xref: /openbmc/linux/arch/ia64/kernel/setup.c (revision 1fa6ac37)
1 /*
2  * Architecture-specific setup.
3  *
4  * Copyright (C) 1998-2001, 2003-2004 Hewlett-Packard Co
5  *	David Mosberger-Tang <davidm@hpl.hp.com>
6  *	Stephane Eranian <eranian@hpl.hp.com>
7  * Copyright (C) 2000, 2004 Intel Corp
8  * 	Rohit Seth <rohit.seth@intel.com>
9  * 	Suresh Siddha <suresh.b.siddha@intel.com>
10  * 	Gordon Jin <gordon.jin@intel.com>
11  * Copyright (C) 1999 VA Linux Systems
12  * Copyright (C) 1999 Walt Drummond <drummond@valinux.com>
13  *
14  * 12/26/04 S.Siddha, G.Jin, R.Seth
15  *			Add multi-threading and multi-core detection
16  * 11/12/01 D.Mosberger Convert get_cpuinfo() to seq_file based show_cpuinfo().
17  * 04/04/00 D.Mosberger renamed cpu_initialized to cpu_online_map
18  * 03/31/00 R.Seth	cpu_initialized and current->processor fixes
19  * 02/04/00 D.Mosberger	some more get_cpuinfo fixes...
20  * 02/01/00 R.Seth	fixed get_cpuinfo for SMP
21  * 01/07/99 S.Eranian	added the support for command line argument
22  * 06/24/99 W.Drummond	added boot_cpu_data.
23  * 05/28/05 Z. Menyhart	Dynamic stride size for "flush_icache_range()"
24  */
25 #include <linux/module.h>
26 #include <linux/init.h>
27 
28 #include <linux/acpi.h>
29 #include <linux/bootmem.h>
30 #include <linux/console.h>
31 #include <linux/delay.h>
32 #include <linux/kernel.h>
33 #include <linux/reboot.h>
34 #include <linux/sched.h>
35 #include <linux/seq_file.h>
36 #include <linux/string.h>
37 #include <linux/threads.h>
38 #include <linux/screen_info.h>
39 #include <linux/dmi.h>
40 #include <linux/serial.h>
41 #include <linux/serial_core.h>
42 #include <linux/efi.h>
43 #include <linux/initrd.h>
44 #include <linux/pm.h>
45 #include <linux/cpufreq.h>
46 #include <linux/kexec.h>
47 #include <linux/crash_dump.h>
48 
49 #include <asm/machvec.h>
50 #include <asm/mca.h>
51 #include <asm/meminit.h>
52 #include <asm/page.h>
53 #include <asm/paravirt.h>
54 #include <asm/paravirt_patch.h>
55 #include <asm/patch.h>
56 #include <asm/pgtable.h>
57 #include <asm/processor.h>
58 #include <asm/sal.h>
59 #include <asm/sections.h>
60 #include <asm/setup.h>
61 #include <asm/smp.h>
62 #include <asm/system.h>
63 #include <asm/tlbflush.h>
64 #include <asm/unistd.h>
65 #include <asm/hpsim.h>
66 
67 #if defined(CONFIG_SMP) && (IA64_CPU_SIZE > PAGE_SIZE)
68 # error "struct cpuinfo_ia64 too big!"
69 #endif
70 
71 #ifdef CONFIG_SMP
72 unsigned long __per_cpu_offset[NR_CPUS];
73 EXPORT_SYMBOL(__per_cpu_offset);
74 #endif
75 
76 DEFINE_PER_CPU(struct cpuinfo_ia64, ia64_cpu_info);
77 DEFINE_PER_CPU(unsigned long, local_per_cpu_offset);
78 unsigned long ia64_cycles_per_usec;
79 struct ia64_boot_param *ia64_boot_param;
80 struct screen_info screen_info;
81 unsigned long vga_console_iobase;
82 unsigned long vga_console_membase;
83 
84 static struct resource data_resource = {
85 	.name	= "Kernel data",
86 	.flags	= IORESOURCE_BUSY | IORESOURCE_MEM
87 };
88 
89 static struct resource code_resource = {
90 	.name	= "Kernel code",
91 	.flags	= IORESOURCE_BUSY | IORESOURCE_MEM
92 };
93 
94 static struct resource bss_resource = {
95 	.name	= "Kernel bss",
96 	.flags	= IORESOURCE_BUSY | IORESOURCE_MEM
97 };
98 
99 unsigned long ia64_max_cacheline_size;
100 
101 int dma_get_cache_alignment(void)
102 {
103         return ia64_max_cacheline_size;
104 }
105 EXPORT_SYMBOL(dma_get_cache_alignment);
106 
107 unsigned long ia64_iobase;	/* virtual address for I/O accesses */
108 EXPORT_SYMBOL(ia64_iobase);
109 struct io_space io_space[MAX_IO_SPACES];
110 EXPORT_SYMBOL(io_space);
111 unsigned int num_io_spaces;
112 
113 /*
114  * "flush_icache_range()" needs to know what processor dependent stride size to use
115  * when it makes i-cache(s) coherent with d-caches.
116  */
117 #define	I_CACHE_STRIDE_SHIFT	5	/* Safest way to go: 32 bytes by 32 bytes */
118 unsigned long ia64_i_cache_stride_shift = ~0;
119 /*
120  * "clflush_cache_range()" needs to know what processor dependent stride size to
121  * use when it flushes cache lines including both d-cache and i-cache.
122  */
123 /* Safest way to go: 32 bytes by 32 bytes */
124 #define	CACHE_STRIDE_SHIFT	5
125 unsigned long ia64_cache_stride_shift = ~0;
126 
127 /*
128  * The merge_mask variable needs to be set to (max(iommu_page_size(iommu)) - 1).  This
129  * mask specifies a mask of address bits that must be 0 in order for two buffers to be
130  * mergeable by the I/O MMU (i.e., the end address of the first buffer and the start
131  * address of the second buffer must be aligned to (merge_mask+1) in order to be
132  * mergeable).  By default, we assume there is no I/O MMU which can merge physically
133  * discontiguous buffers, so we set the merge_mask to ~0UL, which corresponds to a iommu
134  * page-size of 2^64.
135  */
136 unsigned long ia64_max_iommu_merge_mask = ~0UL;
137 EXPORT_SYMBOL(ia64_max_iommu_merge_mask);
138 
139 /*
140  * We use a special marker for the end of memory and it uses the extra (+1) slot
141  */
142 struct rsvd_region rsvd_region[IA64_MAX_RSVD_REGIONS + 1] __initdata;
143 int num_rsvd_regions __initdata;
144 
145 
146 /*
147  * Filter incoming memory segments based on the primitive map created from the boot
148  * parameters. Segments contained in the map are removed from the memory ranges. A
149  * caller-specified function is called with the memory ranges that remain after filtering.
150  * This routine does not assume the incoming segments are sorted.
151  */
152 int __init
153 filter_rsvd_memory (u64 start, u64 end, void *arg)
154 {
155 	u64 range_start, range_end, prev_start;
156 	void (*func)(unsigned long, unsigned long, int);
157 	int i;
158 
159 #if IGNORE_PFN0
160 	if (start == PAGE_OFFSET) {
161 		printk(KERN_WARNING "warning: skipping physical page 0\n");
162 		start += PAGE_SIZE;
163 		if (start >= end) return 0;
164 	}
165 #endif
166 	/*
167 	 * lowest possible address(walker uses virtual)
168 	 */
169 	prev_start = PAGE_OFFSET;
170 	func = arg;
171 
172 	for (i = 0; i < num_rsvd_regions; ++i) {
173 		range_start = max(start, prev_start);
174 		range_end   = min(end, rsvd_region[i].start);
175 
176 		if (range_start < range_end)
177 			call_pernode_memory(__pa(range_start), range_end - range_start, func);
178 
179 		/* nothing more available in this segment */
180 		if (range_end == end) return 0;
181 
182 		prev_start = rsvd_region[i].end;
183 	}
184 	/* end of memory marker allows full processing inside loop body */
185 	return 0;
186 }
187 
188 /*
189  * Similar to "filter_rsvd_memory()", but the reserved memory ranges
190  * are not filtered out.
191  */
192 int __init
193 filter_memory(u64 start, u64 end, void *arg)
194 {
195 	void (*func)(unsigned long, unsigned long, int);
196 
197 #if IGNORE_PFN0
198 	if (start == PAGE_OFFSET) {
199 		printk(KERN_WARNING "warning: skipping physical page 0\n");
200 		start += PAGE_SIZE;
201 		if (start >= end)
202 			return 0;
203 	}
204 #endif
205 	func = arg;
206 	if (start < end)
207 		call_pernode_memory(__pa(start), end - start, func);
208 	return 0;
209 }
210 
211 static void __init
212 sort_regions (struct rsvd_region *rsvd_region, int max)
213 {
214 	int j;
215 
216 	/* simple bubble sorting */
217 	while (max--) {
218 		for (j = 0; j < max; ++j) {
219 			if (rsvd_region[j].start > rsvd_region[j+1].start) {
220 				struct rsvd_region tmp;
221 				tmp = rsvd_region[j];
222 				rsvd_region[j] = rsvd_region[j + 1];
223 				rsvd_region[j + 1] = tmp;
224 			}
225 		}
226 	}
227 }
228 
229 /*
230  * Request address space for all standard resources
231  */
232 static int __init register_memory(void)
233 {
234 	code_resource.start = ia64_tpa(_text);
235 	code_resource.end   = ia64_tpa(_etext) - 1;
236 	data_resource.start = ia64_tpa(_etext);
237 	data_resource.end   = ia64_tpa(_edata) - 1;
238 	bss_resource.start  = ia64_tpa(__bss_start);
239 	bss_resource.end    = ia64_tpa(_end) - 1;
240 	efi_initialize_iomem_resources(&code_resource, &data_resource,
241 			&bss_resource);
242 
243 	return 0;
244 }
245 
246 __initcall(register_memory);
247 
248 
249 #ifdef CONFIG_KEXEC
250 
251 /*
252  * This function checks if the reserved crashkernel is allowed on the specific
253  * IA64 machine flavour. Machines without an IO TLB use swiotlb and require
254  * some memory below 4 GB (i.e. in 32 bit area), see the implementation of
255  * lib/swiotlb.c. The hpzx1 architecture has an IO TLB but cannot use that
256  * in kdump case. See the comment in sba_init() in sba_iommu.c.
257  *
258  * So, the only machvec that really supports loading the kdump kernel
259  * over 4 GB is "sn2".
260  */
261 static int __init check_crashkernel_memory(unsigned long pbase, size_t size)
262 {
263 	if (ia64_platform_is("sn2") || ia64_platform_is("uv"))
264 		return 1;
265 	else
266 		return pbase < (1UL << 32);
267 }
268 
269 static void __init setup_crashkernel(unsigned long total, int *n)
270 {
271 	unsigned long long base = 0, size = 0;
272 	int ret;
273 
274 	ret = parse_crashkernel(boot_command_line, total,
275 			&size, &base);
276 	if (ret == 0 && size > 0) {
277 		if (!base) {
278 			sort_regions(rsvd_region, *n);
279 			base = kdump_find_rsvd_region(size,
280 					rsvd_region, *n);
281 		}
282 
283 		if (!check_crashkernel_memory(base, size)) {
284 			pr_warning("crashkernel: There would be kdump memory "
285 				"at %ld GB but this is unusable because it "
286 				"must\nbe below 4 GB. Change the memory "
287 				"configuration of the machine.\n",
288 				(unsigned long)(base >> 30));
289 			return;
290 		}
291 
292 		if (base != ~0UL) {
293 			printk(KERN_INFO "Reserving %ldMB of memory at %ldMB "
294 					"for crashkernel (System RAM: %ldMB)\n",
295 					(unsigned long)(size >> 20),
296 					(unsigned long)(base >> 20),
297 					(unsigned long)(total >> 20));
298 			rsvd_region[*n].start =
299 				(unsigned long)__va(base);
300 			rsvd_region[*n].end =
301 				(unsigned long)__va(base + size);
302 			(*n)++;
303 			crashk_res.start = base;
304 			crashk_res.end = base + size - 1;
305 		}
306 	}
307 	efi_memmap_res.start = ia64_boot_param->efi_memmap;
308 	efi_memmap_res.end = efi_memmap_res.start +
309 		ia64_boot_param->efi_memmap_size;
310 	boot_param_res.start = __pa(ia64_boot_param);
311 	boot_param_res.end = boot_param_res.start +
312 		sizeof(*ia64_boot_param);
313 }
314 #else
315 static inline void __init setup_crashkernel(unsigned long total, int *n)
316 {}
317 #endif
318 
319 /**
320  * reserve_memory - setup reserved memory areas
321  *
322  * Setup the reserved memory areas set aside for the boot parameters,
323  * initrd, etc.  There are currently %IA64_MAX_RSVD_REGIONS defined,
324  * see arch/ia64/include/asm/meminit.h if you need to define more.
325  */
326 void __init
327 reserve_memory (void)
328 {
329 	int n = 0;
330 	unsigned long total_memory;
331 
332 	/*
333 	 * none of the entries in this table overlap
334 	 */
335 	rsvd_region[n].start = (unsigned long) ia64_boot_param;
336 	rsvd_region[n].end   = rsvd_region[n].start + sizeof(*ia64_boot_param);
337 	n++;
338 
339 	rsvd_region[n].start = (unsigned long) __va(ia64_boot_param->efi_memmap);
340 	rsvd_region[n].end   = rsvd_region[n].start + ia64_boot_param->efi_memmap_size;
341 	n++;
342 
343 	rsvd_region[n].start = (unsigned long) __va(ia64_boot_param->command_line);
344 	rsvd_region[n].end   = (rsvd_region[n].start
345 				+ strlen(__va(ia64_boot_param->command_line)) + 1);
346 	n++;
347 
348 	rsvd_region[n].start = (unsigned long) ia64_imva((void *)KERNEL_START);
349 	rsvd_region[n].end   = (unsigned long) ia64_imva(_end);
350 	n++;
351 
352 	n += paravirt_reserve_memory(&rsvd_region[n]);
353 
354 #ifdef CONFIG_BLK_DEV_INITRD
355 	if (ia64_boot_param->initrd_start) {
356 		rsvd_region[n].start = (unsigned long)__va(ia64_boot_param->initrd_start);
357 		rsvd_region[n].end   = rsvd_region[n].start + ia64_boot_param->initrd_size;
358 		n++;
359 	}
360 #endif
361 
362 #ifdef CONFIG_CRASH_DUMP
363 	if (reserve_elfcorehdr(&rsvd_region[n].start,
364 			       &rsvd_region[n].end) == 0)
365 		n++;
366 #endif
367 
368 	total_memory = efi_memmap_init(&rsvd_region[n].start, &rsvd_region[n].end);
369 	n++;
370 
371 	setup_crashkernel(total_memory, &n);
372 
373 	/* end of memory marker */
374 	rsvd_region[n].start = ~0UL;
375 	rsvd_region[n].end   = ~0UL;
376 	n++;
377 
378 	num_rsvd_regions = n;
379 	BUG_ON(IA64_MAX_RSVD_REGIONS + 1 < n);
380 
381 	sort_regions(rsvd_region, num_rsvd_regions);
382 }
383 
384 
385 /**
386  * find_initrd - get initrd parameters from the boot parameter structure
387  *
388  * Grab the initrd start and end from the boot parameter struct given us by
389  * the boot loader.
390  */
391 void __init
392 find_initrd (void)
393 {
394 #ifdef CONFIG_BLK_DEV_INITRD
395 	if (ia64_boot_param->initrd_start) {
396 		initrd_start = (unsigned long)__va(ia64_boot_param->initrd_start);
397 		initrd_end   = initrd_start+ia64_boot_param->initrd_size;
398 
399 		printk(KERN_INFO "Initial ramdisk at: 0x%lx (%llu bytes)\n",
400 		       initrd_start, ia64_boot_param->initrd_size);
401 	}
402 #endif
403 }
404 
405 static void __init
406 io_port_init (void)
407 {
408 	unsigned long phys_iobase;
409 
410 	/*
411 	 * Set `iobase' based on the EFI memory map or, failing that, the
412 	 * value firmware left in ar.k0.
413 	 *
414 	 * Note that in ia32 mode, IN/OUT instructions use ar.k0 to compute
415 	 * the port's virtual address, so ia32_load_state() loads it with a
416 	 * user virtual address.  But in ia64 mode, glibc uses the
417 	 * *physical* address in ar.k0 to mmap the appropriate area from
418 	 * /dev/mem, and the inX()/outX() interfaces use MMIO.  In both
419 	 * cases, user-mode can only use the legacy 0-64K I/O port space.
420 	 *
421 	 * ar.k0 is not involved in kernel I/O port accesses, which can use
422 	 * any of the I/O port spaces and are done via MMIO using the
423 	 * virtual mmio_base from the appropriate io_space[].
424 	 */
425 	phys_iobase = efi_get_iobase();
426 	if (!phys_iobase) {
427 		phys_iobase = ia64_get_kr(IA64_KR_IO_BASE);
428 		printk(KERN_INFO "No I/O port range found in EFI memory map, "
429 			"falling back to AR.KR0 (0x%lx)\n", phys_iobase);
430 	}
431 	ia64_iobase = (unsigned long) ioremap(phys_iobase, 0);
432 	ia64_set_kr(IA64_KR_IO_BASE, __pa(ia64_iobase));
433 
434 	/* setup legacy IO port space */
435 	io_space[0].mmio_base = ia64_iobase;
436 	io_space[0].sparse = 1;
437 	num_io_spaces = 1;
438 }
439 
440 /**
441  * early_console_setup - setup debugging console
442  *
443  * Consoles started here require little enough setup that we can start using
444  * them very early in the boot process, either right after the machine
445  * vector initialization, or even before if the drivers can detect their hw.
446  *
447  * Returns non-zero if a console couldn't be setup.
448  */
449 static inline int __init
450 early_console_setup (char *cmdline)
451 {
452 	int earlycons = 0;
453 
454 #ifdef CONFIG_SERIAL_SGI_L1_CONSOLE
455 	{
456 		extern int sn_serial_console_early_setup(void);
457 		if (!sn_serial_console_early_setup())
458 			earlycons++;
459 	}
460 #endif
461 #ifdef CONFIG_EFI_PCDP
462 	if (!efi_setup_pcdp_console(cmdline))
463 		earlycons++;
464 #endif
465 	if (!simcons_register())
466 		earlycons++;
467 
468 	return (earlycons) ? 0 : -1;
469 }
470 
471 static inline void
472 mark_bsp_online (void)
473 {
474 #ifdef CONFIG_SMP
475 	/* If we register an early console, allow CPU 0 to printk */
476 	cpu_set(smp_processor_id(), cpu_online_map);
477 #endif
478 }
479 
480 static __initdata int nomca;
481 static __init int setup_nomca(char *s)
482 {
483 	nomca = 1;
484 	return 0;
485 }
486 early_param("nomca", setup_nomca);
487 
488 /*
489  * Note: elfcorehdr_addr is not just limited to vmcore. It is also used by
490  * is_kdump_kernel() to determine if we are booting after a panic. Hence
491  * ifdef it under CONFIG_CRASH_DUMP and not CONFIG_PROC_VMCORE.
492  */
493 #ifdef CONFIG_CRASH_DUMP
494 /* elfcorehdr= specifies the location of elf core header
495  * stored by the crashed kernel.
496  */
497 static int __init parse_elfcorehdr(char *arg)
498 {
499 	if (!arg)
500 		return -EINVAL;
501 
502         elfcorehdr_addr = memparse(arg, &arg);
503 	return 0;
504 }
505 early_param("elfcorehdr", parse_elfcorehdr);
506 
507 int __init reserve_elfcorehdr(u64 *start, u64 *end)
508 {
509 	u64 length;
510 
511 	/* We get the address using the kernel command line,
512 	 * but the size is extracted from the EFI tables.
513 	 * Both address and size are required for reservation
514 	 * to work properly.
515 	 */
516 
517 	if (!is_vmcore_usable())
518 		return -EINVAL;
519 
520 	if ((length = vmcore_find_descriptor_size(elfcorehdr_addr)) == 0) {
521 		vmcore_unusable();
522 		return -EINVAL;
523 	}
524 
525 	*start = (unsigned long)__va(elfcorehdr_addr);
526 	*end = *start + length;
527 	return 0;
528 }
529 
530 #endif /* CONFIG_PROC_VMCORE */
531 
532 void __init
533 setup_arch (char **cmdline_p)
534 {
535 	unw_init();
536 
537 	paravirt_arch_setup_early();
538 
539 	ia64_patch_vtop((u64) __start___vtop_patchlist, (u64) __end___vtop_patchlist);
540 	paravirt_patch_apply();
541 
542 	*cmdline_p = __va(ia64_boot_param->command_line);
543 	strlcpy(boot_command_line, *cmdline_p, COMMAND_LINE_SIZE);
544 
545 	efi_init();
546 	io_port_init();
547 
548 #ifdef CONFIG_IA64_GENERIC
549 	/* machvec needs to be parsed from the command line
550 	 * before parse_early_param() is called to ensure
551 	 * that ia64_mv is initialised before any command line
552 	 * settings may cause console setup to occur
553 	 */
554 	machvec_init_from_cmdline(*cmdline_p);
555 #endif
556 
557 	parse_early_param();
558 
559 	if (early_console_setup(*cmdline_p) == 0)
560 		mark_bsp_online();
561 
562 #ifdef CONFIG_ACPI
563 	/* Initialize the ACPI boot-time table parser */
564 	acpi_table_init();
565 	early_acpi_boot_init();
566 # ifdef CONFIG_ACPI_NUMA
567 	acpi_numa_init();
568 #  ifdef CONFIG_ACPI_HOTPLUG_CPU
569 	prefill_possible_map();
570 #  endif
571 	per_cpu_scan_finalize((cpus_weight(early_cpu_possible_map) == 0 ?
572 		32 : cpus_weight(early_cpu_possible_map)),
573 		additional_cpus > 0 ? additional_cpus : 0);
574 # endif
575 #endif /* CONFIG_APCI_BOOT */
576 
577 #ifdef CONFIG_SMP
578 	smp_build_cpu_map();
579 #endif
580 	find_memory();
581 
582 	/* process SAL system table: */
583 	ia64_sal_init(__va(efi.sal_systab));
584 
585 #ifdef CONFIG_ITANIUM
586 	ia64_patch_rse((u64) __start___rse_patchlist, (u64) __end___rse_patchlist);
587 #else
588 	{
589 		unsigned long num_phys_stacked;
590 
591 		if (ia64_pal_rse_info(&num_phys_stacked, 0) == 0 && num_phys_stacked > 96)
592 			ia64_patch_rse((u64) __start___rse_patchlist, (u64) __end___rse_patchlist);
593 	}
594 #endif
595 
596 #ifdef CONFIG_SMP
597 	cpu_physical_id(0) = hard_smp_processor_id();
598 #endif
599 
600 	cpu_init();	/* initialize the bootstrap CPU */
601 	mmu_context_init();	/* initialize context_id bitmap */
602 
603 #ifdef CONFIG_ACPI
604 	acpi_boot_init();
605 #endif
606 
607 	paravirt_banner();
608 	paravirt_arch_setup_console(cmdline_p);
609 
610 #ifdef CONFIG_VT
611 	if (!conswitchp) {
612 # if defined(CONFIG_DUMMY_CONSOLE)
613 		conswitchp = &dummy_con;
614 # endif
615 # if defined(CONFIG_VGA_CONSOLE)
616 		/*
617 		 * Non-legacy systems may route legacy VGA MMIO range to system
618 		 * memory.  vga_con probes the MMIO hole, so memory looks like
619 		 * a VGA device to it.  The EFI memory map can tell us if it's
620 		 * memory so we can avoid this problem.
621 		 */
622 		if (efi_mem_type(0xA0000) != EFI_CONVENTIONAL_MEMORY)
623 			conswitchp = &vga_con;
624 # endif
625 	}
626 #endif
627 
628 	/* enable IA-64 Machine Check Abort Handling unless disabled */
629 	if (paravirt_arch_setup_nomca())
630 		nomca = 1;
631 	if (!nomca)
632 		ia64_mca_init();
633 
634 	platform_setup(cmdline_p);
635 #ifndef CONFIG_IA64_HP_SIM
636 	check_sal_cache_flush();
637 #endif
638 	paging_init();
639 }
640 
641 /*
642  * Display cpu info for all CPUs.
643  */
644 static int
645 show_cpuinfo (struct seq_file *m, void *v)
646 {
647 #ifdef CONFIG_SMP
648 #	define lpj	c->loops_per_jiffy
649 #	define cpunum	c->cpu
650 #else
651 #	define lpj	loops_per_jiffy
652 #	define cpunum	0
653 #endif
654 	static struct {
655 		unsigned long mask;
656 		const char *feature_name;
657 	} feature_bits[] = {
658 		{ 1UL << 0, "branchlong" },
659 		{ 1UL << 1, "spontaneous deferral"},
660 		{ 1UL << 2, "16-byte atomic ops" }
661 	};
662 	char features[128], *cp, *sep;
663 	struct cpuinfo_ia64 *c = v;
664 	unsigned long mask;
665 	unsigned long proc_freq;
666 	int i, size;
667 
668 	mask = c->features;
669 
670 	/* build the feature string: */
671 	memcpy(features, "standard", 9);
672 	cp = features;
673 	size = sizeof(features);
674 	sep = "";
675 	for (i = 0; i < ARRAY_SIZE(feature_bits) && size > 1; ++i) {
676 		if (mask & feature_bits[i].mask) {
677 			cp += snprintf(cp, size, "%s%s", sep,
678 				       feature_bits[i].feature_name),
679 			sep = ", ";
680 			mask &= ~feature_bits[i].mask;
681 			size = sizeof(features) - (cp - features);
682 		}
683 	}
684 	if (mask && size > 1) {
685 		/* print unknown features as a hex value */
686 		snprintf(cp, size, "%s0x%lx", sep, mask);
687 	}
688 
689 	proc_freq = cpufreq_quick_get(cpunum);
690 	if (!proc_freq)
691 		proc_freq = c->proc_freq / 1000;
692 
693 	seq_printf(m,
694 		   "processor  : %d\n"
695 		   "vendor     : %s\n"
696 		   "arch       : IA-64\n"
697 		   "family     : %u\n"
698 		   "model      : %u\n"
699 		   "model name : %s\n"
700 		   "revision   : %u\n"
701 		   "archrev    : %u\n"
702 		   "features   : %s\n"
703 		   "cpu number : %lu\n"
704 		   "cpu regs   : %u\n"
705 		   "cpu MHz    : %lu.%03lu\n"
706 		   "itc MHz    : %lu.%06lu\n"
707 		   "BogoMIPS   : %lu.%02lu\n",
708 		   cpunum, c->vendor, c->family, c->model,
709 		   c->model_name, c->revision, c->archrev,
710 		   features, c->ppn, c->number,
711 		   proc_freq / 1000, proc_freq % 1000,
712 		   c->itc_freq / 1000000, c->itc_freq % 1000000,
713 		   lpj*HZ/500000, (lpj*HZ/5000) % 100);
714 #ifdef CONFIG_SMP
715 	seq_printf(m, "siblings   : %u\n", cpus_weight(cpu_core_map[cpunum]));
716 	if (c->socket_id != -1)
717 		seq_printf(m, "physical id: %u\n", c->socket_id);
718 	if (c->threads_per_core > 1 || c->cores_per_socket > 1)
719 		seq_printf(m,
720 			   "core id    : %u\n"
721 			   "thread id  : %u\n",
722 			   c->core_id, c->thread_id);
723 #endif
724 	seq_printf(m,"\n");
725 
726 	return 0;
727 }
728 
729 static void *
730 c_start (struct seq_file *m, loff_t *pos)
731 {
732 #ifdef CONFIG_SMP
733 	while (*pos < nr_cpu_ids && !cpu_online(*pos))
734 		++*pos;
735 #endif
736 	return *pos < nr_cpu_ids ? cpu_data(*pos) : NULL;
737 }
738 
739 static void *
740 c_next (struct seq_file *m, void *v, loff_t *pos)
741 {
742 	++*pos;
743 	return c_start(m, pos);
744 }
745 
746 static void
747 c_stop (struct seq_file *m, void *v)
748 {
749 }
750 
751 const struct seq_operations cpuinfo_op = {
752 	.start =	c_start,
753 	.next =		c_next,
754 	.stop =		c_stop,
755 	.show =		show_cpuinfo
756 };
757 
758 #define MAX_BRANDS	8
759 static char brandname[MAX_BRANDS][128];
760 
761 static char * __cpuinit
762 get_model_name(__u8 family, __u8 model)
763 {
764 	static int overflow;
765 	char brand[128];
766 	int i;
767 
768 	memcpy(brand, "Unknown", 8);
769 	if (ia64_pal_get_brand_info(brand)) {
770 		if (family == 0x7)
771 			memcpy(brand, "Merced", 7);
772 		else if (family == 0x1f) switch (model) {
773 			case 0: memcpy(brand, "McKinley", 9); break;
774 			case 1: memcpy(brand, "Madison", 8); break;
775 			case 2: memcpy(brand, "Madison up to 9M cache", 23); break;
776 		}
777 	}
778 	for (i = 0; i < MAX_BRANDS; i++)
779 		if (strcmp(brandname[i], brand) == 0)
780 			return brandname[i];
781 	for (i = 0; i < MAX_BRANDS; i++)
782 		if (brandname[i][0] == '\0')
783 			return strcpy(brandname[i], brand);
784 	if (overflow++ == 0)
785 		printk(KERN_ERR
786 		       "%s: Table overflow. Some processor model information will be missing\n",
787 		       __func__);
788 	return "Unknown";
789 }
790 
791 static void __cpuinit
792 identify_cpu (struct cpuinfo_ia64 *c)
793 {
794 	union {
795 		unsigned long bits[5];
796 		struct {
797 			/* id 0 & 1: */
798 			char vendor[16];
799 
800 			/* id 2 */
801 			u64 ppn;		/* processor serial number */
802 
803 			/* id 3: */
804 			unsigned number		:  8;
805 			unsigned revision	:  8;
806 			unsigned model		:  8;
807 			unsigned family		:  8;
808 			unsigned archrev	:  8;
809 			unsigned reserved	: 24;
810 
811 			/* id 4: */
812 			u64 features;
813 		} field;
814 	} cpuid;
815 	pal_vm_info_1_u_t vm1;
816 	pal_vm_info_2_u_t vm2;
817 	pal_status_t status;
818 	unsigned long impl_va_msb = 50, phys_addr_size = 44;	/* Itanium defaults */
819 	int i;
820 	for (i = 0; i < 5; ++i)
821 		cpuid.bits[i] = ia64_get_cpuid(i);
822 
823 	memcpy(c->vendor, cpuid.field.vendor, 16);
824 #ifdef CONFIG_SMP
825 	c->cpu = smp_processor_id();
826 
827 	/* below default values will be overwritten  by identify_siblings()
828 	 * for Multi-Threading/Multi-Core capable CPUs
829 	 */
830 	c->threads_per_core = c->cores_per_socket = c->num_log = 1;
831 	c->socket_id = -1;
832 
833 	identify_siblings(c);
834 
835 	if (c->threads_per_core > smp_num_siblings)
836 		smp_num_siblings = c->threads_per_core;
837 #endif
838 	c->ppn = cpuid.field.ppn;
839 	c->number = cpuid.field.number;
840 	c->revision = cpuid.field.revision;
841 	c->model = cpuid.field.model;
842 	c->family = cpuid.field.family;
843 	c->archrev = cpuid.field.archrev;
844 	c->features = cpuid.field.features;
845 	c->model_name = get_model_name(c->family, c->model);
846 
847 	status = ia64_pal_vm_summary(&vm1, &vm2);
848 	if (status == PAL_STATUS_SUCCESS) {
849 		impl_va_msb = vm2.pal_vm_info_2_s.impl_va_msb;
850 		phys_addr_size = vm1.pal_vm_info_1_s.phys_add_size;
851 	}
852 	c->unimpl_va_mask = ~((7L<<61) | ((1L << (impl_va_msb + 1)) - 1));
853 	c->unimpl_pa_mask = ~((1L<<63) | ((1L << phys_addr_size) - 1));
854 }
855 
856 /*
857  * Do the following calculations:
858  *
859  * 1. the max. cache line size.
860  * 2. the minimum of the i-cache stride sizes for "flush_icache_range()".
861  * 3. the minimum of the cache stride sizes for "clflush_cache_range()".
862  */
863 static void __cpuinit
864 get_cache_info(void)
865 {
866 	unsigned long line_size, max = 1;
867 	unsigned long l, levels, unique_caches;
868 	pal_cache_config_info_t cci;
869 	long status;
870 
871         status = ia64_pal_cache_summary(&levels, &unique_caches);
872         if (status != 0) {
873                 printk(KERN_ERR "%s: ia64_pal_cache_summary() failed (status=%ld)\n",
874                        __func__, status);
875                 max = SMP_CACHE_BYTES;
876 		/* Safest setup for "flush_icache_range()" */
877 		ia64_i_cache_stride_shift = I_CACHE_STRIDE_SHIFT;
878 		/* Safest setup for "clflush_cache_range()" */
879 		ia64_cache_stride_shift = CACHE_STRIDE_SHIFT;
880 		goto out;
881         }
882 
883 	for (l = 0; l < levels; ++l) {
884 		/* cache_type (data_or_unified)=2 */
885 		status = ia64_pal_cache_config_info(l, 2, &cci);
886 		if (status != 0) {
887 			printk(KERN_ERR "%s: ia64_pal_cache_config_info"
888 				"(l=%lu, 2) failed (status=%ld)\n",
889 				__func__, l, status);
890 			max = SMP_CACHE_BYTES;
891 			/* The safest setup for "flush_icache_range()" */
892 			cci.pcci_stride = I_CACHE_STRIDE_SHIFT;
893 			/* The safest setup for "clflush_cache_range()" */
894 			ia64_cache_stride_shift = CACHE_STRIDE_SHIFT;
895 			cci.pcci_unified = 1;
896 		} else {
897 			if (cci.pcci_stride < ia64_cache_stride_shift)
898 				ia64_cache_stride_shift = cci.pcci_stride;
899 
900 			line_size = 1 << cci.pcci_line_size;
901 			if (line_size > max)
902 				max = line_size;
903 		}
904 
905 		if (!cci.pcci_unified) {
906 			/* cache_type (instruction)=1*/
907 			status = ia64_pal_cache_config_info(l, 1, &cci);
908 			if (status != 0) {
909 				printk(KERN_ERR "%s: ia64_pal_cache_config_info"
910 					"(l=%lu, 1) failed (status=%ld)\n",
911 					__func__, l, status);
912 				/* The safest setup for flush_icache_range() */
913 				cci.pcci_stride = I_CACHE_STRIDE_SHIFT;
914 			}
915 		}
916 		if (cci.pcci_stride < ia64_i_cache_stride_shift)
917 			ia64_i_cache_stride_shift = cci.pcci_stride;
918 	}
919   out:
920 	if (max > ia64_max_cacheline_size)
921 		ia64_max_cacheline_size = max;
922 }
923 
924 /*
925  * cpu_init() initializes state that is per-CPU.  This function acts
926  * as a 'CPU state barrier', nothing should get across.
927  */
928 void __cpuinit
929 cpu_init (void)
930 {
931 	extern void __cpuinit ia64_mmu_init (void *);
932 	static unsigned long max_num_phys_stacked = IA64_NUM_PHYS_STACK_REG;
933 	unsigned long num_phys_stacked;
934 	pal_vm_info_2_u_t vmi;
935 	unsigned int max_ctx;
936 	struct cpuinfo_ia64 *cpu_info;
937 	void *cpu_data;
938 
939 	cpu_data = per_cpu_init();
940 #ifdef CONFIG_SMP
941 	/*
942 	 * insert boot cpu into sibling and core mapes
943 	 * (must be done after per_cpu area is setup)
944 	 */
945 	if (smp_processor_id() == 0) {
946 		cpu_set(0, per_cpu(cpu_sibling_map, 0));
947 		cpu_set(0, cpu_core_map[0]);
948 	} else {
949 		/*
950 		 * Set ar.k3 so that assembly code in MCA handler can compute
951 		 * physical addresses of per cpu variables with a simple:
952 		 *   phys = ar.k3 + &per_cpu_var
953 		 * and the alt-dtlb-miss handler can set per-cpu mapping into
954 		 * the TLB when needed. head.S already did this for cpu0.
955 		 */
956 		ia64_set_kr(IA64_KR_PER_CPU_DATA,
957 			    ia64_tpa(cpu_data) - (long) __per_cpu_start);
958 	}
959 #endif
960 
961 	get_cache_info();
962 
963 	/*
964 	 * We can't pass "local_cpu_data" to identify_cpu() because we haven't called
965 	 * ia64_mmu_init() yet.  And we can't call ia64_mmu_init() first because it
966 	 * depends on the data returned by identify_cpu().  We break the dependency by
967 	 * accessing cpu_data() through the canonical per-CPU address.
968 	 */
969 	cpu_info = cpu_data + ((char *) &__ia64_per_cpu_var(ia64_cpu_info) - __per_cpu_start);
970 	identify_cpu(cpu_info);
971 
972 #ifdef CONFIG_MCKINLEY
973 	{
974 #		define FEATURE_SET 16
975 		struct ia64_pal_retval iprv;
976 
977 		if (cpu_info->family == 0x1f) {
978 			PAL_CALL_PHYS(iprv, PAL_PROC_GET_FEATURES, 0, FEATURE_SET, 0);
979 			if ((iprv.status == 0) && (iprv.v0 & 0x80) && (iprv.v2 & 0x80))
980 				PAL_CALL_PHYS(iprv, PAL_PROC_SET_FEATURES,
981 				              (iprv.v1 | 0x80), FEATURE_SET, 0);
982 		}
983 	}
984 #endif
985 
986 	/* Clear the stack memory reserved for pt_regs: */
987 	memset(task_pt_regs(current), 0, sizeof(struct pt_regs));
988 
989 	ia64_set_kr(IA64_KR_FPU_OWNER, 0);
990 
991 	/*
992 	 * Initialize the page-table base register to a global
993 	 * directory with all zeroes.  This ensure that we can handle
994 	 * TLB-misses to user address-space even before we created the
995 	 * first user address-space.  This may happen, e.g., due to
996 	 * aggressive use of lfetch.fault.
997 	 */
998 	ia64_set_kr(IA64_KR_PT_BASE, __pa(ia64_imva(empty_zero_page)));
999 
1000 	/*
1001 	 * Initialize default control register to defer speculative faults except
1002 	 * for those arising from TLB misses, which are not deferred.  The
1003 	 * kernel MUST NOT depend on a particular setting of these bits (in other words,
1004 	 * the kernel must have recovery code for all speculative accesses).  Turn on
1005 	 * dcr.lc as per recommendation by the architecture team.  Most IA-32 apps
1006 	 * shouldn't be affected by this (moral: keep your ia32 locks aligned and you'll
1007 	 * be fine).
1008 	 */
1009 	ia64_setreg(_IA64_REG_CR_DCR,  (  IA64_DCR_DP | IA64_DCR_DK | IA64_DCR_DX | IA64_DCR_DR
1010 					| IA64_DCR_DA | IA64_DCR_DD | IA64_DCR_LC));
1011 	atomic_inc(&init_mm.mm_count);
1012 	current->active_mm = &init_mm;
1013 	BUG_ON(current->mm);
1014 
1015 	ia64_mmu_init(ia64_imva(cpu_data));
1016 	ia64_mca_cpu_init(ia64_imva(cpu_data));
1017 
1018 	/* Clear ITC to eliminate sched_clock() overflows in human time.  */
1019 	ia64_set_itc(0);
1020 
1021 	/* disable all local interrupt sources: */
1022 	ia64_set_itv(1 << 16);
1023 	ia64_set_lrr0(1 << 16);
1024 	ia64_set_lrr1(1 << 16);
1025 	ia64_setreg(_IA64_REG_CR_PMV, 1 << 16);
1026 	ia64_setreg(_IA64_REG_CR_CMCV, 1 << 16);
1027 
1028 	/* clear TPR & XTP to enable all interrupt classes: */
1029 	ia64_setreg(_IA64_REG_CR_TPR, 0);
1030 
1031 	/* Clear any pending interrupts left by SAL/EFI */
1032 	while (ia64_get_ivr() != IA64_SPURIOUS_INT_VECTOR)
1033 		ia64_eoi();
1034 
1035 #ifdef CONFIG_SMP
1036 	normal_xtp();
1037 #endif
1038 
1039 	/* set ia64_ctx.max_rid to the maximum RID that is supported by all CPUs: */
1040 	if (ia64_pal_vm_summary(NULL, &vmi) == 0) {
1041 		max_ctx = (1U << (vmi.pal_vm_info_2_s.rid_size - 3)) - 1;
1042 		setup_ptcg_sem(vmi.pal_vm_info_2_s.max_purges, NPTCG_FROM_PAL);
1043 	} else {
1044 		printk(KERN_WARNING "cpu_init: PAL VM summary failed, assuming 18 RID bits\n");
1045 		max_ctx = (1U << 15) - 1;	/* use architected minimum */
1046 	}
1047 	while (max_ctx < ia64_ctx.max_ctx) {
1048 		unsigned int old = ia64_ctx.max_ctx;
1049 		if (cmpxchg(&ia64_ctx.max_ctx, old, max_ctx) == old)
1050 			break;
1051 	}
1052 
1053 	if (ia64_pal_rse_info(&num_phys_stacked, NULL) != 0) {
1054 		printk(KERN_WARNING "cpu_init: PAL RSE info failed; assuming 96 physical "
1055 		       "stacked regs\n");
1056 		num_phys_stacked = 96;
1057 	}
1058 	/* size of physical stacked register partition plus 8 bytes: */
1059 	if (num_phys_stacked > max_num_phys_stacked) {
1060 		ia64_patch_phys_stack_reg(num_phys_stacked*8 + 8);
1061 		max_num_phys_stacked = num_phys_stacked;
1062 	}
1063 	platform_cpu_init();
1064 	pm_idle = default_idle;
1065 }
1066 
1067 void __init
1068 check_bugs (void)
1069 {
1070 	ia64_patch_mckinley_e9((unsigned long) __start___mckinley_e9_bundles,
1071 			       (unsigned long) __end___mckinley_e9_bundles);
1072 }
1073 
1074 static int __init run_dmi_scan(void)
1075 {
1076 	dmi_scan_machine();
1077 	return 0;
1078 }
1079 core_initcall(run_dmi_scan);
1080