
942 950
943 951 PRM_POINT("about to create segkpm");
944 952 rw_enter(&kas.a_lock, RW_WRITER);
945 953
946 954 if (seg_attach(&kas, kpm_vbase, kpm_size, segkpm) < 0)
947 955 panic("cannot attach segkpm");
948 956
949 957 b.prot = PROT_READ | PROT_WRITE;
950 958 b.nvcolors = 1;
951 959
952 960 if (segkpm_create(segkpm, (caddr_t)&b) != 0)
953 961 panic("segkpm_create segkpm");
954 962
955 963 rw_exit(&kas.a_lock);
956 964
957 965 kpm_enable = 1;
958 966
959 967 /*
960 968 * As the KPM was disabled while setting up the system, go back and fix
961 969 * CPU zero's access to its user page table. This is a bit gross, but
962 970 * we have a chicken and egg problem otherwise.
963 971 */
964 972 ASSERT(CPU->cpu_hat_info->hci_user_l3ptes == NULL);
965 973 CPU->cpu_hat_info->hci_user_l3ptes =
966 974 (x86pte_t *)hat_kpm_mapin_pfn(CPU->cpu_hat_info->hci_user_l3pfn);
967 975 }
968 976
969 977 /*
970 978 * The debug info page provides enough information to allow external
971 979 * inspectors (e.g. when running under a hypervisor) to bootstrap
972 980 * themselves into allowing full-blown kernel debugging.
973 981 */
974 982 static void
975 983 init_debug_info(void)
976 984 {
977 985 caddr_t mem;
978 986 debug_info_t *di;
979 987
980 988 #ifndef __lint
981 989 ASSERT(sizeof (debug_info_t) < MMU_PAGESIZE);
982 990 #endif
983 991
984 992 mem = BOP_ALLOC(bootops, (caddr_t)DEBUG_INFO_VA, MMU_PAGESIZE,
985 993 MMU_PAGESIZE);
986 994
987 995 if (mem != (caddr_t)DEBUG_INFO_VA)
988 996 panic("BOP_ALLOC() failed");
989 997 bzero(mem, MMU_PAGESIZE);
990 998
991 999 di = (debug_info_t *)mem;
992 1000
993 1001 di->di_magic = DEBUG_INFO_MAGIC;
994 1002 di->di_version = DEBUG_INFO_VERSION;
995 1003 di->di_modules = (uintptr_t)&modules;
996 1004 di->di_s_text = (uintptr_t)s_text;
997 1005 di->di_e_text = (uintptr_t)e_text;
998 1006 di->di_s_data = (uintptr_t)s_data;
999 1007 di->di_e_data = (uintptr_t)e_data;
1000 1008 di->di_hat_htable_off = offsetof(hat_t, hat_htable);
1001 1009 di->di_ht_pfn_off = offsetof(htable_t, ht_pfn);
1002 1010 }
1003 1011
1004 1012 /*
1005 1013 * Build the memlists and other kernel essential memory system data structures.
1006 1014 * This is everything at valloc_base.
1007 1015 */
1008 1016 static void
1009 1017 startup_memlist(void)
1010 1018 {
1011 1019 size_t memlist_sz;
1012 1020 size_t memseg_sz;
1013 1021 size_t pagehash_sz;
1014 1022 size_t pp_sz;
1015 1023 uintptr_t va;
1016 1024 size_t len;
1017 1025 uint_t prot;
1018 1026 pfn_t pfn;
1019 1027 int memblocks;
1020 1028 pfn_t rsvd_high_pfn;
1021 1029 pgcnt_t rsvd_pgcnt;
1022 1030 size_t rsvdmemlist_sz;
1023 1031 int rsvdmemblocks;
1024 1032 caddr_t pagecolor_mem;
1025 1033 size_t pagecolor_memsz;
1026 1034 caddr_t page_ctrs_mem;
1027 1035 size_t page_ctrs_size;
1028 1036 size_t pse_table_alloc_size;
1029 1037 struct memlist *current;
1030 1038 extern void startup_build_mem_nodes(struct memlist *);
1031 1039
1032 1040 /* XX64 fix these - they should be in include files */
1033 1041 extern size_t page_coloring_init(uint_t, int, int);
1034 1042 extern void page_coloring_setup(caddr_t);
1035 1043
1036 1044 PRM_POINT("startup_memlist() starting...");
1037 1045
1038 1046 /*
1039 1047 * Use leftover large page nucleus text/data space for loadable modules.
1040 1048 * Use at most MODTEXT/MODDATA.
1041 1049 */
1042 1050 len = kbm_nucleus_size;
1043 1051 ASSERT(len > MMU_PAGESIZE);
1044 1052
1045 1053 moddata = (caddr_t)ROUND_UP_PAGE(e_data);
1046 1054 e_moddata = (caddr_t)P2ROUNDUP((uintptr_t)e_data, (uintptr_t)len);
1047 1055 if (e_moddata - moddata > MODDATA)
1048 1056 e_moddata = moddata + MODDATA;
1049 1057
1050 1058 modtext = (caddr_t)ROUND_UP_PAGE(e_text);
1051 1059 e_modtext = (caddr_t)P2ROUNDUP((uintptr_t)e_text, (uintptr_t)len);
1052 1060 if (e_modtext - modtext > MODTEXT)
1053 1061 e_modtext = modtext + MODTEXT;
1054 1062
1055 1063 econtig = e_moddata;
1056 1064
1057 1065 PRM_DEBUG(modtext);
1058 1066 PRM_DEBUG(e_modtext);
1059 1067 PRM_DEBUG(moddata);
1060 1068 PRM_DEBUG(e_moddata);
1061 1069 PRM_DEBUG(econtig);
1062 1070
1063 1071 /*
1064 1072 * Examine the boot loader physical memory map to find out:
1065 1073 * - total memory in system - physinstalled
1066 1074 * - the max physical address - physmax
1067 1075 * - the number of discontiguous segments of memory.
1068 1076 */
1069 1077 if (prom_debug)
1070 1078 print_memlist("boot physinstalled",
1071 1079 bootops->boot_mem->physinstalled);
1072 1080 installed_top_size_ex(bootops->boot_mem->physinstalled, &physmax,
1073 1081 &physinstalled, &memblocks);
1074 1082 PRM_DEBUG(physmax);
1075 1083 PRM_DEBUG(physinstalled);
1076 1084 PRM_DEBUG(memblocks);
1077 1085
1078 1086 /*
1079 1087 * Compute maximum physical address for memory DR operations.
1080 1088 * Memory DR operations are unsupported on xpv or 32bit OSes.
1081 1089 */
1082 1090 #ifdef __amd64
1083 1091 if (plat_dr_support_memory()) {
1084 1092 if (plat_dr_physmax == 0) {
1085 1093 uint_t pabits = UINT_MAX;
1086 1094
1087 1095 cpuid_get_addrsize(CPU, &pabits, NULL);
1088 1096 plat_dr_physmax = btop(1ULL << pabits);
1089 1097 }
1090 1098 if (plat_dr_physmax > PHYSMEM_MAX64)
1091 1099 plat_dr_physmax = PHYSMEM_MAX64;
1092 1100 } else
1093 1101 #endif
1094 1102 plat_dr_physmax = 0;
1095 1103
1096 1104 /*
1097 1105 * Examine the bios reserved memory to find out:
1098 1106 * - the number of discontiguous segments of memory.
1099 1107 */
1100 1108 if (prom_debug)
1101 1109 print_memlist("boot reserved mem",
1102 1110 bootops->boot_mem->rsvdmem);
1103 1111 installed_top_size_ex(bootops->boot_mem->rsvdmem, &rsvd_high_pfn,
1104 1112 &rsvd_pgcnt, &rsvdmemblocks);
1105 1113 PRM_DEBUG(rsvd_high_pfn);
1106 1114 PRM_DEBUG(rsvd_pgcnt);
1107 1115 PRM_DEBUG(rsvdmemblocks);
1108 1116
1109 1117 /*
1110 1118 * Initialize hat's mmu parameters.
1111 1119 * Check for enforce-prot-exec in boot environment. It's used to
1112 1120 * enable/disable support for the page table entry NX bit.
1113 1121 * The default is to enforce PROT_EXEC on processors that support NX.
1114 1122 * Boot seems to round up the "len", but 8 seems to be big enough.
1115 1123 */
1116 1124 mmu_init();
1117 1125
1118 1126 #ifdef __i386
1119 1127 /*
1120 1128 * physmax is lowered if there is more memory than can be
1121 1129 * physically addressed in 32 bit (PAE/non-PAE) modes.
1122 1130 */
1123 1131 if (mmu.pae_hat) {
1124 1132 if (PFN_ABOVE64G(physmax)) {
1125 1133 physinstalled -= (physmax - (PFN_64G - 1));
1126 1134 physmax = PFN_64G - 1;
1127 1135 }
1128 1136 } else {
1129 1137 if (PFN_ABOVE4G(physmax)) {
1130 1138 physinstalled -= (physmax - (PFN_4G - 1));
1131 1139 physmax = PFN_4G - 1;
1132 1140 }
1133 1141 }
1134 1142 #endif
1135 1143
1136 1144 startup_build_mem_nodes(bootops->boot_mem->physinstalled);
1137 1145
1138 1146 if (BOP_GETPROPLEN(bootops, "enforce-prot-exec") >= 0) {
1139 1147 int len = BOP_GETPROPLEN(bootops, "enforce-prot-exec");
1140 1148 char value[8];
1141 1149
1142 1150 if (len < 8)
1143 1151 (void) BOP_GETPROP(bootops, "enforce-prot-exec", value);
1144 1152 else
1145 1153 (void) strcpy(value, "");
1146 1154 if (strcmp(value, "off") == 0)
1147 1155 mmu.pt_nx = 0;
1148 1156 }
1149 1157 PRM_DEBUG(mmu.pt_nx);
1150 1158
1151 1159 /*
1152 1160 * We will need page_t's for every page in the system, except for
1153 1161 * memory mapped at or above above the start of the kernel text segment.
1154 1162 *
1155 1163 * pages above e_modtext are attributed to kernel debugger (obp_pages)
1156 1164 */
1157 1165 npages = physinstalled - 1; /* avail_filter() skips page 0, so "- 1" */
1158 1166 obp_pages = 0;
1159 1167 va = KERNEL_TEXT;
1160 1168 while (kbm_probe(&va, &len, &pfn, &prot) != 0) {
1161 1169 npages -= len >> MMU_PAGESHIFT;
1162 1170 if (va >= (uintptr_t)e_moddata)
1163 1171 obp_pages += len >> MMU_PAGESHIFT;
1164 1172 va += len;
1165 1173 }
1166 1174 PRM_DEBUG(npages);
1167 1175 PRM_DEBUG(obp_pages);
1168 1176
1169 1177 /*
1170 1178 * If physmem is patched to be non-zero, use it instead of the computed
1171 1179 * value unless it is larger than the actual amount of memory on hand.
1172 1180 */
1173 1181 if (physmem == 0 || physmem > npages) {
1174 1182 physmem = npages;
1175 1183 } else if (physmem < npages) {
1176 1184 orig_npages = npages;
1177 1185 npages = physmem;
1178 1186 }
1179 1187 PRM_DEBUG(physmem);
1180 1188
1181 1189 /*
1182 1190 * We now compute the sizes of all the initial allocations for
1183 1191 * structures the kernel needs in order do kmem_alloc(). These
1184 1192 * include:
1185 1193 * memsegs
1186 1194 * memlists
1187 1195 * page hash table
1188 1196 * page_t's
1189 1197 * page coloring data structs
1190 1198 */
1191 1199 memseg_sz = sizeof (struct memseg) * (memblocks + POSS_NEW_FRAGMENTS);
1192 1200 ADD_TO_ALLOCATIONS(memseg_base, memseg_sz);
1193 1201 PRM_DEBUG(memseg_sz);
1194 1202
1195 1203 /*
1196 1204 * Reserve space for memlists. There's no real good way to know exactly
1197 1205 * how much room we'll need, but this should be a good upper bound.
1198 1206 */
1199 1207 memlist_sz = ROUND_UP_PAGE(2 * sizeof (struct memlist) *
1200 1208 (memblocks + POSS_NEW_FRAGMENTS));
1201 1209 ADD_TO_ALLOCATIONS(memlist, memlist_sz);
1202 1210 PRM_DEBUG(memlist_sz);
1203 1211
1204 1212 /*
1205 1213 * Reserve space for bios reserved memlists.
1206 1214 */
1207 1215 rsvdmemlist_sz = ROUND_UP_PAGE(2 * sizeof (struct memlist) *
1208 1216 (rsvdmemblocks + POSS_NEW_FRAGMENTS));
1209 1217 ADD_TO_ALLOCATIONS(bios_rsvd, rsvdmemlist_sz);
1210 1218 PRM_DEBUG(rsvdmemlist_sz);
1211 1219
1212 1220 /* LINTED */
1213 1221 ASSERT(P2SAMEHIGHBIT((1 << PP_SHIFT), sizeof (struct page)));
1214 1222 /*
1215 1223 * The page structure hash table size is a power of 2
1216 1224 * such that the average hash chain length is PAGE_HASHAVELEN.
1217 1225 */
1218 1226 page_hashsz = npages / PAGE_HASHAVELEN;
1219 1227 page_hashsz_shift = highbit(page_hashsz);
1220 1228 page_hashsz = 1 << page_hashsz_shift;
1221 1229 pagehash_sz = sizeof (struct page *) * page_hashsz;
1222 1230 ADD_TO_ALLOCATIONS(page_hash, pagehash_sz);
1223 1231 PRM_DEBUG(pagehash_sz);
1224 1232
1225 1233 /*
1226 1234 * Set aside room for the page structures themselves.
1227 1235 */
1228 1236 PRM_DEBUG(npages);
1229 1237 pp_sz = sizeof (struct page) * npages;
1230 1238 ADD_TO_ALLOCATIONS(pp_base, pp_sz);
1231 1239 PRM_DEBUG(pp_sz);
1232 1240
1233 1241 /*
1234 1242 * determine l2 cache info and memory size for page coloring
1235 1243 */
1236 1244 (void) getl2cacheinfo(CPU,
1237 1245 &l2cache_sz, &l2cache_linesz, &l2cache_assoc);
1238 1246 pagecolor_memsz =
1239 1247 page_coloring_init(l2cache_sz, l2cache_linesz, l2cache_assoc);
1240 1248 ADD_TO_ALLOCATIONS(pagecolor_mem, pagecolor_memsz);
1241 1249 PRM_DEBUG(pagecolor_memsz);
1242 1250
1243 1251 page_ctrs_size = page_ctrs_sz();
1244 1252 ADD_TO_ALLOCATIONS(page_ctrs_mem, page_ctrs_size);
1245 1253 PRM_DEBUG(page_ctrs_size);
1246 1254
1247 1255 /*
1248 1256 * Allocate the array that protects pp->p_selock.
1249 1257 */
1250 1258 pse_shift = size_pse_array(physmem, max_ncpus);
1251 1259 pse_table_size = 1 << pse_shift;
1252 1260 pse_table_alloc_size = pse_table_size * sizeof (pad_mutex_t);
1253 1261 ADD_TO_ALLOCATIONS(pse_mutex, pse_table_alloc_size);
1254 1262
1255 1263 #if defined(__amd64)
1256 1264 valloc_sz = ROUND_UP_LPAGE(valloc_sz);
1257 1265 valloc_base = VALLOC_BASE;
1258 1266
1259 1267 /*
1260 1268 * The default values of VALLOC_BASE and SEGKPM_BASE should work
1261 1269 * for values of physmax up to 256GB (1/4 TB). They need adjusting when
1262 1270 * memory is at addresses above 256GB. When adjusted, segkpm_base must
1263 1271 * be aligned on KERNEL_REDZONE_SIZE boundary (span of top level pte).
1264 1272 *
1265 1273 * In the general case (>256GB), we use (4 * physmem) for the
1266 1274 * kernel's virtual addresses, which is divided approximately
1267 1275 * as follows:
1268 1276 * - 1 * physmem for segkpm
1269 1277 * - 1.5 * physmem for segzio
1270 1278 * - 1.5 * physmem for heap
1271 1279 * Total: 4.0 * physmem
1272 1280 *
1273 1281 * Note that the segzio and heap sizes are more than physmem so that
1274 1282 * VA fragmentation does not prevent either of them from being
1275 1283 * able to use nearly all of physmem. The value of 1.5x is determined
1276 1284 * experimentally and may need to change if the workload changes.
1277 1285 */
1278 1286 if (physmax + 1 > mmu_btop(TERABYTE / 4) ||
1279 1287 plat_dr_physmax > mmu_btop(TERABYTE / 4)) {
1280 1288 uint64_t kpm_resv_amount = mmu_ptob(physmax + 1);
1281 1289
1282 1290 if (kpm_resv_amount < mmu_ptob(plat_dr_physmax)) {
1283 1291 kpm_resv_amount = mmu_ptob(plat_dr_physmax);
1284 1292 }
1285 1293
1286 1294 /*
1287 1295 * This is what actually controls the KVA : UVA split.
1288 1296 * The kernel uses high VA, and this is lowering the
1289 1297 * boundary, thus increasing the amount of VA for the kernel.
1290 1298 * This gives the kernel 4 * (amount of physical memory) VA.
1291 1299 *
1292 1300 * The maximum VA is UINT64_MAX and we are using
1293 1301 * 64-bit 2's complement math, so e.g. if you have 512GB
1294 1302 * of memory, segkpm_base = -(4 * 512GB) == -2TB ==
1295 1303 * UINT64_MAX - 2TB (approximately). So the kernel's
1296 1304 * VA is [UINT64_MAX-2TB to UINT64_MAX].
1297 1305 */
1298 1306 segkpm_base = -(P2ROUNDUP((4 * kpm_resv_amount),
1299 1307 KERNEL_REDZONE_SIZE));
1300 1308
1301 1309 /* make sure we leave some space for user apps above hole */
1302 1310 segkpm_base = MAX(segkpm_base, AMD64_VA_HOLE_END + TERABYTE);
1303 1311 if (segkpm_base > SEGKPM_BASE)
1304 1312 segkpm_base = SEGKPM_BASE;
1305 1313 PRM_DEBUG(segkpm_base);
1306 1314
1307 1315 valloc_base = segkpm_base + P2ROUNDUP(kpm_resv_amount, ONE_GIG);
1308 1316 if (valloc_base < segkpm_base)
1309 1317 panic("not enough kernel VA to support memory size");
1310 1318 PRM_DEBUG(valloc_base);
1311 1319 }
1312 1320 #else /* __i386 */
1313 1321 valloc_base = (uintptr_t)(MISC_VA_BASE - valloc_sz);
1314 1322 valloc_base = P2ALIGN(valloc_base, mmu.level_size[1]);
1315 1323 PRM_DEBUG(valloc_base);
1316 1324 #endif /* __i386 */
1317 1325
1318 1326 /*
1319 1327 * do all the initial allocations
1320 1328 */
1321 1329 perform_allocations();
1322 1330
1323 1331 /*
1324 1332 * Build phys_install and phys_avail in kernel memspace.
1325 1333 * - phys_install should be all memory in the system.
1326 1334 * - phys_avail is phys_install minus any memory mapped before this
1327 1335 * point above KERNEL_TEXT.
1328 1336 */
1329 1337 current = phys_install = memlist;
1330 1338 copy_memlist_filter(bootops->boot_mem->physinstalled, ¤t, NULL);
1331 1339 if ((caddr_t)current > (caddr_t)memlist + memlist_sz)
1332 1340 panic("physinstalled was too big!");
1333 1341 if (prom_debug)
1334 1342 print_memlist("phys_install", phys_install);
1335 1343
1336 1344 phys_avail = current;
1337 1345 PRM_POINT("Building phys_avail:\n");
1338 1346 copy_memlist_filter(bootops->boot_mem->physinstalled, ¤t,
1339 1347 avail_filter);
1340 1348 if ((caddr_t)current > (caddr_t)memlist + memlist_sz)
1341 1349 panic("physavail was too big!");
1342 1350 if (prom_debug)
1343 1351 print_memlist("phys_avail", phys_avail);
1344 1352 #ifndef __xpv
1345 1353 /*
1346 1354 * Free unused memlist items, which may be used by memory DR driver
1347 1355 * at runtime.
1348 1356 */
1349 1357 if ((caddr_t)current < (caddr_t)memlist + memlist_sz) {
1350 1358 memlist_free_block((caddr_t)current,
1351 1359 (caddr_t)memlist + memlist_sz - (caddr_t)current);
1352 1360 }
1353 1361 #endif
1354 1362
1355 1363 /*
1356 1364 * Build bios reserved memspace
1357 1365 */
1358 1366 current = bios_rsvd;
1359 1367 copy_memlist_filter(bootops->boot_mem->rsvdmem, ¤t, NULL);
1360 1368 if ((caddr_t)current > (caddr_t)bios_rsvd + rsvdmemlist_sz)
1361 1369 panic("bios_rsvd was too big!");
1362 1370 if (prom_debug)
1363 1371 print_memlist("bios_rsvd", bios_rsvd);
1364 1372 #ifndef __xpv
1365 1373 /*
1366 1374 * Free unused memlist items, which may be used by memory DR driver
1367 1375 * at runtime.
1368 1376 */
1369 1377 if ((caddr_t)current < (caddr_t)bios_rsvd + rsvdmemlist_sz) {
1370 1378 memlist_free_block((caddr_t)current,
1371 1379 (caddr_t)bios_rsvd + rsvdmemlist_sz - (caddr_t)current);
1372 1380 }
1373 1381 #endif
1374 1382
1375 1383 /*
1376 1384 * setup page coloring
1377 1385 */
1378 1386 page_coloring_setup(pagecolor_mem);
1379 1387 page_lock_init(); /* currently a no-op */
1380 1388
1381 1389 /*
1382 1390 * free page list counters
1383 1391 */
1384 1392 (void) page_ctrs_alloc(page_ctrs_mem);
1385 1393
1386 1394 /*
1387 1395 * Size the pcf array based on the number of cpus in the box at
1388 1396 * boot time.
1389 1397 */
1390 1398
1391 1399 pcf_init();
1392 1400
1393 1401 /*
1394 1402 * Initialize the page structures from the memory lists.
1395 1403 */
1396 1404 availrmem_initial = availrmem = freemem = 0;
1397 1405 PRM_POINT("Calling kphysm_init()...");
1398 1406 npages = kphysm_init(pp_base, npages);
1399 1407 PRM_POINT("kphysm_init() done");
1400 1408 PRM_DEBUG(npages);
1401 1409
1402 1410 init_debug_info();
1403 1411
1404 1412 /*
1405 1413 * Now that page_t's have been initialized, remove all the
1406 1414 * initial allocation pages from the kernel free page lists.
1407 1415 */
1408 1416 boot_mapin((caddr_t)valloc_base, valloc_sz);
1409 1417 boot_mapin((caddr_t)MISC_VA_BASE, MISC_VA_SIZE);
1410 1418 PRM_POINT("startup_memlist() done");
1411 1419
1412 1420 PRM_DEBUG(valloc_sz);
1413 1421
1414 1422 #if defined(__amd64)
1415 1423 if ((availrmem >> (30 - MMU_PAGESHIFT)) >=
1416 1424 textrepl_min_gb && l2cache_sz <= 2 << 20) {
1417 1425 extern size_t textrepl_size_thresh;
1418 1426 textrepl_size_thresh = (16 << 20) - 1;
1419 1427 }
1420 1428 #endif
1421 1429 }
1422 1430
1423 1431 /*
1424 1432 * Layout the kernel's part of address space and initialize kmem allocator.
1425 1433 */
1426 1434 static void
1427 1435 startup_kmem(void)
1428 1436 {
1429 1437 extern void page_set_colorequiv_arr(void);
1430 1438 #if !defined(__xpv)
1431 1439 extern uint64_t kpti_kbase;
1432 1440 #endif
1433 1441
1434 1442 PRM_POINT("startup_kmem() starting...");
1435 1443
1436 1444 #if defined(__amd64)
1437 1445 if (eprom_kernelbase && eprom_kernelbase != KERNELBASE)
1438 1446 cmn_err(CE_NOTE, "!kernelbase cannot be changed on 64-bit "
1439 1447 "systems.");
1440 1448 kernelbase = segkpm_base - KERNEL_REDZONE_SIZE;
1441 1449 core_base = (uintptr_t)COREHEAP_BASE;
1442 1450 core_size = (size_t)MISC_VA_BASE - COREHEAP_BASE;
1443 1451 #else /* __i386 */
1444 1452 /*
1445 1453 * We configure kernelbase based on:
1446 1454 *
1447 1455 * 1. user specified kernelbase via eeprom command. Value cannot exceed
1448 1456 * KERNELBASE_MAX. we large page align eprom_kernelbase
1449 1457 *
1450 1458 * 2. Default to KERNELBASE and adjust to 2X less the size for page_t.
1451 1459 * On large memory systems we must lower kernelbase to allow
1452 1460 * enough room for page_t's for all of memory.
1453 1461 *
1454 1462 * The value set here, might be changed a little later.
1455 1463 */
1456 1464 if (eprom_kernelbase) {
1457 1465 kernelbase = eprom_kernelbase & mmu.level_mask[1];
1458 1466 if (kernelbase > KERNELBASE_MAX)
1459 1467 kernelbase = KERNELBASE_MAX;
1460 1468 } else {
1461 1469 kernelbase = (uintptr_t)KERNELBASE;
1462 1470 kernelbase -= ROUND_UP_4MEG(2 * valloc_sz);
1463 1471 }
1464 1472 ASSERT((kernelbase & mmu.level_offset[1]) == 0);
1465 1473 core_base = valloc_base;
1466 1474 core_size = 0;
1467 1475 #endif /* __i386 */
1468 1476
1469 1477 PRM_DEBUG(core_base);
1470 1478 PRM_DEBUG(core_size);
1471 1479 PRM_DEBUG(kernelbase);
1472 1480
1473 1481 #if defined(__i386)
1474 1482 segkp_fromheap = 1;
1475 1483 #endif /* __i386 */
1476 1484
1477 1485 ekernelheap = (char *)core_base;
1478 1486 PRM_DEBUG(ekernelheap);
1479 1487
1480 1488 /*
1481 1489 * Now that we know the real value of kernelbase,
1482 1490 * update variables that were initialized with a value of
1483 1491 * KERNELBASE (in common/conf/param.c).
1484 1492 *
1485 1493 * XXX The problem with this sort of hackery is that the
1486 1494 * compiler just may feel like putting the const declarations
1487 1495 * (in param.c) into the .text section. Perhaps they should
1488 1496 * just be declared as variables there?
1489 1497 */
1490 1498
1491 1499 *(uintptr_t *)&_kernelbase = kernelbase;
1492 1500 *(uintptr_t *)&_userlimit = kernelbase;
1493 1501 #if defined(__amd64)
1494 1502 *(uintptr_t *)&_userlimit -= KERNELBASE - USERLIMIT;
1495 1503 #if !defined(__xpv)
1496 1504 kpti_kbase = kernelbase;
1497 1505 #endif
1498 1506 #else
1499 1507 *(uintptr_t *)&_userlimit32 = _userlimit;
1500 1508 #endif
1501 1509 PRM_DEBUG(_kernelbase);
1502 1510 PRM_DEBUG(_userlimit);
1503 1511 PRM_DEBUG(_userlimit32);
1504 1512
1505 1513 /* We have to re-do this now that we've modified _userlimit. */
1506 1514 mmu_calc_user_slots();
1507 1515
1508 1516 layout_kernel_va();
1509 1517
1510 1518 #if defined(__i386)
1511 1519 /*
1512 1520 * If segmap is too large we can push the bottom of the kernel heap
1513 1521 * higher than the base. Or worse, it could exceed the top of the
1514 1522 * VA space entirely, causing it to wrap around.
1515 1523 */
1516 1524 if (kernelheap >= ekernelheap || (uintptr_t)kernelheap < kernelbase)
1517 1525 panic("too little address space available for kernelheap,"
1518 1526 " use eeprom for lower kernelbase or smaller segmapsize");
1519 1527 #endif /* __i386 */
1520 1528
1521 1529 /*
1522 1530 * Initialize the kernel heap. Note 3rd argument must be > 1st.
1523 1531 */
1524 1532 kernelheap_init(kernelheap, ekernelheap,
1525 1533 kernelheap + MMU_PAGESIZE,
1526 1534 (void *)core_base, (void *)(core_base + core_size));
1527 1535
1528 1536 #if defined(__xpv)
1529 1537 /*
1530 1538 * Link pending events struct into cpu struct
1531 1539 */
1532 1540 CPU->cpu_m.mcpu_evt_pend = &cpu0_evt_data;
1533 1541 #endif
1534 1542 /*
1535 1543 * Initialize kernel memory allocator.
1536 1544 */
1537 1545 kmem_init();
1538 1546
1539 1547 /*
1540 1548 * Factor in colorequiv to check additional 'equivalent' bins
1541 1549 */
1542 1550 page_set_colorequiv_arr();
1543 1551
1544 1552 /*
1545 1553 * print this out early so that we know what's going on
1546 1554 */
1547 1555 print_x86_featureset(x86_featureset);
1548 1556
1549 1557 /*
1550 1558 * Initialize bp_mapin().
1551 1559 */
1552 1560 bp_init(MMU_PAGESIZE, HAT_STORECACHING_OK);
1553 1561
1554 1562 /*
1555 1563 * orig_npages is non-zero if physmem has been configured for less
1556 1564 * than the available memory.
1557 1565 */
1558 1566 if (orig_npages) {
1559 1567 cmn_err(CE_WARN, "!%slimiting physmem to 0x%lx of 0x%lx pages",
1560 1568 (npages == PHYSMEM ? "Due to virtual address space " : ""),
1561 1569 npages, orig_npages);
1562 1570 }
1563 1571 #if defined(__i386)
1564 1572 if (eprom_kernelbase && (eprom_kernelbase != kernelbase))
1565 1573 cmn_err(CE_WARN, "kernelbase value, User specified 0x%lx, "
1566 1574 "System using 0x%lx",
1567 1575 (uintptr_t)eprom_kernelbase, (uintptr_t)kernelbase);
1568 1576 #endif
1569 1577
1570 1578 #ifdef KERNELBASE_ABI_MIN
1571 1579 if (kernelbase < (uintptr_t)KERNELBASE_ABI_MIN) {
1572 1580 cmn_err(CE_NOTE, "!kernelbase set to 0x%lx, system is not "
1573 1581 "i386 ABI compliant.", (uintptr_t)kernelbase);
1574 1582 }
1575 1583 #endif
1576 1584
1577 1585 #ifndef __xpv
1578 1586 if (plat_dr_support_memory()) {
1579 1587 mem_config_init();
1580 1588 }
1581 1589 #else /* __xpv */
1582 1590 /*
1583 1591 * Some of the xen start information has to be relocated up
1584 1592 * into the kernel's permanent address space.
1585 1593 */
1586 1594 PRM_POINT("calling xen_relocate_start_info()");
1587 1595 xen_relocate_start_info();
1588 1596 PRM_POINT("xen_relocate_start_info() done");
1589 1597
1590 1598 /*
1591 1599 * (Update the vcpu pointer in our cpu structure to point into
1592 1600 * the relocated shared info.)
1593 1601 */
1594 1602 CPU->cpu_m.mcpu_vcpu_info =
1595 1603 &HYPERVISOR_shared_info->vcpu_info[CPU->cpu_id];
1596 1604 #endif /* __xpv */
1597 1605
1598 1606 PRM_POINT("startup_kmem() done");
1599 1607 }
1600 1608
1601 1609 #ifndef __xpv
1602 1610 /*
1603 1611 * If we have detected that we are running in an HVM environment, we need
1604 1612 * to prepend the PV driver directory to the module search path.
1605 1613 */
1606 1614 #define HVM_MOD_DIR "/platform/i86hvm/kernel"
1607 1615 static void
1608 1616 update_default_path()
1609 1617 {
1610 1618 char *current, *newpath;
1611 1619 int newlen;
1612 1620
1613 1621 /*
1614 1622 * We are about to resync with krtld. krtld will reset its
1615 1623 * internal module search path iff Solaris has set default_path.
1616 1624 * We want to be sure we're prepending this new directory to the
1617 1625 * right search path.
1618 1626 */
1619 1627 current = (default_path == NULL) ? kobj_module_path : default_path;
1620 1628
1621 1629 newlen = strlen(HVM_MOD_DIR) + strlen(current) + 2;
1622 1630 newpath = kmem_alloc(newlen, KM_SLEEP);
1623 1631 (void) strcpy(newpath, HVM_MOD_DIR);
1624 1632 (void) strcat(newpath, " ");
1625 1633 (void) strcat(newpath, current);
1626 1634
1627 1635 default_path = newpath;
1628 1636 }
1629 1637 #endif
1630 1638
1631 1639 static void
1632 1640 startup_modules(void)
1633 1641 {
1634 1642 int cnt;
1635 1643 extern void prom_setup(void);
1636 1644 int32_t v, h;
1637 1645 char d[11];
1638 1646 char *cp;
1639 1647 cmi_hdl_t hdl;
1640 1648
1641 1649 PRM_POINT("startup_modules() starting...");
1642 1650
1643 1651 #ifndef __xpv
1644 1652 /*
1645 1653 * Initialize ten-micro second timer so that drivers will
1646 1654 * not get short changed in their init phase. This was
1647 1655 * not getting called until clkinit which, on fast cpu's
1648 1656 * caused the drv_usecwait to be way too short.
1649 1657 */
1650 1658 microfind();
1651 1659
1652 1660 if ((get_hwenv() & HW_XEN_HVM) != 0)
1653 1661 update_default_path();
1654 1662 #endif
1655 1663
1656 1664 /*
1657 1665 * Read the GMT lag from /etc/rtc_config.
1658 1666 */
1659 1667 sgmtl(process_rtc_config_file());
1660 1668
1661 1669 /*
1662 1670 * Calculate default settings of system parameters based upon
1663 1671 * maxusers, yet allow to be overridden via the /etc/system file.
1664 1672 */
1665 1673 param_calc(0);
1666 1674
1667 1675 mod_setup();
1668 1676
1669 1677 /*
1670 1678 * Initialize system parameters.
1671 1679 */
1672 1680 param_init();
1673 1681
1674 1682 /*
1675 1683 * Initialize the default brands
1676 1684 */
1677 1685 brand_init();
1678 1686
1679 1687 /*
1680 1688 * maxmem is the amount of physical memory we're playing with.
1681 1689 */
1682 1690 maxmem = physmem;
1683 1691
1684 1692 /*
1685 1693 * Initialize segment management stuff.
1686 1694 */
1687 1695 seg_init();
1688 1696
1689 1697 if (modload("fs", "specfs") == -1)
1690 1698 halt("Can't load specfs");
1691 1699
1692 1700 if (modload("fs", "devfs") == -1)
1693 1701 halt("Can't load devfs");
1694 1702
1695 1703 if (modload("fs", "dev") == -1)
1696 1704 halt("Can't load dev");
1697 1705
1698 1706 if (modload("fs", "procfs") == -1)
1699 1707 halt("Can't load procfs");
1700 1708
1701 1709 (void) modloadonly("sys", "lbl_edition");
1702 1710
1703 1711 dispinit();
1704 1712
1705 1713 /* Read cluster configuration data. */
1706 1714 clconf_init();
1707 1715
1708 1716 #if defined(__xpv)
1709 1717 (void) ec_init();
1710 1718 gnttab_init();
1711 1719 (void) xs_early_init();
1712 1720 #endif /* __xpv */
1713 1721
1714 1722 /*
1715 1723 * Create a kernel device tree. First, create rootnex and
1716 1724 * then invoke bus specific code to probe devices.
1717 1725 */
1718 1726 setup_ddi();
1719 1727
1720 1728 #ifdef __xpv
1721 1729 if (DOMAIN_IS_INITDOMAIN(xen_info))
1722 1730 #endif
1723 1731 {
1724 1732 id_t smid;
1725 1733 smbios_system_t smsys;
1726 1734 smbios_info_t sminfo;
1727 1735 char *mfg;
1728 1736 /*
1729 1737 * Load the System Management BIOS into the global ksmbios
1730 1738 * handle, if an SMBIOS is present on this system.
1731 1739 * Also set "si-hw-provider" property, if not already set.
1732 1740 */
1733 1741 ksmbios = smbios_open(NULL, SMB_VERSION, ksmbios_flags, NULL);
1734 1742 if (ksmbios != NULL &&
1735 1743 ((smid = smbios_info_system(ksmbios, &smsys)) != SMB_ERR) &&
1736 1744 (smbios_info_common(ksmbios, smid, &sminfo)) != SMB_ERR) {
1737 1745 mfg = (char *)sminfo.smbi_manufacturer;
1738 1746 if (BOP_GETPROPLEN(bootops, "si-hw-provider") < 0) {
1739 1747 extern char hw_provider[];
1740 1748 int i;
1741 1749 for (i = 0; i < SYS_NMLN; i++) {
1742 1750 if (isprint(mfg[i]))
1743 1751 hw_provider[i] = mfg[i];
1744 1752 else {
1745 1753 hw_provider[i] = '\0';
1746 1754 break;
1747 1755 }
1748 1756 }
1749 1757 hw_provider[SYS_NMLN - 1] = '\0';
1750 1758 }
1751 1759 }
1752 1760 }
1753 1761
1754 1762
1755 1763 /*
1756 1764 * Originally clconf_init() apparently needed the hostid. But
1757 1765 * this no longer appears to be true - it uses its own nodeid.
1758 1766 * By placing the hostid logic here, we are able to make use of
1759 1767 * the SMBIOS UUID.
1760 1768 */
1761 1769 if ((h = set_soft_hostid()) == HW_INVALID_HOSTID) {
1762 1770 cmn_err(CE_WARN, "Unable to set hostid");
1763 1771 } else {
1764 1772 for (v = h, cnt = 0; cnt < 10; cnt++) {
1765 1773 d[cnt] = (char)(v % 10);
1766 1774 v /= 10;
1767 1775 if (v == 0)
1768 1776 break;
1769 1777 }
1770 1778 for (cp = hw_serial; cnt >= 0; cnt--)
1771 1779 *cp++ = d[cnt] + '0';
1772 1780 *cp = 0;
1773 1781 }
1774 1782
1775 1783 /*
1776 1784 * Set up the CPU module subsystem for the boot cpu in the native
1777 1785 * case, and all physical cpu resource in the xpv dom0 case.
1778 1786 * Modifies the device tree, so this must be done after
1779 1787 * setup_ddi().
1780 1788 */
1781 1789 #ifdef __xpv
1782 1790 /*
1783 1791 * If paravirtualized and on dom0 then we initialize all physical
1784 1792 * cpu handles now; if paravirtualized on a domU then do not
1785 1793 * initialize.
1786 1794 */
1787 1795 if (DOMAIN_IS_INITDOMAIN(xen_info)) {
1788 1796 xen_mc_lcpu_cookie_t cpi;
1789 1797
1790 1798 for (cpi = xen_physcpu_next(NULL); cpi != NULL;
1791 1799 cpi = xen_physcpu_next(cpi)) {
1792 1800 if ((hdl = cmi_init(CMI_HDL_SOLARIS_xVM_MCA,
1793 1801 xen_physcpu_chipid(cpi), xen_physcpu_coreid(cpi),
1794 1802 xen_physcpu_strandid(cpi))) != NULL &&
1795 1803 is_x86_feature(x86_featureset, X86FSET_MCA))
1796 1804 cmi_mca_init(hdl);
1797 1805 }
1798 1806 }
1799 1807 #else
1800 1808 /*
1801 1809 * Initialize a handle for the boot cpu - others will initialize
1802 1810 * as they startup.
1803 1811 */
1804 1812 if ((hdl = cmi_init(CMI_HDL_NATIVE, cmi_ntv_hwchipid(CPU),
1805 1813 cmi_ntv_hwcoreid(CPU), cmi_ntv_hwstrandid(CPU))) != NULL) {
1806 1814 if (is_x86_feature(x86_featureset, X86FSET_MCA))
1807 1815 cmi_mca_init(hdl);
1808 1816 CPU->cpu_m.mcpu_cmi_hdl = hdl;
1809 1817 }
1810 1818 #endif /* __xpv */
1811 1819
1812 1820 /*
1813 1821 * Fake a prom tree such that /dev/openprom continues to work
1814 1822 */
1815 1823 PRM_POINT("startup_modules: calling prom_setup...");
1816 1824 prom_setup();
1817 1825 PRM_POINT("startup_modules: done");
1818 1826
1819 1827 /*
1820 1828 * Load all platform specific modules
1821 1829 */
1822 1830 PRM_POINT("startup_modules: calling psm_modload...");
1823 1831 psm_modload();
1824 1832
1825 1833 PRM_POINT("startup_modules() done");
1826 1834 }
1827 1835
1828 1836 /*
1829 1837 * claim a "setaside" boot page for use in the kernel
1830 1838 */
1831 1839 page_t *
1832 1840 boot_claim_page(pfn_t pfn)
1833 1841 {
1834 1842 page_t *pp;
1835 1843
1836 1844 pp = page_numtopp_nolock(pfn);
1837 1845 ASSERT(pp != NULL);
1838 1846
1839 1847 if (PP_ISBOOTPAGES(pp)) {
1840 1848 if (pp->p_next != NULL)
1841 1849 pp->p_next->p_prev = pp->p_prev;
1842 1850 if (pp->p_prev == NULL)
1843 1851 bootpages = pp->p_next;
1844 1852 else
1845 1853 pp->p_prev->p_next = pp->p_next;
1846 1854 } else {
1847 1855 /*
1848 1856 * htable_attach() expects a base pagesize page
1849 1857 */
1850 1858 if (pp->p_szc != 0)
1851 1859 page_boot_demote(pp);
1852 1860 pp = page_numtopp(pfn, SE_EXCL);
1853 1861 }
1854 1862 return (pp);
1855 1863 }
1856 1864
1857 1865 /*
1858 1866 * Walk through the pagetables looking for pages mapped in by boot. If the
1859 1867 * setaside flag is set the pages are expected to be returned to the
1860 1868 * kernel later in boot, so we add them to the bootpages list.
1861 1869 */
1862 1870 static void
1863 1871 protect_boot_range(uintptr_t low, uintptr_t high, int setaside)
1864 1872 {
1865 1873 uintptr_t va = low;
1866 1874 size_t len;
1867 1875 uint_t prot;
1868 1876 pfn_t pfn;
1869 1877 page_t *pp;
1870 1878 pgcnt_t boot_protect_cnt = 0;
1871 1879
1872 1880 while (kbm_probe(&va, &len, &pfn, &prot) != 0 && va < high) {
1873 1881 if (va + len >= high)
1874 1882 panic("0x%lx byte mapping at 0x%p exceeds boot's "
1875 1883 "legal range.", len, (void *)va);
1876 1884
1877 1885 while (len > 0) {
1878 1886 pp = page_numtopp_alloc(pfn);
1879 1887 if (pp != NULL) {
1880 1888 if (setaside == 0)
1881 1889 panic("Unexpected mapping by boot. "
1882 1890 "addr=%p pfn=%lx\n",
1883 1891 (void *)va, pfn);
1884 1892
1885 1893 pp->p_next = bootpages;
1886 1894 pp->p_prev = NULL;
1887 1895 PP_SETBOOTPAGES(pp);
1888 1896 if (bootpages != NULL) {
1889 1897 bootpages->p_prev = pp;
1890 1898 }
1891 1899 bootpages = pp;
1892 1900 ++boot_protect_cnt;
1893 1901 }
1894 1902
1895 1903 ++pfn;
1896 1904 len -= MMU_PAGESIZE;
1897 1905 va += MMU_PAGESIZE;
1898 1906 }
1899 1907 }
1900 1908 PRM_DEBUG(boot_protect_cnt);
1901 1909 }
1902 1910
1903 1911 /*
1904 1912 *
1905 1913 */
1906 1914 static void
1907 1915 layout_kernel_va(void)
1908 1916 {
1909 1917 PRM_POINT("layout_kernel_va() starting...");
1910 1918 /*
1911 1919 * Establish the final size of the kernel's heap, size of segmap,
1912 1920 * segkp, etc.
1913 1921 */
1914 1922
1915 1923 #if defined(__amd64)
1916 1924
1917 1925 kpm_vbase = (caddr_t)segkpm_base;
1918 1926 if (physmax + 1 < plat_dr_physmax) {
1919 1927 kpm_size = ROUND_UP_LPAGE(mmu_ptob(plat_dr_physmax));
1920 1928 } else {
1921 1929 kpm_size = ROUND_UP_LPAGE(mmu_ptob(physmax + 1));
1922 1930 }
1923 1931 if ((uintptr_t)kpm_vbase + kpm_size > (uintptr_t)valloc_base)
1924 1932 panic("not enough room for kpm!");
1925 1933 PRM_DEBUG(kpm_size);
1926 1934 PRM_DEBUG(kpm_vbase);
1927 1935
1928 1936 /*
1929 1937 * By default we create a seg_kp in 64 bit kernels, it's a little
1930 1938 * faster to access than embedding it in the heap.
1931 1939 */
1932 1940 segkp_base = (caddr_t)valloc_base + valloc_sz;
1933 1941 if (!segkp_fromheap) {
1934 1942 size_t sz = mmu_ptob(segkpsize);
1935 1943
1936 1944 /*
1937 1945 * determine size of segkp
1938 1946 */
1939 1947 if (sz < SEGKPMINSIZE || sz > SEGKPMAXSIZE) {
1940 1948 sz = SEGKPDEFSIZE;
1941 1949 cmn_err(CE_WARN, "!Illegal value for segkpsize. "
1942 1950 "segkpsize has been reset to %ld pages",
1943 1951 mmu_btop(sz));
1944 1952 }
1945 1953 sz = MIN(sz, MAX(SEGKPMINSIZE, mmu_ptob(physmem)));
1946 1954
1947 1955 segkpsize = mmu_btop(ROUND_UP_LPAGE(sz));
1948 1956 }
1949 1957 PRM_DEBUG(segkp_base);
1950 1958 PRM_DEBUG(segkpsize);
1951 1959
1952 1960 /*
1953 1961 * segzio is used for ZFS cached data. It uses a distinct VA
1954 1962 * segment (from kernel heap) so that we can easily tell not to
1955 1963 * include it in kernel crash dumps on 64 bit kernels. The trick is
1956 1964 * to give it lots of VA, but not constrain the kernel heap.
1957 1965 * We can use 1.5x physmem for segzio, leaving approximately
1958 1966 * another 1.5x physmem for heap. See also the comment in
1959 1967 * startup_memlist().
1960 1968 */
1961 1969 segzio_base = segkp_base + mmu_ptob(segkpsize);
1962 1970 if (segzio_fromheap) {
1963 1971 segziosize = 0;
1964 1972 } else {
1965 1973 size_t physmem_size = mmu_ptob(physmem);
1966 1974 size_t size = (segziosize == 0) ?
1967 1975 physmem_size * 3 / 2 : mmu_ptob(segziosize);
1968 1976
1969 1977 if (size < SEGZIOMINSIZE)
1970 1978 size = SEGZIOMINSIZE;
1971 1979 segziosize = mmu_btop(ROUND_UP_LPAGE(size));
1972 1980 }
1973 1981 PRM_DEBUG(segziosize);
1974 1982 PRM_DEBUG(segzio_base);
1975 1983
1976 1984 /*
1977 1985 * Put the range of VA for device mappings next, kmdb knows to not
1978 1986 * grep in this range of addresses.
1979 1987 */
1980 1988 toxic_addr =
1981 1989 ROUND_UP_LPAGE((uintptr_t)segzio_base + mmu_ptob(segziosize));
1982 1990 PRM_DEBUG(toxic_addr);
1983 1991 segmap_start = ROUND_UP_LPAGE(toxic_addr + toxic_size);
1984 1992 #else /* __i386 */
1985 1993 segmap_start = ROUND_UP_LPAGE(kernelbase);
1986 1994 #endif /* __i386 */
1987 1995 PRM_DEBUG(segmap_start);
1988 1996
1989 1997 /*
1990 1998 * Users can change segmapsize through eeprom. If the variable
1991 1999 * is tuned through eeprom, there is no upper bound on the
1992 2000 * size of segmap.
1993 2001 */
1994 2002 segmapsize = MAX(ROUND_UP_LPAGE(segmapsize), SEGMAPDEFAULT);
1995 2003
1996 2004 #if defined(__i386)
1997 2005 /*
1998 2006 * 32-bit systems don't have segkpm or segkp, so segmap appears at
1999 2007 * the bottom of the kernel's address range. Set aside space for a
2000 2008 * small red zone just below the start of segmap.
2001 2009 */
2002 2010 segmap_start += KERNEL_REDZONE_SIZE;
2003 2011 segmapsize -= KERNEL_REDZONE_SIZE;
2004 2012 #endif
2005 2013
2006 2014 PRM_DEBUG(segmap_start);
2007 2015 PRM_DEBUG(segmapsize);
2008 2016 kernelheap = (caddr_t)ROUND_UP_LPAGE(segmap_start + segmapsize);
2009 2017 PRM_DEBUG(kernelheap);
2010 2018 PRM_POINT("layout_kernel_va() done...");
2011 2019 }
2012 2020
2013 2021 /*
2014 2022 * Finish initializing the VM system, now that we are no longer
2015 2023 * relying on the boot time memory allocators.
2016 2024 */
2017 2025 static void
2018 2026 startup_vm(void)
2019 2027 {
2020 2028 struct segmap_crargs a;
2021 2029
2022 2030 extern int use_brk_lpg, use_stk_lpg;
2023 2031
2024 2032 PRM_POINT("startup_vm() starting...");
2025 2033
2026 2034 /*
2027 2035 * Initialize the hat layer.
2028 2036 */
2029 2037 hat_init();
2030 2038
2031 2039 /*
2032 2040 * Do final allocations of HAT data structures that need to
2033 2041 * be allocated before quiescing the boot loader.
2034 2042 */
2035 2043 PRM_POINT("Calling hat_kern_alloc()...");
2036 2044 hat_kern_alloc((caddr_t)segmap_start, segmapsize, ekernelheap);
2037 2045 PRM_POINT("hat_kern_alloc() done");
2038 2046
2039 2047 #ifndef __xpv
2040 2048 /*
2041 2049 * Setup Page Attribute Table
2042 2050 */
2043 2051 pat_sync();
2044 2052 #endif
2045 2053
2046 2054 /*
2047 2055 * The next two loops are done in distinct steps in order
2048 2056 * to be sure that any page that is doubly mapped (both above
2049 2057 * KERNEL_TEXT and below kernelbase) is dealt with correctly.
2050 2058 * Note this may never happen, but it might someday.
2051 2059 */
2052 2060 bootpages = NULL;
2053 2061 PRM_POINT("Protecting boot pages");
2054 2062
2055 2063 /*
2056 2064 * Protect any pages mapped above KERNEL_TEXT that somehow have
2057 2065 * page_t's. This can only happen if something weird allocated
2058 2066 * in this range (like kadb/kmdb).
2059 2067 */
2060 2068 protect_boot_range(KERNEL_TEXT, (uintptr_t)-1, 0);
2061 2069
2062 2070 /*
2063 2071 * Before we can take over memory allocation/mapping from the boot
2064 2072 * loader we must remove from our free page lists any boot allocated
2065 2073 * pages that stay mapped until release_bootstrap().
2066 2074 */
2067 2075 protect_boot_range(0, kernelbase, 1);
2068 2076
2069 2077
2070 2078 /*
2071 2079 * Switch to running on regular HAT (not boot_mmu)
2072 2080 */
2073 2081 PRM_POINT("Calling hat_kern_setup()...");
2074 2082 hat_kern_setup();
2075 2083
2076 2084 /*
2077 2085 * It is no longer safe to call BOP_ALLOC(), so make sure we don't.
2078 2086 */
2079 2087 bop_no_more_mem();
2080 2088
2081 2089 PRM_POINT("hat_kern_setup() done");
2082 2090
2083 2091 hat_cpu_online(CPU);
2084 2092
2085 2093 /*
2086 2094 * Initialize VM system
2087 2095 */
2088 2096 PRM_POINT("Calling kvm_init()...");
2089 2097 kvm_init();
2090 2098 PRM_POINT("kvm_init() done");
2091 2099
2092 2100 /*
2093 2101 * Tell kmdb that the VM system is now working
2094 2102 */
2095 2103 if (boothowto & RB_DEBUG)
2096 2104 kdi_dvec_vmready();
2097 2105
2098 2106 #if defined(__xpv)
2099 2107 /*
2100 2108 * Populate the I/O pool on domain 0
2101 2109 */
2102 2110 if (DOMAIN_IS_INITDOMAIN(xen_info)) {
2103 2111 extern long populate_io_pool(void);
2104 2112 long init_io_pool_cnt;
2105 2113
2106 2114 PRM_POINT("Populating reserve I/O page pool");
2107 2115 init_io_pool_cnt = populate_io_pool();
2108 2116 PRM_DEBUG(init_io_pool_cnt);
2109 2117 }
2110 2118 #endif
2111 2119 /*
2112 2120 * Mangle the brand string etc.
2113 2121 */
2114 2122 cpuid_pass3(CPU);
2115 2123
2116 2124 #if defined(__amd64)
2117 2125
2118 2126 /*
2119 2127 * Create the device arena for toxic (to dtrace/kmdb) mappings.
2120 2128 */
2121 2129 device_arena = vmem_create("device", (void *)toxic_addr,
2122 2130 toxic_size, MMU_PAGESIZE, NULL, NULL, NULL, 0, VM_SLEEP);
2123 2131
2124 2132 #else /* __i386 */
2125 2133
2126 2134 /*
2127 2135 * allocate the bit map that tracks toxic pages
2128 2136 */
2129 2137 toxic_bit_map_len = btop((ulong_t)(valloc_base - kernelbase));
2130 2138 PRM_DEBUG(toxic_bit_map_len);
2131 2139 toxic_bit_map =
2132 2140 kmem_zalloc(BT_SIZEOFMAP(toxic_bit_map_len), KM_NOSLEEP);
2133 2141 ASSERT(toxic_bit_map != NULL);
2134 2142 PRM_DEBUG(toxic_bit_map);
2135 2143
2136 2144 #endif /* __i386 */
2137 2145
2138 2146
2139 2147 /*
2140 2148 * Now that we've got more VA, as well as the ability to allocate from
2141 2149 * it, tell the debugger.
2142 2150 */
2143 2151 if (boothowto & RB_DEBUG)
2144 2152 kdi_dvec_memavail();
2145 2153
2146 2154 #if !defined(__xpv)
2147 2155 /*
2148 2156 * Map page pfn=0 for drivers, such as kd, that need to pick up
2149 2157 * parameters left there by controllers/BIOS.
2150 2158 */
2151 2159 PRM_POINT("setup up p0_va");
2152 2160 p0_va = i86devmap(0, 1, PROT_READ);
2153 2161 PRM_DEBUG(p0_va);
2154 2162 #endif
2155 2163
2156 2164 cmn_err(CE_CONT, "?mem = %luK (0x%lx)\n",
2157 2165 physinstalled << (MMU_PAGESHIFT - 10), ptob(physinstalled));
2158 2166
2159 2167 /*
2160 2168 * disable automatic large pages for small memory systems or
2161 2169 * when the disable flag is set.
2162 2170 *
2163 2171 * Do not yet consider page sizes larger than 2m/4m.
2164 2172 */
2165 2173 if (!auto_lpg_disable && mmu.max_page_level > 0) {
2166 2174 max_uheap_lpsize = LEVEL_SIZE(1);
2167 2175 max_ustack_lpsize = LEVEL_SIZE(1);
2168 2176 max_privmap_lpsize = LEVEL_SIZE(1);
2169 2177 max_uidata_lpsize = LEVEL_SIZE(1);
2170 2178 max_utext_lpsize = LEVEL_SIZE(1);
2171 2179 max_shm_lpsize = LEVEL_SIZE(1);
2172 2180 }
2173 2181 if (physmem < privm_lpg_min_physmem || mmu.max_page_level == 0 ||
2174 2182 auto_lpg_disable) {
2175 2183 use_brk_lpg = 0;
2176 2184 use_stk_lpg = 0;
2177 2185 }
2178 2186 mcntl0_lpsize = LEVEL_SIZE(mmu.umax_page_level);
2179 2187
2180 2188 PRM_POINT("Calling hat_init_finish()...");
2181 2189 hat_init_finish();
2182 2190 PRM_POINT("hat_init_finish() done");
2183 2191
2184 2192 /*
2185 2193 * Initialize the segkp segment type.
2186 2194 */
2187 2195 rw_enter(&kas.a_lock, RW_WRITER);
2188 2196 PRM_POINT("Attaching segkp");
2189 2197 if (segkp_fromheap) {
2190 2198 segkp->s_as = &kas;
2191 2199 } else if (seg_attach(&kas, (caddr_t)segkp_base, mmu_ptob(segkpsize),
2192 2200 segkp) < 0) {
2193 2201 panic("startup: cannot attach segkp");
2194 2202 /*NOTREACHED*/
2195 2203 }
2196 2204 PRM_POINT("Doing segkp_create()");
2197 2205 if (segkp_create(segkp) != 0) {
2198 2206 panic("startup: segkp_create failed");
2199 2207 /*NOTREACHED*/
2200 2208 }
2201 2209 PRM_DEBUG(segkp);
2202 2210 rw_exit(&kas.a_lock);
2203 2211
2204 2212 /*
2205 2213 * kpm segment
2206 2214 */
2207 2215 segmap_kpm = 0;
2208 2216 if (kpm_desired)
2209 2217 kpm_init();
2210 2218
2211 2219 /*
2212 2220 * Now create segmap segment.
2213 2221 */
2214 2222 rw_enter(&kas.a_lock, RW_WRITER);
2215 2223 if (seg_attach(&kas, (caddr_t)segmap_start, segmapsize, segmap) < 0) {
2216 2224 panic("cannot attach segmap");
2217 2225 /*NOTREACHED*/
2218 2226 }
2219 2227 PRM_DEBUG(segmap);
2220 2228
2221 2229 a.prot = PROT_READ | PROT_WRITE;
2222 2230 a.shmsize = 0;
2223 2231 a.nfreelist = segmapfreelists;
2224 2232
2225 2233 if (segmap_create(segmap, (caddr_t)&a) != 0)
2226 2234 panic("segmap_create segmap");
2227 2235 rw_exit(&kas.a_lock);
2228 2236
2229 2237 setup_vaddr_for_ppcopy(CPU);
2230 2238
2231 2239 segdev_init();
2232 2240 #if defined(__xpv)
2233 2241 if (DOMAIN_IS_INITDOMAIN(xen_info))
2234 2242 #endif
2235 2243 pmem_init();
2236 2244
2237 2245 PRM_POINT("startup_vm() done");
2238 2246 }
2239 2247
2240 2248 /*
2241 2249 * Load a tod module for the non-standard tod part found on this system.
2242 2250 */
2243 2251 static void
2244 2252 load_tod_module(char *todmod)
2245 2253 {
2246 2254 if (modload("tod", todmod) == -1)
2247 2255 halt("Can't load TOD module");
2248 2256 }
2249 2257
2250 2258 static void
2251 2259 startup_end(void)
2252 2260 {
2253 2261 int i;
2254 2262 extern void setx86isalist(void);
2255 2263 extern void cpu_event_init(void);
2256 2264
2257 2265 PRM_POINT("startup_end() starting...");
2258 2266
2259 2267 /*
2260 2268 * Perform tasks that get done after most of the VM
2261 2269 * initialization has been done but before the clock
2262 2270 * and other devices get started.
2263 2271 */
2264 2272 kern_setup1();
2265 2273
2266 2274 /*
2267 2275 * Perform CPC initialization for this CPU.
2268 2276 */
2269 2277 kcpc_hw_init(CPU);
2270 2278
2271 2279 /*
2272 2280 * Initialize cpu event framework.
2273 2281 */
2274 2282 cpu_event_init();
2275 2283
2276 2284 #if defined(OPTERON_WORKAROUND_6323525)
2277 2285 if (opteron_workaround_6323525)
2278 2286 patch_workaround_6323525();
2279 2287 #endif
2280 2288 /*
2281 2289 * If needed, load TOD module now so that ddi_get_time(9F) etc. work
2282 2290 * (For now, "needed" is defined as set tod_module_name in /etc/system)
2283 2291 */
2284 2292 if (tod_module_name != NULL) {
2285 2293 PRM_POINT("load_tod_module()");
2286 2294 load_tod_module(tod_module_name);
2287 2295 }
2288 2296
2289 2297 #if defined(__xpv)
2290 2298 /*
2291 2299 * Forceload interposing TOD module for the hypervisor.
2292 2300 */
2293 2301 PRM_POINT("load_tod_module()");
2294 2302 load_tod_module("xpvtod");
2295 2303 #endif
2296 2304
2297 2305 /*
2298 2306 * Configure the system.
2299 2307 */
2300 2308 PRM_POINT("Calling configure()...");
2301 2309 configure(); /* set up devices */
2302 2310 PRM_POINT("configure() done");
2303 2311
2304 2312 /*
2305 2313 * We can now setup for XSAVE because fpu_probe is done in configure().
2306 2314 */
2307 2315 if (fp_save_mech == FP_XSAVE) {
2308 2316 xsave_setup_msr(CPU);
2309 2317 }
2310 2318
2311 2319 /*
2312 2320 * Set the isa_list string to the defined instruction sets we
2313 2321 * support.
2314 2322 */
2315 2323 setx86isalist();
2316 2324 cpu_intr_alloc(CPU, NINTR_THREADS);
2317 2325 psm_install();
2318 2326
2319 2327 /*
2320 2328 * We're done with bootops. We don't unmap the bootstrap yet because
2321 2329 * we're still using bootsvcs.
2322 2330 */
2323 2331 PRM_POINT("NULLing out bootops");
2324 2332 *bootopsp = (struct bootops *)NULL;
2325 2333 bootops = (struct bootops *)NULL;
2326 2334
2327 2335 #if defined(__xpv)
2328 2336 ec_init_debug_irq();
2329 2337 xs_domu_init();
2330 2338 #endif
2331 2339
2332 2340 #if !defined(__xpv)
2333 2341 /*
2334 2342 * Intel IOMMU has been setup/initialized in ddi_impl.c
2335 2343 * Start it up now.
2336 2344 */
2337 2345 immu_startup();
2338 2346
2339 2347 /*
2340 2348 * Now that we're no longer going to drop into real mode for a BIOS call
2341 2349 * via bootops, we can enable PCID (which requires CR0.PG).
2342 2350 */
2343 2351 enable_pcid();
2344 2352 #endif
2345 2353
2346 2354 PRM_POINT("Enabling interrupts");
2347 2355 (*picinitf)();
2348 2356 sti();
2349 2357 #if defined(__xpv)
2350 2358 ASSERT(CPU->cpu_m.mcpu_vcpu_info->evtchn_upcall_mask == 0);
2351 2359 xen_late_startup();
2352 2360 #endif
2353 2361
2354 2362 (void) add_avsoftintr((void *)&softlevel1_hdl, 1, softlevel1,
2355 2363 "softlevel1", NULL, NULL); /* XXX to be moved later */
2356 2364
2357 2365 /*
2358 2366 * Register software interrupt handlers for ddi_periodic_add(9F).
2359 2367 * Software interrupts up to the level 10 are supported.
2360 2368 */
2361 2369 for (i = DDI_IPL_1; i <= DDI_IPL_10; i++) {
2362 2370 (void) add_avsoftintr((void *)&softlevel_hdl[i-1], i,
2363 2371 (avfunc)(uintptr_t)ddi_periodic_softintr, "ddi_periodic",
2364 2372 (caddr_t)(uintptr_t)i, NULL);
2365 2373 }
2366 2374
2367 2375 #if !defined(__xpv)
2368 2376 if (modload("drv", "amd_iommu") < 0) {
2369 2377 PRM_POINT("No AMD IOMMU present\n");
2370 2378 } else if (ddi_hold_installed_driver(ddi_name_to_major(
2371 2379 "amd_iommu")) == NULL) {
2372 2380 prom_printf("ERROR: failed to attach AMD IOMMU\n");
2373 2381 }
2374 2382 #endif
2375 2383 post_startup_cpu_fixups();
2376 2384
2377 2385 PRM_POINT("startup_end() done");
2378 2386 }
2379 2387
2380 2388 /*
2381 2389 * Don't remove the following 2 variables. They are necessary
2382 2390 * for reading the hostid from the legacy file (/kernel/misc/sysinit).
2383 2391 */
2384 2392 char *_hs1107 = hw_serial;
2385 2393 ulong_t _bdhs34;
2386 2394
2387 2395 void
2388 2396 post_startup(void)
2389 2397 {
2390 2398 extern void cpupm_init(cpu_t *);
2391 2399 extern void cpu_event_init_cpu(cpu_t *);
2392 2400
2393 2401 /*
2394 2402 * Set the system wide, processor-specific flags to be passed
2395 2403 * to userland via the aux vector for performance hints and
2396 2404 * instruction set extensions.
2397 2405 */
2398 2406 bind_hwcap();
2399 2407
2400 2408 #ifdef __xpv
2401 2409 if (DOMAIN_IS_INITDOMAIN(xen_info))
2402 2410 #endif
2403 2411 {
2404 2412 #if defined(__xpv)
2405 2413 xpv_panic_init();
2406 2414 #else
2407 2415 /*
2408 2416 * Startup the memory scrubber.
2409 2417 * XXPV This should be running somewhere ..
2410 2418 */
2411 2419 if ((get_hwenv() & HW_VIRTUAL) == 0)
2412 2420 memscrub_init();
2413 2421 #endif
2414 2422 }
2415 2423
2416 2424 /*
2417 2425 * Complete CPU module initialization
2418 2426 */
2419 2427 cmi_post_startup();
2420 2428
2421 2429 /*
2422 2430 * Perform forceloading tasks for /etc/system.
2423 2431 */
2424 2432 (void) mod_sysctl(SYS_FORCELOAD, NULL);
2425 2433
2426 2434 /*
2427 2435 * ON4.0: Force /proc module in until clock interrupt handle fixed
2428 2436 * ON4.0: This must be fixed or restated in /etc/systems.
2429 2437 */
2430 2438 (void) modload("fs", "procfs");
2431 2439
2432 2440 (void) i_ddi_attach_hw_nodes("pit_beep");
2433 2441
2434 2442 #if defined(__i386)
2435 2443 /*
2436 2444 * Check for required functional Floating Point hardware,
2437 2445 * unless FP hardware explicitly disabled.
2438 2446 */
2439 2447 if (fpu_exists && (fpu_pentium_fdivbug || fp_kind == FP_NO))
2440 2448 halt("No working FP hardware found");
2441 2449 #endif
2442 2450
2443 2451 maxmem = freemem;
2444 2452
2445 2453 cpu_event_init_cpu(CPU);
2446 2454 cpupm_init(CPU);
2447 2455 (void) mach_cpu_create_device_node(CPU, NULL);
2448 2456
2449 2457 pg_init();
2450 2458 }
2451 2459
2452 2460 static int
2453 2461 pp_in_range(page_t *pp, uint64_t low_addr, uint64_t high_addr)
2454 2462 {
2455 2463 return ((pp->p_pagenum >= btop(low_addr)) &&
2456 2464 (pp->p_pagenum < btopr(high_addr)));
2457 2465 }
2458 2466
2459 2467 static int
2460 2468 pp_in_module(page_t *pp, const rd_existing_t *modranges)
2461 2469 {
2462 2470 uint_t i;
2463 2471
2464 2472 for (i = 0; modranges[i].phys != 0; i++) {
2465 2473 if (pp_in_range(pp, modranges[i].phys,
2466 2474 modranges[i].phys + modranges[i].size))
2467 2475 return (1);
2468 2476 }
2469 2477
2470 2478 return (0);
2471 2479 }
2472 2480
2473 2481 void
2474 2482 release_bootstrap(void)
2475 2483 {
2476 2484 int root_is_ramdisk;
2477 2485 page_t *pp;
2478 2486 extern void kobj_boot_unmountroot(void);
2479 2487 extern dev_t rootdev;
2480 2488 uint_t i;
2481 2489 char propname[32];
2482 2490 rd_existing_t *modranges;
2483 2491 #if !defined(__xpv)
2484 2492 pfn_t pfn;
2485 2493 #endif
2486 2494
2487 2495 /*
2488 2496 * Save the bootfs module ranges so that we can reserve them below
2489 2497 * for the real bootfs.
2490 2498 */
2491 2499 modranges = kmem_alloc(sizeof (rd_existing_t) * MAX_BOOT_MODULES,
2492 2500 KM_SLEEP);
2493 2501 for (i = 0; ; i++) {
2494 2502 uint64_t start, size;
2495 2503
2496 2504 modranges[i].phys = 0;
2497 2505
2498 2506 (void) snprintf(propname, sizeof (propname),
2499 2507 "module-addr-%u", i);
2500 2508 if (do_bsys_getproplen(NULL, propname) <= 0)
2501 2509 break;
2502 2510 (void) do_bsys_getprop(NULL, propname, &start);
2503 2511
2504 2512 (void) snprintf(propname, sizeof (propname),
2505 2513 "module-size-%u", i);
2506 2514 if (do_bsys_getproplen(NULL, propname) <= 0)
2507 2515 break;
2508 2516 (void) do_bsys_getprop(NULL, propname, &size);
2509 2517
2510 2518 modranges[i].phys = start;
2511 2519 modranges[i].size = size;
2512 2520 }
2513 2521
2514 2522 /* unmount boot ramdisk and release kmem usage */
2515 2523 kobj_boot_unmountroot();
2516 2524
2517 2525 /*
2518 2526 * We're finished using the boot loader so free its pages.
2519 2527 */
2520 2528 PRM_POINT("Unmapping lower boot pages");
2521 2529
2522 2530 clear_boot_mappings(0, _userlimit);
2523 2531
2524 2532 postbootkernelbase = kernelbase;
2525 2533
2526 2534 /*
2527 2535 * If root isn't on ramdisk, destroy the hardcoded
2528 2536 * ramdisk node now and release the memory. Else,
2529 2537 * ramdisk memory is kept in rd_pages.
2530 2538 */
2531 2539 root_is_ramdisk = (getmajor(rootdev) == ddi_name_to_major("ramdisk"));
2532 2540 if (!root_is_ramdisk) {
2533 2541 dev_info_t *dip = ddi_find_devinfo("ramdisk", -1, 0);
2534 2542 ASSERT(dip && ddi_get_parent(dip) == ddi_root_node());
2535 2543 ndi_rele_devi(dip); /* held from ddi_find_devinfo */
2536 2544 (void) ddi_remove_child(dip, 0);
2537 2545 }
2538 2546
2539 2547 PRM_POINT("Releasing boot pages");
2540 2548 while (bootpages) {
2541 2549 extern uint64_t ramdisk_start, ramdisk_end;
2542 2550 pp = bootpages;
2543 2551 bootpages = pp->p_next;
2544 2552
2545 2553
2546 2554 /* Keep pages for the lower 64K */
2547 2555 if (pp_in_range(pp, 0, 0x40000)) {
2548 2556 pp->p_next = lower_pages;
2549 2557 lower_pages = pp;
2550 2558 lower_pages_count++;
2551 2559 continue;
2552 2560 }
2553 2561
2554 2562 if (root_is_ramdisk && pp_in_range(pp, ramdisk_start,
2555 2563 ramdisk_end) || pp_in_module(pp, modranges)) {
2556 2564 pp->p_next = rd_pages;
2557 2565 rd_pages = pp;
2558 2566 continue;
2559 2567 }
2560 2568 pp->p_next = (struct page *)0;
2561 2569 pp->p_prev = (struct page *)0;
2562 2570 PP_CLRBOOTPAGES(pp);
2563 2571 page_free(pp, 1);
2564 2572 }
2565 2573 PRM_POINT("Boot pages released");
2566 2574
2567 2575 kmem_free(modranges, sizeof (rd_existing_t) * 99);
2568 2576
2569 2577 #if !defined(__xpv)
2570 2578 /* XXPV -- note this following bunch of code needs to be revisited in Xen 3.0 */
2571 2579 /*
2572 2580 * Find 1 page below 1 MB so that other processors can boot up or
2573 2581 * so that any processor can resume.
2574 2582 * Make sure it has a kernel VA as well as a 1:1 mapping.
2575 2583 * We should have just free'd one up.
2576 2584 */
2577 2585
2578 2586 /*
2579 2587 * 0x10 pages is 64K. Leave the bottom 64K alone
2580 2588 * for BIOS.
2581 2589 */
2582 2590 for (pfn = 0x10; pfn < btop(1*1024*1024); pfn++) {
2583 2591 if (page_numtopp_alloc(pfn) == NULL)
2584 2592 continue;
2585 2593 rm_platter_va = i86devmap(pfn, 1,
2586 2594 PROT_READ | PROT_WRITE | PROT_EXEC);
2587 2595 rm_platter_pa = ptob(pfn);
2588 2596 break;
2589 2597 }
2590 2598 if (pfn == btop(1*1024*1024) && use_mp)
2591 2599 panic("No page below 1M available for starting "
2592 2600 "other processors or for resuming from system-suspend");
2593 2601 #endif /* !__xpv */
2594 2602 }
2595 2603
2596 2604 /*
2597 2605 * Initialize the platform-specific parts of a page_t.
2598 2606 */
2599 2607 void
2600 2608 add_physmem_cb(page_t *pp, pfn_t pnum)
2601 2609 {
2602 2610 pp->p_pagenum = pnum;
2603 2611 pp->p_mapping = NULL;
2604 2612 pp->p_embed = 0;
2605 2613 pp->p_share = 0;
2606 2614 pp->p_mlentry = 0;
2607 2615 }
2608 2616
2609 2617 /*
2610 2618 * kphysm_init() initializes physical memory.
2611 2619 */
2612 2620 static pgcnt_t
2613 2621 kphysm_init(page_t *pp, pgcnt_t npages)
2614 2622 {
2615 2623 struct memlist *pmem;
2616 2624 struct memseg *cur_memseg;
2617 2625 pfn_t base_pfn;
2618 2626 pfn_t end_pfn;
2619 2627 pgcnt_t num;
2620 2628 pgcnt_t pages_done = 0;
2621 2629 uint64_t addr;
2622 2630 uint64_t size;
2623 2631 extern pfn_t ddiphysmin;
2624 2632 extern int mnode_xwa;
2625 2633 int ms = 0, me = 0;
2626 2634
2627 2635 ASSERT(page_hash != NULL && page_hashsz != 0);
2628 2636
2629 2637 cur_memseg = memseg_base;
2630 2638 for (pmem = phys_avail; pmem && npages; pmem = pmem->ml_next) {
2631 2639 /*
2632 2640 * In a 32 bit kernel can't use higher memory if we're
2633 2641 * not booting in PAE mode. This check takes care of that.
2634 2642 */
2635 2643 addr = pmem->ml_address;
2636 2644 size = pmem->ml_size;
2637 2645 if (btop(addr) > physmax)
2638 2646 continue;
2639 2647
2640 2648 /*
2641 2649 * align addr and size - they may not be at page boundaries
2642 2650 */
2643 2651 if ((addr & MMU_PAGEOFFSET) != 0) {
2644 2652 addr += MMU_PAGEOFFSET;
2645 2653 addr &= ~(uint64_t)MMU_PAGEOFFSET;
2646 2654 size -= addr - pmem->ml_address;
2647 2655 }
2648 2656
2649 2657 /* only process pages below or equal to physmax */
2650 2658 if ((btop(addr + size) - 1) > physmax)
2651 2659 size = ptob(physmax - btop(addr) + 1);
2652 2660
2653 2661 num = btop(size);
2654 2662 if (num == 0)
2655 2663 continue;
2656 2664
2657 2665 if (num > npages)
2658 2666 num = npages;
2659 2667
2660 2668 npages -= num;
2661 2669 pages_done += num;
2662 2670 base_pfn = btop(addr);
2663 2671
2664 2672 if (prom_debug)
2665 2673 prom_printf("MEMSEG addr=0x%" PRIx64
2666 2674 " pgs=0x%lx pfn 0x%lx-0x%lx\n",
2667 2675 addr, num, base_pfn, base_pfn + num);
2668 2676
2669 2677 /*
2670 2678 * Ignore pages below ddiphysmin to simplify ddi memory
2671 2679 * allocation with non-zero addr_lo requests.
2672 2680 */
2673 2681 if (base_pfn < ddiphysmin) {
2674 2682 if (base_pfn + num <= ddiphysmin)
2675 2683 continue;
2676 2684 pp += (ddiphysmin - base_pfn);
2677 2685 num -= (ddiphysmin - base_pfn);
2678 2686 base_pfn = ddiphysmin;
2679 2687 }
2680 2688
2681 2689 /*
2682 2690 * mnode_xwa is greater than 1 when large pages regions can
2683 2691 * cross memory node boundaries. To prevent the formation
2684 2692 * of these large pages, configure the memsegs based on the
2685 2693 * memory node ranges which had been made non-contiguous.
2686 2694 */
2687 2695 end_pfn = base_pfn + num - 1;
2688 2696 if (mnode_xwa > 1) {
2689 2697 ms = PFN_2_MEM_NODE(base_pfn);
2690 2698 me = PFN_2_MEM_NODE(end_pfn);
2691 2699
2692 2700 if (ms != me) {
2693 2701 /*
2694 2702 * current range spans more than 1 memory node.
2695 2703 * Set num to only the pfn range in the start
2696 2704 * memory node.
2697 2705 */
2698 2706 num = mem_node_config[ms].physmax - base_pfn
2699 2707 + 1;
2700 2708 ASSERT(end_pfn > mem_node_config[ms].physmax);
2701 2709 }
2702 2710 }
2703 2711
2704 2712 for (;;) {
2705 2713 /*
2706 2714 * Build the memsegs entry
2707 2715 */
2708 2716 cur_memseg->pages = pp;
2709 2717 cur_memseg->epages = pp + num;
2710 2718 cur_memseg->pages_base = base_pfn;
2711 2719 cur_memseg->pages_end = base_pfn + num;
2712 2720
2713 2721 /*
2714 2722 * Insert into memseg list in decreasing pfn range
2715 2723 * order. Low memory is typically more fragmented such
2716 2724 * that this ordering keeps the larger ranges at the
2717 2725 * front of the list for code that searches memseg.
2718 2726 * This ASSERTS that the memsegs coming in from boot
2719 2727 * are in increasing physical address order and not
2720 2728 * contiguous.
2721 2729 */
2722 2730 if (memsegs != NULL) {
2723 2731 ASSERT(cur_memseg->pages_base >=
2724 2732 memsegs->pages_end);
2725 2733 cur_memseg->next = memsegs;
2726 2734 }
2727 2735 memsegs = cur_memseg;
2728 2736
2729 2737 /*
2730 2738 * add_physmem() initializes the PSM part of the page
2731 2739 * struct by calling the PSM back with add_physmem_cb().
2732 2740 * In addition it coalesces pages into larger pages as
2733 2741 * it initializes them.
2734 2742 */
2735 2743 add_physmem(pp, num, base_pfn);
2736 2744 cur_memseg++;
2737 2745 availrmem_initial += num;
2738 2746 availrmem += num;
2739 2747
2740 2748 pp += num;
2741 2749 if (ms >= me)
2742 2750 break;
2743 2751
2744 2752 /* process next memory node range */
2745 2753 ms++;
2746 2754 base_pfn = mem_node_config[ms].physbase;
2747 2755
2748 2756 if (mnode_xwa > 1) {
2749 2757 num = MIN(mem_node_config[ms].physmax,
2750 2758 end_pfn) - base_pfn + 1;
2751 2759 } else {
2752 2760 num = mem_node_config[ms].physmax -
2753 2761 base_pfn + 1;
2754 2762 }
2755 2763 }
2756 2764 }
2757 2765
2758 2766 PRM_DEBUG(availrmem_initial);
2759 2767 PRM_DEBUG(availrmem);
2760 2768 PRM_DEBUG(freemem);
2761 2769 build_pfn_hash();
2762 2770 return (pages_done);
2763 2771 }
2764 2772
2765 2773 /*
2766 2774 * Kernel VM initialization.
2767 2775 */
2768 2776 static void
2769 2777 kvm_init(void)
2770 2778 {
2771 2779 ASSERT((((uintptr_t)s_text) & MMU_PAGEOFFSET) == 0);
2772 2780
2773 2781 /*
2774 2782 * Put the kernel segments in kernel address space.
2775 2783 */
2776 2784 rw_enter(&kas.a_lock, RW_WRITER);
2777 2785 as_avlinit(&kas);
2778 2786
2779 2787 (void) seg_attach(&kas, s_text, e_moddata - s_text, &ktextseg);
2780 2788 (void) segkmem_create(&ktextseg);
2781 2789
2782 2790 (void) seg_attach(&kas, (caddr_t)valloc_base, valloc_sz, &kvalloc);
2783 2791 (void) segkmem_create(&kvalloc);
2784 2792
2785 2793 (void) seg_attach(&kas, kernelheap,
2786 2794 ekernelheap - kernelheap, &kvseg);
2787 2795 (void) segkmem_create(&kvseg);
2788 2796
2789 2797 if (core_size > 0) {
2790 2798 PRM_POINT("attaching kvseg_core");
2791 2799 (void) seg_attach(&kas, (caddr_t)core_base, core_size,
2792 2800 &kvseg_core);
2793 2801 (void) segkmem_create(&kvseg_core);
2794 2802 }
2795 2803
2796 2804 if (segziosize > 0) {
2797 2805 PRM_POINT("attaching segzio");
2798 2806 (void) seg_attach(&kas, segzio_base, mmu_ptob(segziosize),
2799 2807 &kzioseg);
2800 2808 (void) segkmem_zio_create(&kzioseg);
2801 2809
2802 2810 /* create zio area covering new segment */
2803 2811 segkmem_zio_init(segzio_base, mmu_ptob(segziosize));
2804 2812 }
2805 2813
2806 2814 (void) seg_attach(&kas, kdi_segdebugbase, kdi_segdebugsize, &kdebugseg);
2807 2815 (void) segkmem_create(&kdebugseg);
2808 2816
2809 2817 rw_exit(&kas.a_lock);
2810 2818
2811 2819 /*
2812 2820 * Ensure that the red zone at kernelbase is never accessible.
2813 2821 */
2814 2822 PRM_POINT("protecting redzone");
2815 2823 (void) as_setprot(&kas, (caddr_t)kernelbase, KERNEL_REDZONE_SIZE, 0);
2816 2824
2817 2825 /*
2818 2826 * Make the text writable so that it can be hot patched by DTrace.
2819 2827 */
2820 2828 (void) as_setprot(&kas, s_text, e_modtext - s_text,
2821 2829 PROT_READ | PROT_WRITE | PROT_EXEC);
2822 2830
2823 2831 /*
2824 2832 * Make data writable until end.
2825 2833 */
2826 2834 (void) as_setprot(&kas, s_data, e_moddata - s_data,
2827 2835 PROT_READ | PROT_WRITE | PROT_EXEC);
2828 2836 }
2829 2837
2830 2838 #ifndef __xpv
2831 2839 /*
2832 2840 * Solaris adds an entry for Write Combining caching to the PAT
2833 2841 */
2834 2842 static uint64_t pat_attr_reg = PAT_DEFAULT_ATTRIBUTE;
2835 2843
2836 2844 void
2837 2845 pat_sync(void)
2838 2846 {
2839 2847 ulong_t cr0, cr0_orig, cr4;
2840 2848
2841 2849 if (!is_x86_feature(x86_featureset, X86FSET_PAT))
2842 2850 return;
2843 2851 cr0_orig = cr0 = getcr0();
2844 2852 cr4 = getcr4();
2845 2853
2846 2854 /* disable caching and flush all caches and TLBs */
2847 2855 cr0 |= CR0_CD;
2848 2856 cr0 &= ~CR0_NW;
2849 2857 setcr0(cr0);
2850 2858 invalidate_cache();
2851 2859 if (cr4 & CR4_PGE) {
2852 2860 setcr4(cr4 & ~(ulong_t)CR4_PGE);
2853 2861 setcr4(cr4);
2854 2862 } else {
2855 2863 reload_cr3();
2856 2864 }
2857 2865
2858 2866 /* add our entry to the PAT */
2859 2867 wrmsr(REG_PAT, pat_attr_reg);
2860 2868
2861 2869 /* flush TLBs and cache again, then reenable cr0 caching */
2862 2870 if (cr4 & CR4_PGE) {
2863 2871 setcr4(cr4 & ~(ulong_t)CR4_PGE);
2864 2872 setcr4(cr4);
2865 2873 } else {
2866 2874 reload_cr3();
2867 2875 }
2868 2876 invalidate_cache();
2869 2877 setcr0(cr0_orig);
2870 2878 }
2871 2879
2872 2880 #endif /* !__xpv */
2873 2881
2874 2882 #if defined(_SOFT_HOSTID)
2875 2883 /*
2876 2884 * On platforms that do not have a hardware serial number, attempt
2877 2885 * to set one based on the contents of /etc/hostid. If this file does
2878 2886 * not exist, assume that we are to generate a new hostid and set
2879 2887 * it in the kernel, for subsequent saving by a userland process
2880 2888 * once the system is up and the root filesystem is mounted r/w.
2881 2889 *
2882 2890 * In order to gracefully support upgrade on OpenSolaris, if
2883 2891 * /etc/hostid does not exist, we will attempt to get a serial number
2884 2892 * using the legacy method (/kernel/misc/sysinit).
2885 2893 *
2886 2894 * If that isn't present, we attempt to use an SMBIOS UUID, which is
2887 2895 * a hardware serial number. Note that we don't automatically trust
2888 2896 * all SMBIOS UUIDs (some older platforms are defective and ship duplicate
2889 2897 * UUIDs in violation of the standard), we check against a blacklist.
2890 2898 *
2891 2899 * In an attempt to make the hostid less prone to abuse
2892 2900 * (for license circumvention, etc), we store it in /etc/hostid
2893 2901 * in rot47 format.
2894 2902 */
2895 2903 extern volatile unsigned long tenmicrodata;
2896 2904 static int atoi(char *);
2897 2905
2898 2906 /*
2899 2907 * Set this to non-zero in /etc/system if you think your SMBIOS returns a
2900 2908 * UUID that is not unique. (Also report it so that the smbios_uuid_blacklist
2901 2909 * array can be updated.)
2902 2910 */
2903 2911 int smbios_broken_uuid = 0;
2904 2912
2905 2913 /*
2906 2914 * List of known bad UUIDs. This is just the lower 32-bit values, since
2907 2915 * that's what we use for the host id. If your hostid falls here, you need
2908 2916 * to contact your hardware OEM for a fix for your BIOS.
2909 2917 */
2910 2918 static unsigned char
2911 2919 smbios_uuid_blacklist[][16] = {
2912 2920
2913 2921 { /* Reported bad UUID (Google search) */
2914 2922 0x00, 0x02, 0x00, 0x03, 0x00, 0x04, 0x00, 0x05,
2915 2923 0x00, 0x06, 0x00, 0x07, 0x00, 0x08, 0x00, 0x09,
2916 2924 },
2917 2925 { /* Known bad DELL UUID */
2918 2926 0x4C, 0x4C, 0x45, 0x44, 0x00, 0x00, 0x20, 0x10,
2919 2927 0x80, 0x20, 0x80, 0xC0, 0x4F, 0x20, 0x20, 0x20,
2920 2928 },
2921 2929 { /* Uninitialized flash */
2922 2930 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff,
2923 2931 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff
2924 2932 },
2925 2933 { /* All zeros */
2926 2934 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
2927 2935 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00
2928 2936 },
2929 2937 };
2930 2938
2931 2939 static int32_t
2932 2940 uuid_to_hostid(const uint8_t *uuid)
2933 2941 {
2934 2942 /*
2935 2943 * Although the UUIDs are 128-bits, they may not distribute entropy
2936 2944 * evenly. We would like to use SHA or MD5, but those are located
2937 2945 * in loadable modules and not available this early in boot. As we
2938 2946 * don't need the values to be cryptographically strong, we just
2939 2947 * generate 32-bit vaue by xor'ing the various sequences together,
2940 2948 * which ensures that the entire UUID contributes to the hostid.
2941 2949 */
2942 2950 uint32_t id = 0;
2943 2951
2944 2952 /* first check against the blacklist */
2945 2953 for (int i = 0; i < (sizeof (smbios_uuid_blacklist) / 16); i++) {
2946 2954 if (bcmp(smbios_uuid_blacklist[0], uuid, 16) == 0) {
2947 2955 cmn_err(CE_CONT, "?Broken SMBIOS UUID. "
2948 2956 "Contact BIOS manufacturer for repair.\n");
2949 2957 return ((int32_t)HW_INVALID_HOSTID);
2950 2958 }
2951 2959 }
2952 2960
2953 2961 for (int i = 0; i < 16; i++)
2954 2962 id ^= ((uuid[i]) << (8 * (i % sizeof (id))));
2955 2963
2956 2964 /* Make sure return value is positive */
2957 2965 return (id & 0x7fffffff);
2958 2966 }
2959 2967
2960 2968 static int32_t
2961 2969 set_soft_hostid(void)
2962 2970 {
2963 2971 struct _buf *file;
2964 2972 char tokbuf[MAXNAMELEN];
2965 2973 token_t token;
2966 2974 int done = 0;
2967 2975 u_longlong_t tmp;
2968 2976 int i;
2969 2977 int32_t hostid = (int32_t)HW_INVALID_HOSTID;
2970 2978 unsigned char *c;
2971 2979 hrtime_t tsc;
2972 2980 smbios_system_t smsys;
2973 2981
2974 2982 /*
2975 2983 * If /etc/hostid file not found, we'd like to get a pseudo
2976 2984 * random number to use at the hostid. A nice way to do this
2977 2985 * is to read the real time clock. To remain xen-compatible,
2978 2986 * we can't poke the real hardware, so we use tsc_read() to
2979 2987 * read the real time clock. However, there is an ominous
2980 2988 * warning in tsc_read that says it can return zero, so we
2981 2989 * deal with that possibility by falling back to using the
2982 2990 * (hopefully random enough) value in tenmicrodata.
2983 2991 */
2984 2992
2985 2993 if ((file = kobj_open_file(hostid_file)) == (struct _buf *)-1) {
2986 2994 /*
2987 2995 * hostid file not found - try to load sysinit module
2988 2996 * and see if it has a nonzero hostid value...use that
2989 2997 * instead of generating a new hostid here if so.
2990 2998 */
2991 2999 if ((i = modload("misc", "sysinit")) != -1) {
2992 3000 if (strlen(hw_serial) > 0)
2993 3001 hostid = (int32_t)atoi(hw_serial);
2994 3002 (void) modunload(i);
2995 3003 }
2996 3004
2997 3005 /*
2998 3006 * We try to use the SMBIOS UUID. But not if it is blacklisted
2999 3007 * in /etc/system.
3000 3008 */
3001 3009 if ((hostid == HW_INVALID_HOSTID) &&
3002 3010 (smbios_broken_uuid == 0) &&
3003 3011 (ksmbios != NULL) &&
3004 3012 (smbios_info_system(ksmbios, &smsys) != SMB_ERR) &&
3005 3013 (smsys.smbs_uuidlen >= 16)) {
3006 3014 hostid = uuid_to_hostid(smsys.smbs_uuid);
3007 3015 }
3008 3016
3009 3017 /*
3010 3018 * Generate a "random" hostid using the clock. These
3011 3019 * hostids will change on each boot if the value is not
3012 3020 * saved to a persistent /etc/hostid file.
3013 3021 */
3014 3022 if (hostid == HW_INVALID_HOSTID) {
3015 3023 tsc = tsc_read();
3016 3024 if (tsc == 0) /* tsc_read can return zero sometimes */
3017 3025 hostid = (int32_t)tenmicrodata & 0x0CFFFFF;
3018 3026 else
3019 3027 hostid = (int32_t)tsc & 0x0CFFFFF;
3020 3028 }
3021 3029 } else {
3022 3030 /* hostid file found */
3023 3031 while (!done) {
3024 3032 token = kobj_lex(file, tokbuf, sizeof (tokbuf));
3025 3033
3026 3034 switch (token) {
3027 3035 case POUND:
3028 3036 /*
3029 3037 * skip comments
3030 3038 */
3031 3039 kobj_find_eol(file);
3032 3040 break;
3033 3041 case STRING:
3034 3042 /*
3035 3043 * un-rot47 - obviously this
3036 3044 * nonsense is ascii-specific
3037 3045 */
3038 3046 for (c = (unsigned char *)tokbuf;
3039 3047 *c != '\0'; c++) {
3040 3048 *c += 47;
3041 3049 if (*c > '~')
3042 3050 *c -= 94;
3043 3051 else if (*c < '!')
3044 3052 *c += 94;
3045 3053 }
3046 3054 /*
3047 3055 * now we should have a real number
3048 3056 */
3049 3057
3050 3058 if (kobj_getvalue(tokbuf, &tmp) != 0)
3051 3059 kobj_file_err(CE_WARN, file,
3052 3060 "Bad value %s for hostid",
3053 3061 tokbuf);
3054 3062 else
3055 3063 hostid = (int32_t)tmp;
3056 3064
3057 3065 break;
3058 3066 case EOF:
3059 3067 done = 1;
3060 3068 /* FALLTHROUGH */
3061 3069 case NEWLINE:
3062 3070 kobj_newline(file);
3063 3071 break;
3064 3072 default:
3065 3073 break;
3066 3074
3067 3075 }
3068 3076 }
3069 3077 if (hostid == HW_INVALID_HOSTID) /* didn't find a hostid */
3070 3078 kobj_file_err(CE_WARN, file,
3071 3079 "hostid missing or corrupt");
3072 3080
3073 3081 kobj_close_file(file);
3074 3082 }
3075 3083 /*
3076 3084 * hostid is now the value read from /etc/hostid, or the
3077 3085 * new hostid we generated in this routine or HW_INVALID_HOSTID if not
3078 3086 * set.
3079 3087 */
3080 3088 return (hostid);
3081 3089 }
3082 3090
3083 3091 static int
3084 3092 atoi(char *p)
3085 3093 {
3086 3094 int i = 0;
3087 3095
3088 3096 while (*p != '\0')
3089 3097 i = 10 * i + (*p++ - '0');
3090 3098
3091 3099 return (i);
3092 3100 }
3093 3101
3094 3102 #endif /* _SOFT_HOSTID */
3095 3103
3096 3104 void
3097 3105 get_system_configuration(void)
3098 3106 {
3099 3107 char prop[32];
3100 3108 u_longlong_t nodes_ll, cpus_pernode_ll, lvalue;
3101 3109
3102 3110 if (BOP_GETPROPLEN(bootops, "nodes") > sizeof (prop) ||
3103 3111 BOP_GETPROP(bootops, "nodes", prop) < 0 ||
3104 3112 kobj_getvalue(prop, &nodes_ll) == -1 ||
3105 3113 nodes_ll > MAXNODES ||
3106 3114 BOP_GETPROPLEN(bootops, "cpus_pernode") > sizeof (prop) ||
3107 3115 BOP_GETPROP(bootops, "cpus_pernode", prop) < 0 ||
3108 3116 kobj_getvalue(prop, &cpus_pernode_ll) == -1) {
3109 3117 system_hardware.hd_nodes = 1;
3110 3118 system_hardware.hd_cpus_per_node = 0;
3111 3119 } else {
3112 3120 system_hardware.hd_nodes = (int)nodes_ll;
3113 3121 system_hardware.hd_cpus_per_node = (int)cpus_pernode_ll;
3114 3122 }
3115 3123
3116 3124 if (BOP_GETPROPLEN(bootops, "kernelbase") > sizeof (prop) ||
3117 3125 BOP_GETPROP(bootops, "kernelbase", prop) < 0 ||
3118 3126 kobj_getvalue(prop, &lvalue) == -1)
3119 3127 eprom_kernelbase = 0;
3120 3128 else
3121 3129 eprom_kernelbase = (uintptr_t)lvalue;
3122 3130
3123 3131 if (BOP_GETPROPLEN(bootops, "segmapsize") > sizeof (prop) ||
3124 3132 BOP_GETPROP(bootops, "segmapsize", prop) < 0 ||
3125 3133 kobj_getvalue(prop, &lvalue) == -1)
3126 3134 segmapsize = SEGMAPDEFAULT;
3127 3135 else
3128 3136 segmapsize = (uintptr_t)lvalue;
3129 3137
3130 3138 if (BOP_GETPROPLEN(bootops, "segmapfreelists") > sizeof (prop) ||
3131 3139 BOP_GETPROP(bootops, "segmapfreelists", prop) < 0 ||
3132 3140 kobj_getvalue(prop, &lvalue) == -1)
3133 3141 segmapfreelists = 0; /* use segmap driver default */
3134 3142 else
3135 3143 segmapfreelists = (int)lvalue;
3136 3144
3137 3145 /* physmem used to be here, but moved much earlier to fakebop.c */
3138 3146 }
3139 3147
3140 3148 /*
3141 3149 * Add to a memory list.
3142 3150 * start = start of new memory segment
3143 3151 * len = length of new memory segment in bytes
3144 3152 * new = pointer to a new struct memlist
3145 3153 * memlistp = memory list to which to add segment.
3146 3154 */
3147 3155 void
3148 3156 memlist_add(
3149 3157 uint64_t start,
3150 3158 uint64_t len,
3151 3159 struct memlist *new,
3152 3160 struct memlist **memlistp)
3153 3161 {
3154 3162 struct memlist *cur;
3155 3163 uint64_t end = start + len;
3156 3164
3157 3165 new->ml_address = start;
3158 3166 new->ml_size = len;
3159 3167
3160 3168 cur = *memlistp;
3161 3169
3162 3170 while (cur) {
3163 3171 if (cur->ml_address >= end) {
3164 3172 new->ml_next = cur;
3165 3173 *memlistp = new;
3166 3174 new->ml_prev = cur->ml_prev;
3167 3175 cur->ml_prev = new;
3168 3176 return;
3169 3177 }
3170 3178 ASSERT(cur->ml_address + cur->ml_size <= start);
3171 3179 if (cur->ml_next == NULL) {
3172 3180 cur->ml_next = new;
3173 3181 new->ml_prev = cur;
3174 3182 new->ml_next = NULL;
3175 3183 return;
3176 3184 }
3177 3185 memlistp = &cur->ml_next;
3178 3186 cur = cur->ml_next;
3179 3187 }
3180 3188 }
3181 3189
3182 3190 void
3183 3191 kobj_vmem_init(vmem_t **text_arena, vmem_t **data_arena)
3184 3192 {
3185 3193 size_t tsize = e_modtext - modtext;
3186 3194 size_t dsize = e_moddata - moddata;
3187 3195
3188 3196 *text_arena = vmem_create("module_text", tsize ? modtext : NULL, tsize,
3189 3197 1, segkmem_alloc, segkmem_free, heaptext_arena, 0, VM_SLEEP);
3190 3198 *data_arena = vmem_create("module_data", dsize ? moddata : NULL, dsize,
3191 3199 1, segkmem_alloc, segkmem_free, heap32_arena, 0, VM_SLEEP);
3192 3200 }
3193 3201
3194 3202 caddr_t
3195 3203 kobj_text_alloc(vmem_t *arena, size_t size)
3196 3204 {
3197 3205 return (vmem_alloc(arena, size, VM_SLEEP | VM_BESTFIT));
3198 3206 }
3199 3207
3200 3208 /*ARGSUSED*/
3201 3209 caddr_t
3202 3210 kobj_texthole_alloc(caddr_t addr, size_t size)
3203 3211 {
3204 3212 panic("unexpected call to kobj_texthole_alloc()");
3205 3213 /*NOTREACHED*/
3206 3214 return (0);
3207 3215 }
3208 3216
3209 3217 /*ARGSUSED*/
3210 3218 void
3211 3219 kobj_texthole_free(caddr_t addr, size_t size)
3212 3220 {
3213 3221 panic("unexpected call to kobj_texthole_free()");
3214 3222 }
3215 3223
3216 3224 /*
3217 3225 * This is called just after configure() in startup().
3218 3226 *
3219 3227 * The ISALIST concept is a bit hopeless on Intel, because
3220 3228 * there's no guarantee of an ever-more-capable processor
3221 3229 * given that various parts of the instruction set may appear
3222 3230 * and disappear between different implementations.
3223 3231 *
3224 3232 * While it would be possible to correct it and even enhance
3225 3233 * it somewhat, the explicit hardware capability bitmask allows
3226 3234 * more flexibility.
3227 3235 *
3228 3236 * So, we just leave this alone.
3229 3237 */
3230 3238 void
3231 3239 setx86isalist(void)
3232 3240 {
3233 3241 char *tp;
3234 3242 size_t len;
3235 3243 extern char *isa_list;
3236 3244
3237 3245 #define TBUFSIZE 1024
3238 3246
3239 3247 tp = kmem_alloc(TBUFSIZE, KM_SLEEP);
3240 3248 *tp = '\0';
3241 3249
3242 3250 #if defined(__amd64)
3243 3251 (void) strcpy(tp, "amd64 ");
3244 3252 #endif
3245 3253
3246 3254 switch (x86_vendor) {
3247 3255 case X86_VENDOR_Intel:
3248 3256 case X86_VENDOR_AMD:
3249 3257 case X86_VENDOR_TM:
3250 3258 if (is_x86_feature(x86_featureset, X86FSET_CMOV)) {
3251 3259 /*
3252 3260 * Pentium Pro or later
3253 3261 */
3254 3262 (void) strcat(tp, "pentium_pro");
3255 3263 (void) strcat(tp,
3256 3264 is_x86_feature(x86_featureset, X86FSET_MMX) ?
3257 3265 "+mmx pentium_pro " : " ");
3258 3266 }
3259 3267 /*FALLTHROUGH*/
3260 3268 case X86_VENDOR_Cyrix:
3261 3269 /*
3262 3270 * The Cyrix 6x86 does not have any Pentium features
3263 3271 * accessible while not at privilege level 0.
3264 3272 */
3265 3273 if (is_x86_feature(x86_featureset, X86FSET_CPUID)) {
3266 3274 (void) strcat(tp, "pentium");
3267 3275 (void) strcat(tp,
3268 3276 is_x86_feature(x86_featureset, X86FSET_MMX) ?
3269 3277 "+mmx pentium " : " ");
3270 3278 }
3271 3279 break;
3272 3280 default:
3273 3281 break;
3274 3282 }
3275 3283 (void) strcat(tp, "i486 i386 i86");
3276 3284 len = strlen(tp) + 1; /* account for NULL at end of string */
3277 3285 isa_list = strcpy(kmem_alloc(len, KM_SLEEP), tp);
3278 3286 kmem_free(tp, TBUFSIZE);
3279 3287
3280 3288 #undef TBUFSIZE
3281 3289 }
3282 3290
3283 3291
3284 3292 #ifdef __amd64
3285 3293
3286 3294 void *
3287 3295 device_arena_alloc(size_t size, int vm_flag)
3288 3296 {
3289 3297 return (vmem_alloc(device_arena, size, vm_flag));
3290 3298 }
3291 3299
3292 3300 void
3293 3301 device_arena_free(void *vaddr, size_t size)
3294 3302 {
3295 3303 vmem_free(device_arena, vaddr, size);
3296 3304 }
3297 3305
3298 3306 #else /* __i386 */
3299 3307
3300 3308 void *
3301 3309 device_arena_alloc(size_t size, int vm_flag)
3302 3310 {
3303 3311 caddr_t vaddr;
3304 3312 uintptr_t v;
3305 3313 size_t start;
3306 3314 size_t end;
3307 3315
3308 3316 vaddr = vmem_alloc(heap_arena, size, vm_flag);
3309 3317 if (vaddr == NULL)
3310 3318 return (NULL);
3311 3319
3312 3320 v = (uintptr_t)vaddr;
3313 3321 ASSERT(v >= kernelbase);
3314 3322 ASSERT(v + size <= valloc_base);
3315 3323
3316 3324 start = btop(v - kernelbase);
3317 3325 end = btop(v + size - 1 - kernelbase);
3318 3326 ASSERT(start < toxic_bit_map_len);
3319 3327 ASSERT(end < toxic_bit_map_len);
3320 3328
3321 3329 while (start <= end) {
3322 3330 BT_ATOMIC_SET(toxic_bit_map, start);
3323 3331 ++start;
3324 3332 }
3325 3333 return (vaddr);
3326 3334 }
3327 3335
3328 3336 void
3329 3337 device_arena_free(void *vaddr, size_t size)
3330 3338 {
3331 3339 uintptr_t v = (uintptr_t)vaddr;
3332 3340 size_t start;
3333 3341 size_t end;
3334 3342
3335 3343 ASSERT(v >= kernelbase);
3336 3344 ASSERT(v + size <= valloc_base);
3337 3345
3338 3346 start = btop(v - kernelbase);
3339 3347 end = btop(v + size - 1 - kernelbase);
3340 3348 ASSERT(start < toxic_bit_map_len);
3341 3349 ASSERT(end < toxic_bit_map_len);
3342 3350
3343 3351 while (start <= end) {
3344 3352 ASSERT(BT_TEST(toxic_bit_map, start) != 0);
3345 3353 BT_ATOMIC_CLEAR(toxic_bit_map, start);
3346 3354 ++start;
3347 3355 }
3348 3356 vmem_free(heap_arena, vaddr, size);
3349 3357 }
3350 3358
3351 3359 /*
3352 3360 * returns 1st address in range that is in device arena, or NULL
3353 3361 * if len is not NULL it returns the length of the toxic range
3354 3362 */
3355 3363 void *
3356 3364 device_arena_contains(void *vaddr, size_t size, size_t *len)
3357 3365 {
3358 3366 uintptr_t v = (uintptr_t)vaddr;
3359 3367 uintptr_t eaddr = v + size;
3360 3368 size_t start;
3361 3369 size_t end;
3362 3370
3363 3371 /*
3364 3372 * if called very early by kmdb, just return NULL
3365 3373 */
3366 3374 if (toxic_bit_map == NULL)
3367 3375 return (NULL);
3368 3376
3369 3377 /*
3370 3378 * First check if we're completely outside the bitmap range.
3371 3379 */
3372 3380 if (v >= valloc_base || eaddr < kernelbase)
3373 3381 return (NULL);
3374 3382
3375 3383 /*
3376 3384 * Trim ends of search to look at only what the bitmap covers.
3377 3385 */
3378 3386 if (v < kernelbase)
3379 3387 v = kernelbase;
3380 3388 start = btop(v - kernelbase);
3381 3389 end = btop(eaddr - kernelbase);
3382 3390 if (end >= toxic_bit_map_len)
3383 3391 end = toxic_bit_map_len;
3384 3392
3385 3393 if (bt_range(toxic_bit_map, &start, &end, end) == 0)
3386 3394 return (NULL);
3387 3395
3388 3396 v = kernelbase + ptob(start);
3389 3397 if (len != NULL)
3390 3398 *len = ptob(end - start);
3391 3399 return ((void *)v);
3392 3400 }
3393 3401
3394 3402 #endif /* __i386 */
|
↓ open down ↓ |
2650 lines elided |
↑ open up ↑ |
XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX