xref: /openbmc/linux/arch/ia64/kernel/setup.c (revision ce6e71ad)
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/config.h>
26 #include <linux/module.h>
27 #include <linux/init.h>
28 
29 #include <linux/acpi.h>
30 #include <linux/bootmem.h>
31 #include <linux/console.h>
32 #include <linux/delay.h>
33 #include <linux/kernel.h>
34 #include <linux/reboot.h>
35 #include <linux/sched.h>
36 #include <linux/seq_file.h>
37 #include <linux/string.h>
38 #include <linux/threads.h>
39 #include <linux/tty.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/platform.h>
45 #include <linux/pm.h>
46 
47 #include <asm/ia32.h>
48 #include <asm/machvec.h>
49 #include <asm/mca.h>
50 #include <asm/meminit.h>
51 #include <asm/page.h>
52 #include <asm/patch.h>
53 #include <asm/pgtable.h>
54 #include <asm/processor.h>
55 #include <asm/sal.h>
56 #include <asm/sections.h>
57 #include <asm/serial.h>
58 #include <asm/setup.h>
59 #include <asm/smp.h>
60 #include <asm/system.h>
61 #include <asm/unistd.h>
62 
63 #if defined(CONFIG_SMP) && (IA64_CPU_SIZE > PAGE_SIZE)
64 # error "struct cpuinfo_ia64 too big!"
65 #endif
66 
67 #ifdef CONFIG_SMP
68 unsigned long __per_cpu_offset[NR_CPUS];
69 EXPORT_SYMBOL(__per_cpu_offset);
70 #endif
71 
72 DEFINE_PER_CPU(struct cpuinfo_ia64, cpu_info);
73 DEFINE_PER_CPU(unsigned long, local_per_cpu_offset);
74 DEFINE_PER_CPU(unsigned long, ia64_phys_stacked_size_p8);
75 unsigned long ia64_cycles_per_usec;
76 struct ia64_boot_param *ia64_boot_param;
77 struct screen_info screen_info;
78 unsigned long vga_console_iobase;
79 unsigned long vga_console_membase;
80 
81 unsigned long ia64_max_cacheline_size;
82 unsigned long ia64_iobase;	/* virtual address for I/O accesses */
83 EXPORT_SYMBOL(ia64_iobase);
84 struct io_space io_space[MAX_IO_SPACES];
85 EXPORT_SYMBOL(io_space);
86 unsigned int num_io_spaces;
87 
88 /*
89  * "flush_icache_range()" needs to know what processor dependent stride size to use
90  * when it makes i-cache(s) coherent with d-caches.
91  */
92 #define	I_CACHE_STRIDE_SHIFT	5	/* Safest way to go: 32 bytes by 32 bytes */
93 unsigned long ia64_i_cache_stride_shift = ~0;
94 
95 /*
96  * The merge_mask variable needs to be set to (max(iommu_page_size(iommu)) - 1).  This
97  * mask specifies a mask of address bits that must be 0 in order for two buffers to be
98  * mergeable by the I/O MMU (i.e., the end address of the first buffer and the start
99  * address of the second buffer must be aligned to (merge_mask+1) in order to be
100  * mergeable).  By default, we assume there is no I/O MMU which can merge physically
101  * discontiguous buffers, so we set the merge_mask to ~0UL, which corresponds to a iommu
102  * page-size of 2^64.
103  */
104 unsigned long ia64_max_iommu_merge_mask = ~0UL;
105 EXPORT_SYMBOL(ia64_max_iommu_merge_mask);
106 
107 /*
108  * We use a special marker for the end of memory and it uses the extra (+1) slot
109  */
110 struct rsvd_region rsvd_region[IA64_MAX_RSVD_REGIONS + 1];
111 int num_rsvd_regions;
112 
113 
114 /*
115  * Filter incoming memory segments based on the primitive map created from the boot
116  * parameters. Segments contained in the map are removed from the memory ranges. A
117  * caller-specified function is called with the memory ranges that remain after filtering.
118  * This routine does not assume the incoming segments are sorted.
119  */
120 int
121 filter_rsvd_memory (unsigned long start, unsigned long end, void *arg)
122 {
123 	unsigned long range_start, range_end, prev_start;
124 	void (*func)(unsigned long, unsigned long, int);
125 	int i;
126 
127 #if IGNORE_PFN0
128 	if (start == PAGE_OFFSET) {
129 		printk(KERN_WARNING "warning: skipping physical page 0\n");
130 		start += PAGE_SIZE;
131 		if (start >= end) return 0;
132 	}
133 #endif
134 	/*
135 	 * lowest possible address(walker uses virtual)
136 	 */
137 	prev_start = PAGE_OFFSET;
138 	func = arg;
139 
140 	for (i = 0; i < num_rsvd_regions; ++i) {
141 		range_start = max(start, prev_start);
142 		range_end   = min(end, rsvd_region[i].start);
143 
144 		if (range_start < range_end)
145 			call_pernode_memory(__pa(range_start), range_end - range_start, func);
146 
147 		/* nothing more available in this segment */
148 		if (range_end == end) return 0;
149 
150 		prev_start = rsvd_region[i].end;
151 	}
152 	/* end of memory marker allows full processing inside loop body */
153 	return 0;
154 }
155 
156 static void
157 sort_regions (struct rsvd_region *rsvd_region, int max)
158 {
159 	int j;
160 
161 	/* simple bubble sorting */
162 	while (max--) {
163 		for (j = 0; j < max; ++j) {
164 			if (rsvd_region[j].start > rsvd_region[j+1].start) {
165 				struct rsvd_region tmp;
166 				tmp = rsvd_region[j];
167 				rsvd_region[j] = rsvd_region[j + 1];
168 				rsvd_region[j + 1] = tmp;
169 			}
170 		}
171 	}
172 }
173 
174 /**
175  * reserve_memory - setup reserved memory areas
176  *
177  * Setup the reserved memory areas set aside for the boot parameters,
178  * initrd, etc.  There are currently %IA64_MAX_RSVD_REGIONS defined,
179  * see include/asm-ia64/meminit.h if you need to define more.
180  */
181 void
182 reserve_memory (void)
183 {
184 	int n = 0;
185 
186 	/*
187 	 * none of the entries in this table overlap
188 	 */
189 	rsvd_region[n].start = (unsigned long) ia64_boot_param;
190 	rsvd_region[n].end   = rsvd_region[n].start + sizeof(*ia64_boot_param);
191 	n++;
192 
193 	rsvd_region[n].start = (unsigned long) __va(ia64_boot_param->efi_memmap);
194 	rsvd_region[n].end   = rsvd_region[n].start + ia64_boot_param->efi_memmap_size;
195 	n++;
196 
197 	rsvd_region[n].start = (unsigned long) __va(ia64_boot_param->command_line);
198 	rsvd_region[n].end   = (rsvd_region[n].start
199 				+ strlen(__va(ia64_boot_param->command_line)) + 1);
200 	n++;
201 
202 	rsvd_region[n].start = (unsigned long) ia64_imva((void *)KERNEL_START);
203 	rsvd_region[n].end   = (unsigned long) ia64_imva(_end);
204 	n++;
205 
206 #ifdef CONFIG_BLK_DEV_INITRD
207 	if (ia64_boot_param->initrd_start) {
208 		rsvd_region[n].start = (unsigned long)__va(ia64_boot_param->initrd_start);
209 		rsvd_region[n].end   = rsvd_region[n].start + ia64_boot_param->initrd_size;
210 		n++;
211 	}
212 #endif
213 
214 	/* end of memory marker */
215 	rsvd_region[n].start = ~0UL;
216 	rsvd_region[n].end   = ~0UL;
217 	n++;
218 
219 	num_rsvd_regions = n;
220 
221 	sort_regions(rsvd_region, num_rsvd_regions);
222 }
223 
224 /**
225  * find_initrd - get initrd parameters from the boot parameter structure
226  *
227  * Grab the initrd start and end from the boot parameter struct given us by
228  * the boot loader.
229  */
230 void
231 find_initrd (void)
232 {
233 #ifdef CONFIG_BLK_DEV_INITRD
234 	if (ia64_boot_param->initrd_start) {
235 		initrd_start = (unsigned long)__va(ia64_boot_param->initrd_start);
236 		initrd_end   = initrd_start+ia64_boot_param->initrd_size;
237 
238 		printk(KERN_INFO "Initial ramdisk at: 0x%lx (%lu bytes)\n",
239 		       initrd_start, ia64_boot_param->initrd_size);
240 	}
241 #endif
242 }
243 
244 static void __init
245 io_port_init (void)
246 {
247 	extern unsigned long ia64_iobase;
248 	unsigned long phys_iobase;
249 
250 	/*
251 	 *  Set `iobase' to the appropriate address in region 6 (uncached access range).
252 	 *
253 	 *  The EFI memory map is the "preferred" location to get the I/O port space base,
254 	 *  rather the relying on AR.KR0. This should become more clear in future SAL
255 	 *  specs. We'll fall back to getting it out of AR.KR0 if no appropriate entry is
256 	 *  found in the memory map.
257 	 */
258 	phys_iobase = efi_get_iobase();
259 	if (phys_iobase)
260 		/* set AR.KR0 since this is all we use it for anyway */
261 		ia64_set_kr(IA64_KR_IO_BASE, phys_iobase);
262 	else {
263 		phys_iobase = ia64_get_kr(IA64_KR_IO_BASE);
264 		printk(KERN_INFO "No I/O port range found in EFI memory map, falling back "
265 		       "to AR.KR0\n");
266 		printk(KERN_INFO "I/O port base = 0x%lx\n", phys_iobase);
267 	}
268 	ia64_iobase = (unsigned long) ioremap(phys_iobase, 0);
269 
270 	/* setup legacy IO port space */
271 	io_space[0].mmio_base = ia64_iobase;
272 	io_space[0].sparse = 1;
273 	num_io_spaces = 1;
274 }
275 
276 /**
277  * early_console_setup - setup debugging console
278  *
279  * Consoles started here require little enough setup that we can start using
280  * them very early in the boot process, either right after the machine
281  * vector initialization, or even before if the drivers can detect their hw.
282  *
283  * Returns non-zero if a console couldn't be setup.
284  */
285 static inline int __init
286 early_console_setup (char *cmdline)
287 {
288 	int earlycons = 0;
289 
290 #ifdef CONFIG_SERIAL_SGI_L1_CONSOLE
291 	{
292 		extern int sn_serial_console_early_setup(void);
293 		if (!sn_serial_console_early_setup())
294 			earlycons++;
295 	}
296 #endif
297 #ifdef CONFIG_EFI_PCDP
298 	if (!efi_setup_pcdp_console(cmdline))
299 		earlycons++;
300 #endif
301 #ifdef CONFIG_SERIAL_8250_CONSOLE
302 	if (!early_serial_console_init(cmdline))
303 		earlycons++;
304 #endif
305 
306 	return (earlycons) ? 0 : -1;
307 }
308 
309 static inline void
310 mark_bsp_online (void)
311 {
312 #ifdef CONFIG_SMP
313 	/* If we register an early console, allow CPU 0 to printk */
314 	cpu_set(smp_processor_id(), cpu_online_map);
315 #endif
316 }
317 
318 #ifdef CONFIG_SMP
319 static void
320 check_for_logical_procs (void)
321 {
322 	pal_logical_to_physical_t info;
323 	s64 status;
324 
325 	status = ia64_pal_logical_to_phys(0, &info);
326 	if (status == -1) {
327 		printk(KERN_INFO "No logical to physical processor mapping "
328 		       "available\n");
329 		return;
330 	}
331 	if (status) {
332 		printk(KERN_ERR "ia64_pal_logical_to_phys failed with %ld\n",
333 		       status);
334 		return;
335 	}
336 	/*
337 	 * Total number of siblings that BSP has.  Though not all of them
338 	 * may have booted successfully. The correct number of siblings
339 	 * booted is in info.overview_num_log.
340 	 */
341 	smp_num_siblings = info.overview_tpc;
342 	smp_num_cpucores = info.overview_cpp;
343 }
344 #endif
345 
346 void __init
347 setup_arch (char **cmdline_p)
348 {
349 	unw_init();
350 
351 	ia64_patch_vtop((u64) __start___vtop_patchlist, (u64) __end___vtop_patchlist);
352 
353 	*cmdline_p = __va(ia64_boot_param->command_line);
354 	strlcpy(saved_command_line, *cmdline_p, COMMAND_LINE_SIZE);
355 
356 	efi_init();
357 	io_port_init();
358 
359 #ifdef CONFIG_IA64_GENERIC
360 	{
361 		const char *mvec_name = strstr (*cmdline_p, "machvec=");
362 		char str[64];
363 
364 		if (mvec_name) {
365 			const char *end;
366 			size_t len;
367 
368 			mvec_name += 8;
369 			end = strchr (mvec_name, ' ');
370 			if (end)
371 				len = end - mvec_name;
372 			else
373 				len = strlen (mvec_name);
374 			len = min(len, sizeof (str) - 1);
375 			strncpy (str, mvec_name, len);
376 			str[len] = '\0';
377 			mvec_name = str;
378 		} else
379 			mvec_name = acpi_get_sysname();
380 		machvec_init(mvec_name);
381 	}
382 #endif
383 
384 	if (early_console_setup(*cmdline_p) == 0)
385 		mark_bsp_online();
386 
387 #ifdef CONFIG_ACPI
388 	/* Initialize the ACPI boot-time table parser */
389 	acpi_table_init();
390 # ifdef CONFIG_ACPI_NUMA
391 	acpi_numa_init();
392 # endif
393 #else
394 # ifdef CONFIG_SMP
395 	smp_build_cpu_map();	/* happens, e.g., with the Ski simulator */
396 # endif
397 #endif /* CONFIG_APCI_BOOT */
398 
399 	find_memory();
400 
401 	/* process SAL system table: */
402 	ia64_sal_init(efi.sal_systab);
403 
404 #ifdef CONFIG_SMP
405 	cpu_physical_id(0) = hard_smp_processor_id();
406 
407 	cpu_set(0, cpu_sibling_map[0]);
408 	cpu_set(0, cpu_core_map[0]);
409 
410 	check_for_logical_procs();
411 	if (smp_num_cpucores > 1)
412 		printk(KERN_INFO
413 		       "cpu package is Multi-Core capable: number of cores=%d\n",
414 		       smp_num_cpucores);
415 	if (smp_num_siblings > 1)
416 		printk(KERN_INFO
417 		       "cpu package is Multi-Threading capable: number of siblings=%d\n",
418 		       smp_num_siblings);
419 #endif
420 
421 	cpu_init();	/* initialize the bootstrap CPU */
422 
423 #ifdef CONFIG_ACPI
424 	acpi_boot_init();
425 #endif
426 
427 #ifdef CONFIG_VT
428 	if (!conswitchp) {
429 # if defined(CONFIG_DUMMY_CONSOLE)
430 		conswitchp = &dummy_con;
431 # endif
432 # if defined(CONFIG_VGA_CONSOLE)
433 		/*
434 		 * Non-legacy systems may route legacy VGA MMIO range to system
435 		 * memory.  vga_con probes the MMIO hole, so memory looks like
436 		 * a VGA device to it.  The EFI memory map can tell us if it's
437 		 * memory so we can avoid this problem.
438 		 */
439 		if (efi_mem_type(0xA0000) != EFI_CONVENTIONAL_MEMORY)
440 			conswitchp = &vga_con;
441 # endif
442 	}
443 #endif
444 
445 	/* enable IA-64 Machine Check Abort Handling unless disabled */
446 	if (!strstr(saved_command_line, "nomca"))
447 		ia64_mca_init();
448 
449 	platform_setup(cmdline_p);
450 	paging_init();
451 }
452 
453 /*
454  * Display cpu info for all cpu's.
455  */
456 static int
457 show_cpuinfo (struct seq_file *m, void *v)
458 {
459 #ifdef CONFIG_SMP
460 #	define lpj	c->loops_per_jiffy
461 #	define cpunum	c->cpu
462 #else
463 #	define lpj	loops_per_jiffy
464 #	define cpunum	0
465 #endif
466 	static struct {
467 		unsigned long mask;
468 		const char *feature_name;
469 	} feature_bits[] = {
470 		{ 1UL << 0, "branchlong" },
471 		{ 1UL << 1, "spontaneous deferral"},
472 		{ 1UL << 2, "16-byte atomic ops" }
473 	};
474 	char family[32], features[128], *cp, sep;
475 	struct cpuinfo_ia64 *c = v;
476 	unsigned long mask;
477 	int i;
478 
479 	mask = c->features;
480 
481 	switch (c->family) {
482 	      case 0x07:	memcpy(family, "Itanium", 8); break;
483 	      case 0x1f:	memcpy(family, "Itanium 2", 10); break;
484 	      default:		sprintf(family, "%u", c->family); break;
485 	}
486 
487 	/* build the feature string: */
488 	memcpy(features, " standard", 10);
489 	cp = features;
490 	sep = 0;
491 	for (i = 0; i < (int) ARRAY_SIZE(feature_bits); ++i) {
492 		if (mask & feature_bits[i].mask) {
493 			if (sep)
494 				*cp++ = sep;
495 			sep = ',';
496 			*cp++ = ' ';
497 			strcpy(cp, feature_bits[i].feature_name);
498 			cp += strlen(feature_bits[i].feature_name);
499 			mask &= ~feature_bits[i].mask;
500 		}
501 	}
502 	if (mask) {
503 		/* print unknown features as a hex value: */
504 		if (sep)
505 			*cp++ = sep;
506 		sprintf(cp, " 0x%lx", mask);
507 	}
508 
509 	seq_printf(m,
510 		   "processor  : %d\n"
511 		   "vendor     : %s\n"
512 		   "arch       : IA-64\n"
513 		   "family     : %s\n"
514 		   "model      : %u\n"
515 		   "revision   : %u\n"
516 		   "archrev    : %u\n"
517 		   "features   :%s\n"	/* don't change this---it _is_ right! */
518 		   "cpu number : %lu\n"
519 		   "cpu regs   : %u\n"
520 		   "cpu MHz    : %lu.%06lu\n"
521 		   "itc MHz    : %lu.%06lu\n"
522 		   "BogoMIPS   : %lu.%02lu\n",
523 		   cpunum, c->vendor, family, c->model, c->revision, c->archrev,
524 		   features, c->ppn, c->number,
525 		   c->proc_freq / 1000000, c->proc_freq % 1000000,
526 		   c->itc_freq / 1000000, c->itc_freq % 1000000,
527 		   lpj*HZ/500000, (lpj*HZ/5000) % 100);
528 #ifdef CONFIG_SMP
529 	seq_printf(m, "siblings   : %u\n", cpus_weight(cpu_core_map[cpunum]));
530 	if (c->threads_per_core > 1 || c->cores_per_socket > 1)
531 		seq_printf(m,
532 		   	   "physical id: %u\n"
533 		   	   "core id    : %u\n"
534 		   	   "thread id  : %u\n",
535 		   	   c->socket_id, c->core_id, c->thread_id);
536 #endif
537 	seq_printf(m,"\n");
538 
539 	return 0;
540 }
541 
542 static void *
543 c_start (struct seq_file *m, loff_t *pos)
544 {
545 #ifdef CONFIG_SMP
546 	while (*pos < NR_CPUS && !cpu_isset(*pos, cpu_online_map))
547 		++*pos;
548 #endif
549 	return *pos < NR_CPUS ? cpu_data(*pos) : NULL;
550 }
551 
552 static void *
553 c_next (struct seq_file *m, void *v, loff_t *pos)
554 {
555 	++*pos;
556 	return c_start(m, pos);
557 }
558 
559 static void
560 c_stop (struct seq_file *m, void *v)
561 {
562 }
563 
564 struct seq_operations cpuinfo_op = {
565 	.start =	c_start,
566 	.next =		c_next,
567 	.stop =		c_stop,
568 	.show =		show_cpuinfo
569 };
570 
571 void
572 identify_cpu (struct cpuinfo_ia64 *c)
573 {
574 	union {
575 		unsigned long bits[5];
576 		struct {
577 			/* id 0 & 1: */
578 			char vendor[16];
579 
580 			/* id 2 */
581 			u64 ppn;		/* processor serial number */
582 
583 			/* id 3: */
584 			unsigned number		:  8;
585 			unsigned revision	:  8;
586 			unsigned model		:  8;
587 			unsigned family		:  8;
588 			unsigned archrev	:  8;
589 			unsigned reserved	: 24;
590 
591 			/* id 4: */
592 			u64 features;
593 		} field;
594 	} cpuid;
595 	pal_vm_info_1_u_t vm1;
596 	pal_vm_info_2_u_t vm2;
597 	pal_status_t status;
598 	unsigned long impl_va_msb = 50, phys_addr_size = 44;	/* Itanium defaults */
599 	int i;
600 
601 	for (i = 0; i < 5; ++i)
602 		cpuid.bits[i] = ia64_get_cpuid(i);
603 
604 	memcpy(c->vendor, cpuid.field.vendor, 16);
605 #ifdef CONFIG_SMP
606 	c->cpu = smp_processor_id();
607 
608 	/* below default values will be overwritten  by identify_siblings()
609 	 * for Multi-Threading/Multi-Core capable cpu's
610 	 */
611 	c->threads_per_core = c->cores_per_socket = c->num_log = 1;
612 	c->socket_id = -1;
613 
614 	identify_siblings(c);
615 #endif
616 	c->ppn = cpuid.field.ppn;
617 	c->number = cpuid.field.number;
618 	c->revision = cpuid.field.revision;
619 	c->model = cpuid.field.model;
620 	c->family = cpuid.field.family;
621 	c->archrev = cpuid.field.archrev;
622 	c->features = cpuid.field.features;
623 
624 	status = ia64_pal_vm_summary(&vm1, &vm2);
625 	if (status == PAL_STATUS_SUCCESS) {
626 		impl_va_msb = vm2.pal_vm_info_2_s.impl_va_msb;
627 		phys_addr_size = vm1.pal_vm_info_1_s.phys_add_size;
628 	}
629 	c->unimpl_va_mask = ~((7L<<61) | ((1L << (impl_va_msb + 1)) - 1));
630 	c->unimpl_pa_mask = ~((1L<<63) | ((1L << phys_addr_size) - 1));
631 }
632 
633 void
634 setup_per_cpu_areas (void)
635 {
636 	/* start_kernel() requires this... */
637 }
638 
639 /*
640  * Calculate the max. cache line size.
641  *
642  * In addition, the minimum of the i-cache stride sizes is calculated for
643  * "flush_icache_range()".
644  */
645 static void
646 get_max_cacheline_size (void)
647 {
648 	unsigned long line_size, max = 1;
649 	u64 l, levels, unique_caches;
650         pal_cache_config_info_t cci;
651         s64 status;
652 
653         status = ia64_pal_cache_summary(&levels, &unique_caches);
654         if (status != 0) {
655                 printk(KERN_ERR "%s: ia64_pal_cache_summary() failed (status=%ld)\n",
656                        __FUNCTION__, status);
657                 max = SMP_CACHE_BYTES;
658 		/* Safest setup for "flush_icache_range()" */
659 		ia64_i_cache_stride_shift = I_CACHE_STRIDE_SHIFT;
660 		goto out;
661         }
662 
663 	for (l = 0; l < levels; ++l) {
664 		status = ia64_pal_cache_config_info(l, /* cache_type (data_or_unified)= */ 2,
665 						    &cci);
666 		if (status != 0) {
667 			printk(KERN_ERR
668 			       "%s: ia64_pal_cache_config_info(l=%lu, 2) failed (status=%ld)\n",
669 			       __FUNCTION__, l, status);
670 			max = SMP_CACHE_BYTES;
671 			/* The safest setup for "flush_icache_range()" */
672 			cci.pcci_stride = I_CACHE_STRIDE_SHIFT;
673 			cci.pcci_unified = 1;
674 		}
675 		line_size = 1 << cci.pcci_line_size;
676 		if (line_size > max)
677 			max = line_size;
678 		if (!cci.pcci_unified) {
679 			status = ia64_pal_cache_config_info(l,
680 						    /* cache_type (instruction)= */ 1,
681 						    &cci);
682 			if (status != 0) {
683 				printk(KERN_ERR
684 				"%s: ia64_pal_cache_config_info(l=%lu, 1) failed (status=%ld)\n",
685 					__FUNCTION__, l, status);
686 				/* The safest setup for "flush_icache_range()" */
687 				cci.pcci_stride = I_CACHE_STRIDE_SHIFT;
688 			}
689 		}
690 		if (cci.pcci_stride < ia64_i_cache_stride_shift)
691 			ia64_i_cache_stride_shift = cci.pcci_stride;
692 	}
693   out:
694 	if (max > ia64_max_cacheline_size)
695 		ia64_max_cacheline_size = max;
696 }
697 
698 /*
699  * cpu_init() initializes state that is per-CPU.  This function acts
700  * as a 'CPU state barrier', nothing should get across.
701  */
702 void
703 cpu_init (void)
704 {
705 	extern void __devinit ia64_mmu_init (void *);
706 	unsigned long num_phys_stacked;
707 	pal_vm_info_2_u_t vmi;
708 	unsigned int max_ctx;
709 	struct cpuinfo_ia64 *cpu_info;
710 	void *cpu_data;
711 
712 	cpu_data = per_cpu_init();
713 
714 	/*
715 	 * We set ar.k3 so that assembly code in MCA handler can compute
716 	 * physical addresses of per cpu variables with a simple:
717 	 *   phys = ar.k3 + &per_cpu_var
718 	 */
719 	ia64_set_kr(IA64_KR_PER_CPU_DATA,
720 		    ia64_tpa(cpu_data) - (long) __per_cpu_start);
721 
722 	get_max_cacheline_size();
723 
724 	/*
725 	 * We can't pass "local_cpu_data" to identify_cpu() because we haven't called
726 	 * ia64_mmu_init() yet.  And we can't call ia64_mmu_init() first because it
727 	 * depends on the data returned by identify_cpu().  We break the dependency by
728 	 * accessing cpu_data() through the canonical per-CPU address.
729 	 */
730 	cpu_info = cpu_data + ((char *) &__ia64_per_cpu_var(cpu_info) - __per_cpu_start);
731 	identify_cpu(cpu_info);
732 
733 #ifdef CONFIG_MCKINLEY
734 	{
735 #		define FEATURE_SET 16
736 		struct ia64_pal_retval iprv;
737 
738 		if (cpu_info->family == 0x1f) {
739 			PAL_CALL_PHYS(iprv, PAL_PROC_GET_FEATURES, 0, FEATURE_SET, 0);
740 			if ((iprv.status == 0) && (iprv.v0 & 0x80) && (iprv.v2 & 0x80))
741 				PAL_CALL_PHYS(iprv, PAL_PROC_SET_FEATURES,
742 				              (iprv.v1 | 0x80), FEATURE_SET, 0);
743 		}
744 	}
745 #endif
746 
747 	/* Clear the stack memory reserved for pt_regs: */
748 	memset(ia64_task_regs(current), 0, sizeof(struct pt_regs));
749 
750 	ia64_set_kr(IA64_KR_FPU_OWNER, 0);
751 
752 	/*
753 	 * Initialize the page-table base register to a global
754 	 * directory with all zeroes.  This ensure that we can handle
755 	 * TLB-misses to user address-space even before we created the
756 	 * first user address-space.  This may happen, e.g., due to
757 	 * aggressive use of lfetch.fault.
758 	 */
759 	ia64_set_kr(IA64_KR_PT_BASE, __pa(ia64_imva(empty_zero_page)));
760 
761 	/*
762 	 * Initialize default control register to defer speculative faults except
763 	 * for those arising from TLB misses, which are not deferred.  The
764 	 * kernel MUST NOT depend on a particular setting of these bits (in other words,
765 	 * the kernel must have recovery code for all speculative accesses).  Turn on
766 	 * dcr.lc as per recommendation by the architecture team.  Most IA-32 apps
767 	 * shouldn't be affected by this (moral: keep your ia32 locks aligned and you'll
768 	 * be fine).
769 	 */
770 	ia64_setreg(_IA64_REG_CR_DCR,  (  IA64_DCR_DP | IA64_DCR_DK | IA64_DCR_DX | IA64_DCR_DR
771 					| IA64_DCR_DA | IA64_DCR_DD | IA64_DCR_LC));
772 	atomic_inc(&init_mm.mm_count);
773 	current->active_mm = &init_mm;
774 	if (current->mm)
775 		BUG();
776 
777 	ia64_mmu_init(ia64_imva(cpu_data));
778 	ia64_mca_cpu_init(ia64_imva(cpu_data));
779 
780 #ifdef CONFIG_IA32_SUPPORT
781 	ia32_cpu_init();
782 #endif
783 
784 	/* Clear ITC to eliminiate sched_clock() overflows in human time.  */
785 	ia64_set_itc(0);
786 
787 	/* disable all local interrupt sources: */
788 	ia64_set_itv(1 << 16);
789 	ia64_set_lrr0(1 << 16);
790 	ia64_set_lrr1(1 << 16);
791 	ia64_setreg(_IA64_REG_CR_PMV, 1 << 16);
792 	ia64_setreg(_IA64_REG_CR_CMCV, 1 << 16);
793 
794 	/* clear TPR & XTP to enable all interrupt classes: */
795 	ia64_setreg(_IA64_REG_CR_TPR, 0);
796 #ifdef CONFIG_SMP
797 	normal_xtp();
798 #endif
799 
800 	/* set ia64_ctx.max_rid to the maximum RID that is supported by all CPUs: */
801 	if (ia64_pal_vm_summary(NULL, &vmi) == 0)
802 		max_ctx = (1U << (vmi.pal_vm_info_2_s.rid_size - 3)) - 1;
803 	else {
804 		printk(KERN_WARNING "cpu_init: PAL VM summary failed, assuming 18 RID bits\n");
805 		max_ctx = (1U << 15) - 1;	/* use architected minimum */
806 	}
807 	while (max_ctx < ia64_ctx.max_ctx) {
808 		unsigned int old = ia64_ctx.max_ctx;
809 		if (cmpxchg(&ia64_ctx.max_ctx, old, max_ctx) == old)
810 			break;
811 	}
812 
813 	if (ia64_pal_rse_info(&num_phys_stacked, NULL) != 0) {
814 		printk(KERN_WARNING "cpu_init: PAL RSE info failed; assuming 96 physical "
815 		       "stacked regs\n");
816 		num_phys_stacked = 96;
817 	}
818 	/* size of physical stacked register partition plus 8 bytes: */
819 	__get_cpu_var(ia64_phys_stacked_size_p8) = num_phys_stacked*8 + 8;
820 	platform_cpu_init();
821 	pm_idle = default_idle;
822 }
823 
824 void
825 check_bugs (void)
826 {
827 	ia64_patch_mckinley_e9((unsigned long) __start___mckinley_e9_bundles,
828 			       (unsigned long) __end___mckinley_e9_bundles);
829 }
830