xref: /openbmc/linux/arch/ia64/kernel/setup.c (revision 44c45120)
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 	unsigned long phys_iobase;
248 
249 	/*
250 	 * Set `iobase' based on the EFI memory map or, failing that, the
251 	 * value firmware left in ar.k0.
252 	 *
253 	 * Note that in ia32 mode, IN/OUT instructions use ar.k0 to compute
254 	 * the port's virtual address, so ia32_load_state() loads it with a
255 	 * user virtual address.  But in ia64 mode, glibc uses the
256 	 * *physical* address in ar.k0 to mmap the appropriate area from
257 	 * /dev/mem, and the inX()/outX() interfaces use MMIO.  In both
258 	 * cases, user-mode can only use the legacy 0-64K I/O port space.
259 	 *
260 	 * ar.k0 is not involved in kernel I/O port accesses, which can use
261 	 * any of the I/O port spaces and are done via MMIO using the
262 	 * virtual mmio_base from the appropriate io_space[].
263 	 */
264 	phys_iobase = efi_get_iobase();
265 	if (!phys_iobase) {
266 		phys_iobase = ia64_get_kr(IA64_KR_IO_BASE);
267 		printk(KERN_INFO "No I/O port range found in EFI memory map, "
268 			"falling back to AR.KR0 (0x%lx)\n", phys_iobase);
269 	}
270 	ia64_iobase = (unsigned long) ioremap(phys_iobase, 0);
271 	ia64_set_kr(IA64_KR_IO_BASE, __pa(ia64_iobase));
272 
273 	/* setup legacy IO port space */
274 	io_space[0].mmio_base = ia64_iobase;
275 	io_space[0].sparse = 1;
276 	num_io_spaces = 1;
277 }
278 
279 /**
280  * early_console_setup - setup debugging console
281  *
282  * Consoles started here require little enough setup that we can start using
283  * them very early in the boot process, either right after the machine
284  * vector initialization, or even before if the drivers can detect their hw.
285  *
286  * Returns non-zero if a console couldn't be setup.
287  */
288 static inline int __init
289 early_console_setup (char *cmdline)
290 {
291 	int earlycons = 0;
292 
293 #ifdef CONFIG_SERIAL_SGI_L1_CONSOLE
294 	{
295 		extern int sn_serial_console_early_setup(void);
296 		if (!sn_serial_console_early_setup())
297 			earlycons++;
298 	}
299 #endif
300 #ifdef CONFIG_EFI_PCDP
301 	if (!efi_setup_pcdp_console(cmdline))
302 		earlycons++;
303 #endif
304 #ifdef CONFIG_SERIAL_8250_CONSOLE
305 	if (!early_serial_console_init(cmdline))
306 		earlycons++;
307 #endif
308 
309 	return (earlycons) ? 0 : -1;
310 }
311 
312 static inline void
313 mark_bsp_online (void)
314 {
315 #ifdef CONFIG_SMP
316 	/* If we register an early console, allow CPU 0 to printk */
317 	cpu_set(smp_processor_id(), cpu_online_map);
318 #endif
319 }
320 
321 #ifdef CONFIG_SMP
322 static void
323 check_for_logical_procs (void)
324 {
325 	pal_logical_to_physical_t info;
326 	s64 status;
327 
328 	status = ia64_pal_logical_to_phys(0, &info);
329 	if (status == -1) {
330 		printk(KERN_INFO "No logical to physical processor mapping "
331 		       "available\n");
332 		return;
333 	}
334 	if (status) {
335 		printk(KERN_ERR "ia64_pal_logical_to_phys failed with %ld\n",
336 		       status);
337 		return;
338 	}
339 	/*
340 	 * Total number of siblings that BSP has.  Though not all of them
341 	 * may have booted successfully. The correct number of siblings
342 	 * booted is in info.overview_num_log.
343 	 */
344 	smp_num_siblings = info.overview_tpc;
345 	smp_num_cpucores = info.overview_cpp;
346 }
347 #endif
348 
349 void __init
350 setup_arch (char **cmdline_p)
351 {
352 	unw_init();
353 
354 	ia64_patch_vtop((u64) __start___vtop_patchlist, (u64) __end___vtop_patchlist);
355 
356 	*cmdline_p = __va(ia64_boot_param->command_line);
357 	strlcpy(saved_command_line, *cmdline_p, COMMAND_LINE_SIZE);
358 
359 	efi_init();
360 	io_port_init();
361 
362 #ifdef CONFIG_IA64_GENERIC
363 	{
364 		const char *mvec_name = strstr (*cmdline_p, "machvec=");
365 		char str[64];
366 
367 		if (mvec_name) {
368 			const char *end;
369 			size_t len;
370 
371 			mvec_name += 8;
372 			end = strchr (mvec_name, ' ');
373 			if (end)
374 				len = end - mvec_name;
375 			else
376 				len = strlen (mvec_name);
377 			len = min(len, sizeof (str) - 1);
378 			strncpy (str, mvec_name, len);
379 			str[len] = '\0';
380 			mvec_name = str;
381 		} else
382 			mvec_name = acpi_get_sysname();
383 		machvec_init(mvec_name);
384 	}
385 #endif
386 
387 	if (early_console_setup(*cmdline_p) == 0)
388 		mark_bsp_online();
389 
390 #ifdef CONFIG_ACPI
391 	/* Initialize the ACPI boot-time table parser */
392 	acpi_table_init();
393 # ifdef CONFIG_ACPI_NUMA
394 	acpi_numa_init();
395 # endif
396 #else
397 # ifdef CONFIG_SMP
398 	smp_build_cpu_map();	/* happens, e.g., with the Ski simulator */
399 # endif
400 #endif /* CONFIG_APCI_BOOT */
401 
402 	find_memory();
403 
404 	/* process SAL system table: */
405 	ia64_sal_init(efi.sal_systab);
406 
407 #ifdef CONFIG_SMP
408 	cpu_physical_id(0) = hard_smp_processor_id();
409 
410 	cpu_set(0, cpu_sibling_map[0]);
411 	cpu_set(0, cpu_core_map[0]);
412 
413 	check_for_logical_procs();
414 	if (smp_num_cpucores > 1)
415 		printk(KERN_INFO
416 		       "cpu package is Multi-Core capable: number of cores=%d\n",
417 		       smp_num_cpucores);
418 	if (smp_num_siblings > 1)
419 		printk(KERN_INFO
420 		       "cpu package is Multi-Threading capable: number of siblings=%d\n",
421 		       smp_num_siblings);
422 #endif
423 
424 	cpu_init();	/* initialize the bootstrap CPU */
425 
426 #ifdef CONFIG_ACPI
427 	acpi_boot_init();
428 #endif
429 
430 #ifdef CONFIG_VT
431 	if (!conswitchp) {
432 # if defined(CONFIG_DUMMY_CONSOLE)
433 		conswitchp = &dummy_con;
434 # endif
435 # if defined(CONFIG_VGA_CONSOLE)
436 		/*
437 		 * Non-legacy systems may route legacy VGA MMIO range to system
438 		 * memory.  vga_con probes the MMIO hole, so memory looks like
439 		 * a VGA device to it.  The EFI memory map can tell us if it's
440 		 * memory so we can avoid this problem.
441 		 */
442 		if (efi_mem_type(0xA0000) != EFI_CONVENTIONAL_MEMORY)
443 			conswitchp = &vga_con;
444 # endif
445 	}
446 #endif
447 
448 	/* enable IA-64 Machine Check Abort Handling unless disabled */
449 	if (!strstr(saved_command_line, "nomca"))
450 		ia64_mca_init();
451 
452 	platform_setup(cmdline_p);
453 	paging_init();
454 }
455 
456 /*
457  * Display cpu info for all cpu's.
458  */
459 static int
460 show_cpuinfo (struct seq_file *m, void *v)
461 {
462 #ifdef CONFIG_SMP
463 #	define lpj	c->loops_per_jiffy
464 #	define cpunum	c->cpu
465 #else
466 #	define lpj	loops_per_jiffy
467 #	define cpunum	0
468 #endif
469 	static struct {
470 		unsigned long mask;
471 		const char *feature_name;
472 	} feature_bits[] = {
473 		{ 1UL << 0, "branchlong" },
474 		{ 1UL << 1, "spontaneous deferral"},
475 		{ 1UL << 2, "16-byte atomic ops" }
476 	};
477 	char family[32], features[128], *cp, sep;
478 	struct cpuinfo_ia64 *c = v;
479 	unsigned long mask;
480 	int i;
481 
482 	mask = c->features;
483 
484 	switch (c->family) {
485 	      case 0x07:	memcpy(family, "Itanium", 8); break;
486 	      case 0x1f:	memcpy(family, "Itanium 2", 10); break;
487 	      default:		sprintf(family, "%u", c->family); break;
488 	}
489 
490 	/* build the feature string: */
491 	memcpy(features, " standard", 10);
492 	cp = features;
493 	sep = 0;
494 	for (i = 0; i < (int) ARRAY_SIZE(feature_bits); ++i) {
495 		if (mask & feature_bits[i].mask) {
496 			if (sep)
497 				*cp++ = sep;
498 			sep = ',';
499 			*cp++ = ' ';
500 			strcpy(cp, feature_bits[i].feature_name);
501 			cp += strlen(feature_bits[i].feature_name);
502 			mask &= ~feature_bits[i].mask;
503 		}
504 	}
505 	if (mask) {
506 		/* print unknown features as a hex value: */
507 		if (sep)
508 			*cp++ = sep;
509 		sprintf(cp, " 0x%lx", mask);
510 	}
511 
512 	seq_printf(m,
513 		   "processor  : %d\n"
514 		   "vendor     : %s\n"
515 		   "arch       : IA-64\n"
516 		   "family     : %s\n"
517 		   "model      : %u\n"
518 		   "revision   : %u\n"
519 		   "archrev    : %u\n"
520 		   "features   :%s\n"	/* don't change this---it _is_ right! */
521 		   "cpu number : %lu\n"
522 		   "cpu regs   : %u\n"
523 		   "cpu MHz    : %lu.%06lu\n"
524 		   "itc MHz    : %lu.%06lu\n"
525 		   "BogoMIPS   : %lu.%02lu\n",
526 		   cpunum, c->vendor, family, c->model, c->revision, c->archrev,
527 		   features, c->ppn, c->number,
528 		   c->proc_freq / 1000000, c->proc_freq % 1000000,
529 		   c->itc_freq / 1000000, c->itc_freq % 1000000,
530 		   lpj*HZ/500000, (lpj*HZ/5000) % 100);
531 #ifdef CONFIG_SMP
532 	seq_printf(m, "siblings   : %u\n", c->num_log);
533 	if (c->threads_per_core > 1 || c->cores_per_socket > 1)
534 		seq_printf(m,
535 		   	   "physical id: %u\n"
536 		   	   "core id    : %u\n"
537 		   	   "thread id  : %u\n",
538 		   	   c->socket_id, c->core_id, c->thread_id);
539 #endif
540 	seq_printf(m,"\n");
541 
542 	return 0;
543 }
544 
545 static void *
546 c_start (struct seq_file *m, loff_t *pos)
547 {
548 #ifdef CONFIG_SMP
549 	while (*pos < NR_CPUS && !cpu_isset(*pos, cpu_online_map))
550 		++*pos;
551 #endif
552 	return *pos < NR_CPUS ? cpu_data(*pos) : NULL;
553 }
554 
555 static void *
556 c_next (struct seq_file *m, void *v, loff_t *pos)
557 {
558 	++*pos;
559 	return c_start(m, pos);
560 }
561 
562 static void
563 c_stop (struct seq_file *m, void *v)
564 {
565 }
566 
567 struct seq_operations cpuinfo_op = {
568 	.start =	c_start,
569 	.next =		c_next,
570 	.stop =		c_stop,
571 	.show =		show_cpuinfo
572 };
573 
574 void
575 identify_cpu (struct cpuinfo_ia64 *c)
576 {
577 	union {
578 		unsigned long bits[5];
579 		struct {
580 			/* id 0 & 1: */
581 			char vendor[16];
582 
583 			/* id 2 */
584 			u64 ppn;		/* processor serial number */
585 
586 			/* id 3: */
587 			unsigned number		:  8;
588 			unsigned revision	:  8;
589 			unsigned model		:  8;
590 			unsigned family		:  8;
591 			unsigned archrev	:  8;
592 			unsigned reserved	: 24;
593 
594 			/* id 4: */
595 			u64 features;
596 		} field;
597 	} cpuid;
598 	pal_vm_info_1_u_t vm1;
599 	pal_vm_info_2_u_t vm2;
600 	pal_status_t status;
601 	unsigned long impl_va_msb = 50, phys_addr_size = 44;	/* Itanium defaults */
602 	int i;
603 
604 	for (i = 0; i < 5; ++i)
605 		cpuid.bits[i] = ia64_get_cpuid(i);
606 
607 	memcpy(c->vendor, cpuid.field.vendor, 16);
608 #ifdef CONFIG_SMP
609 	c->cpu = smp_processor_id();
610 
611 	/* below default values will be overwritten  by identify_siblings()
612 	 * for Multi-Threading/Multi-Core capable cpu's
613 	 */
614 	c->threads_per_core = c->cores_per_socket = c->num_log = 1;
615 	c->socket_id = -1;
616 
617 	identify_siblings(c);
618 #endif
619 	c->ppn = cpuid.field.ppn;
620 	c->number = cpuid.field.number;
621 	c->revision = cpuid.field.revision;
622 	c->model = cpuid.field.model;
623 	c->family = cpuid.field.family;
624 	c->archrev = cpuid.field.archrev;
625 	c->features = cpuid.field.features;
626 
627 	status = ia64_pal_vm_summary(&vm1, &vm2);
628 	if (status == PAL_STATUS_SUCCESS) {
629 		impl_va_msb = vm2.pal_vm_info_2_s.impl_va_msb;
630 		phys_addr_size = vm1.pal_vm_info_1_s.phys_add_size;
631 	}
632 	c->unimpl_va_mask = ~((7L<<61) | ((1L << (impl_va_msb + 1)) - 1));
633 	c->unimpl_pa_mask = ~((1L<<63) | ((1L << phys_addr_size) - 1));
634 }
635 
636 void
637 setup_per_cpu_areas (void)
638 {
639 	/* start_kernel() requires this... */
640 }
641 
642 /*
643  * Calculate the max. cache line size.
644  *
645  * In addition, the minimum of the i-cache stride sizes is calculated for
646  * "flush_icache_range()".
647  */
648 static void
649 get_max_cacheline_size (void)
650 {
651 	unsigned long line_size, max = 1;
652 	u64 l, levels, unique_caches;
653         pal_cache_config_info_t cci;
654         s64 status;
655 
656         status = ia64_pal_cache_summary(&levels, &unique_caches);
657         if (status != 0) {
658                 printk(KERN_ERR "%s: ia64_pal_cache_summary() failed (status=%ld)\n",
659                        __FUNCTION__, status);
660                 max = SMP_CACHE_BYTES;
661 		/* Safest setup for "flush_icache_range()" */
662 		ia64_i_cache_stride_shift = I_CACHE_STRIDE_SHIFT;
663 		goto out;
664         }
665 
666 	for (l = 0; l < levels; ++l) {
667 		status = ia64_pal_cache_config_info(l, /* cache_type (data_or_unified)= */ 2,
668 						    &cci);
669 		if (status != 0) {
670 			printk(KERN_ERR
671 			       "%s: ia64_pal_cache_config_info(l=%lu, 2) failed (status=%ld)\n",
672 			       __FUNCTION__, l, status);
673 			max = SMP_CACHE_BYTES;
674 			/* The safest setup for "flush_icache_range()" */
675 			cci.pcci_stride = I_CACHE_STRIDE_SHIFT;
676 			cci.pcci_unified = 1;
677 		}
678 		line_size = 1 << cci.pcci_line_size;
679 		if (line_size > max)
680 			max = line_size;
681 		if (!cci.pcci_unified) {
682 			status = ia64_pal_cache_config_info(l,
683 						    /* cache_type (instruction)= */ 1,
684 						    &cci);
685 			if (status != 0) {
686 				printk(KERN_ERR
687 				"%s: ia64_pal_cache_config_info(l=%lu, 1) failed (status=%ld)\n",
688 					__FUNCTION__, l, status);
689 				/* The safest setup for "flush_icache_range()" */
690 				cci.pcci_stride = I_CACHE_STRIDE_SHIFT;
691 			}
692 		}
693 		if (cci.pcci_stride < ia64_i_cache_stride_shift)
694 			ia64_i_cache_stride_shift = cci.pcci_stride;
695 	}
696   out:
697 	if (max > ia64_max_cacheline_size)
698 		ia64_max_cacheline_size = max;
699 }
700 
701 /*
702  * cpu_init() initializes state that is per-CPU.  This function acts
703  * as a 'CPU state barrier', nothing should get across.
704  */
705 void
706 cpu_init (void)
707 {
708 	extern void __devinit ia64_mmu_init (void *);
709 	unsigned long num_phys_stacked;
710 	pal_vm_info_2_u_t vmi;
711 	unsigned int max_ctx;
712 	struct cpuinfo_ia64 *cpu_info;
713 	void *cpu_data;
714 
715 	cpu_data = per_cpu_init();
716 
717 	/*
718 	 * We set ar.k3 so that assembly code in MCA handler can compute
719 	 * physical addresses of per cpu variables with a simple:
720 	 *   phys = ar.k3 + &per_cpu_var
721 	 */
722 	ia64_set_kr(IA64_KR_PER_CPU_DATA,
723 		    ia64_tpa(cpu_data) - (long) __per_cpu_start);
724 
725 	get_max_cacheline_size();
726 
727 	/*
728 	 * We can't pass "local_cpu_data" to identify_cpu() because we haven't called
729 	 * ia64_mmu_init() yet.  And we can't call ia64_mmu_init() first because it
730 	 * depends on the data returned by identify_cpu().  We break the dependency by
731 	 * accessing cpu_data() through the canonical per-CPU address.
732 	 */
733 	cpu_info = cpu_data + ((char *) &__ia64_per_cpu_var(cpu_info) - __per_cpu_start);
734 	identify_cpu(cpu_info);
735 
736 #ifdef CONFIG_MCKINLEY
737 	{
738 #		define FEATURE_SET 16
739 		struct ia64_pal_retval iprv;
740 
741 		if (cpu_info->family == 0x1f) {
742 			PAL_CALL_PHYS(iprv, PAL_PROC_GET_FEATURES, 0, FEATURE_SET, 0);
743 			if ((iprv.status == 0) && (iprv.v0 & 0x80) && (iprv.v2 & 0x80))
744 				PAL_CALL_PHYS(iprv, PAL_PROC_SET_FEATURES,
745 				              (iprv.v1 | 0x80), FEATURE_SET, 0);
746 		}
747 	}
748 #endif
749 
750 	/* Clear the stack memory reserved for pt_regs: */
751 	memset(ia64_task_regs(current), 0, sizeof(struct pt_regs));
752 
753 	ia64_set_kr(IA64_KR_FPU_OWNER, 0);
754 
755 	/*
756 	 * Initialize the page-table base register to a global
757 	 * directory with all zeroes.  This ensure that we can handle
758 	 * TLB-misses to user address-space even before we created the
759 	 * first user address-space.  This may happen, e.g., due to
760 	 * aggressive use of lfetch.fault.
761 	 */
762 	ia64_set_kr(IA64_KR_PT_BASE, __pa(ia64_imva(empty_zero_page)));
763 
764 	/*
765 	 * Initialize default control register to defer speculative faults except
766 	 * for those arising from TLB misses, which are not deferred.  The
767 	 * kernel MUST NOT depend on a particular setting of these bits (in other words,
768 	 * the kernel must have recovery code for all speculative accesses).  Turn on
769 	 * dcr.lc as per recommendation by the architecture team.  Most IA-32 apps
770 	 * shouldn't be affected by this (moral: keep your ia32 locks aligned and you'll
771 	 * be fine).
772 	 */
773 	ia64_setreg(_IA64_REG_CR_DCR,  (  IA64_DCR_DP | IA64_DCR_DK | IA64_DCR_DX | IA64_DCR_DR
774 					| IA64_DCR_DA | IA64_DCR_DD | IA64_DCR_LC));
775 	atomic_inc(&init_mm.mm_count);
776 	current->active_mm = &init_mm;
777 	if (current->mm)
778 		BUG();
779 
780 	ia64_mmu_init(ia64_imva(cpu_data));
781 	ia64_mca_cpu_init(ia64_imva(cpu_data));
782 
783 #ifdef CONFIG_IA32_SUPPORT
784 	ia32_cpu_init();
785 #endif
786 
787 	/* Clear ITC to eliminiate sched_clock() overflows in human time.  */
788 	ia64_set_itc(0);
789 
790 	/* disable all local interrupt sources: */
791 	ia64_set_itv(1 << 16);
792 	ia64_set_lrr0(1 << 16);
793 	ia64_set_lrr1(1 << 16);
794 	ia64_setreg(_IA64_REG_CR_PMV, 1 << 16);
795 	ia64_setreg(_IA64_REG_CR_CMCV, 1 << 16);
796 
797 	/* clear TPR & XTP to enable all interrupt classes: */
798 	ia64_setreg(_IA64_REG_CR_TPR, 0);
799 #ifdef CONFIG_SMP
800 	normal_xtp();
801 #endif
802 
803 	/* set ia64_ctx.max_rid to the maximum RID that is supported by all CPUs: */
804 	if (ia64_pal_vm_summary(NULL, &vmi) == 0)
805 		max_ctx = (1U << (vmi.pal_vm_info_2_s.rid_size - 3)) - 1;
806 	else {
807 		printk(KERN_WARNING "cpu_init: PAL VM summary failed, assuming 18 RID bits\n");
808 		max_ctx = (1U << 15) - 1;	/* use architected minimum */
809 	}
810 	while (max_ctx < ia64_ctx.max_ctx) {
811 		unsigned int old = ia64_ctx.max_ctx;
812 		if (cmpxchg(&ia64_ctx.max_ctx, old, max_ctx) == old)
813 			break;
814 	}
815 
816 	if (ia64_pal_rse_info(&num_phys_stacked, NULL) != 0) {
817 		printk(KERN_WARNING "cpu_init: PAL RSE info failed; assuming 96 physical "
818 		       "stacked regs\n");
819 		num_phys_stacked = 96;
820 	}
821 	/* size of physical stacked register partition plus 8 bytes: */
822 	__get_cpu_var(ia64_phys_stacked_size_p8) = num_phys_stacked*8 + 8;
823 	platform_cpu_init();
824 	pm_idle = default_idle;
825 }
826 
827 void
828 check_bugs (void)
829 {
830 	ia64_patch_mckinley_e9((unsigned long) __start___mckinley_e9_bundles,
831 			       (unsigned long) __end___mckinley_e9_bundles);
832 }
833