xref: /openbmc/linux/arch/powerpc/kernel/setup_64.c (revision bbecb07f)
1 /*
2  *
3  * Common boot and setup code.
4  *
5  * Copyright (C) 2001 PPC64 Team, IBM Corp
6  *
7  *      This program is free software; you can redistribute it and/or
8  *      modify it under the terms of the GNU General Public License
9  *      as published by the Free Software Foundation; either version
10  *      2 of the License, or (at your option) any later version.
11  */
12 
13 #define DEBUG
14 
15 #include <linux/export.h>
16 #include <linux/string.h>
17 #include <linux/sched.h>
18 #include <linux/init.h>
19 #include <linux/kernel.h>
20 #include <linux/reboot.h>
21 #include <linux/delay.h>
22 #include <linux/initrd.h>
23 #include <linux/seq_file.h>
24 #include <linux/ioport.h>
25 #include <linux/console.h>
26 #include <linux/utsname.h>
27 #include <linux/tty.h>
28 #include <linux/root_dev.h>
29 #include <linux/notifier.h>
30 #include <linux/cpu.h>
31 #include <linux/unistd.h>
32 #include <linux/serial.h>
33 #include <linux/serial_8250.h>
34 #include <linux/bootmem.h>
35 #include <linux/pci.h>
36 #include <linux/lockdep.h>
37 #include <linux/memblock.h>
38 #include <linux/memory.h>
39 #include <linux/nmi.h>
40 
41 #include <asm/io.h>
42 #include <asm/kdump.h>
43 #include <asm/prom.h>
44 #include <asm/processor.h>
45 #include <asm/pgtable.h>
46 #include <asm/smp.h>
47 #include <asm/elf.h>
48 #include <asm/machdep.h>
49 #include <asm/paca.h>
50 #include <asm/time.h>
51 #include <asm/cputable.h>
52 #include <asm/dt_cpu_ftrs.h>
53 #include <asm/sections.h>
54 #include <asm/btext.h>
55 #include <asm/nvram.h>
56 #include <asm/setup.h>
57 #include <asm/rtas.h>
58 #include <asm/iommu.h>
59 #include <asm/serial.h>
60 #include <asm/cache.h>
61 #include <asm/page.h>
62 #include <asm/mmu.h>
63 #include <asm/firmware.h>
64 #include <asm/xmon.h>
65 #include <asm/udbg.h>
66 #include <asm/kexec.h>
67 #include <asm/code-patching.h>
68 #include <asm/livepatch.h>
69 #include <asm/opal.h>
70 #include <asm/cputhreads.h>
71 
72 #include "setup.h"
73 
74 #ifdef DEBUG
75 #define DBG(fmt...) udbg_printf(fmt)
76 #else
77 #define DBG(fmt...)
78 #endif
79 
80 int spinning_secondaries;
81 u64 ppc64_pft_size;
82 
83 struct ppc64_caches ppc64_caches = {
84 	.l1d = {
85 		.block_size = 0x40,
86 		.log_block_size = 6,
87 	},
88 	.l1i = {
89 		.block_size = 0x40,
90 		.log_block_size = 6
91 	},
92 };
93 EXPORT_SYMBOL_GPL(ppc64_caches);
94 
95 #if defined(CONFIG_PPC_BOOK3E) && defined(CONFIG_SMP)
96 void __init setup_tlb_core_data(void)
97 {
98 	int cpu;
99 
100 	BUILD_BUG_ON(offsetof(struct tlb_core_data, lock) != 0);
101 
102 	for_each_possible_cpu(cpu) {
103 		int first = cpu_first_thread_sibling(cpu);
104 
105 		/*
106 		 * If we boot via kdump on a non-primary thread,
107 		 * make sure we point at the thread that actually
108 		 * set up this TLB.
109 		 */
110 		if (cpu_first_thread_sibling(boot_cpuid) == first)
111 			first = boot_cpuid;
112 
113 		paca[cpu].tcd_ptr = &paca[first].tcd;
114 
115 		/*
116 		 * If we have threads, we need either tlbsrx.
117 		 * or e6500 tablewalk mode, or else TLB handlers
118 		 * will be racy and could produce duplicate entries.
119 		 * Should we panic instead?
120 		 */
121 		WARN_ONCE(smt_enabled_at_boot >= 2 &&
122 			  !mmu_has_feature(MMU_FTR_USE_TLBRSRV) &&
123 			  book3e_htw_mode != PPC_HTW_E6500,
124 			  "%s: unsupported MMU configuration\n", __func__);
125 	}
126 }
127 #endif
128 
129 #ifdef CONFIG_SMP
130 
131 static char *smt_enabled_cmdline;
132 
133 /* Look for ibm,smt-enabled OF option */
134 void __init check_smt_enabled(void)
135 {
136 	struct device_node *dn;
137 	const char *smt_option;
138 
139 	/* Default to enabling all threads */
140 	smt_enabled_at_boot = threads_per_core;
141 
142 	/* Allow the command line to overrule the OF option */
143 	if (smt_enabled_cmdline) {
144 		if (!strcmp(smt_enabled_cmdline, "on"))
145 			smt_enabled_at_boot = threads_per_core;
146 		else if (!strcmp(smt_enabled_cmdline, "off"))
147 			smt_enabled_at_boot = 0;
148 		else {
149 			int smt;
150 			int rc;
151 
152 			rc = kstrtoint(smt_enabled_cmdline, 10, &smt);
153 			if (!rc)
154 				smt_enabled_at_boot =
155 					min(threads_per_core, smt);
156 		}
157 	} else {
158 		dn = of_find_node_by_path("/options");
159 		if (dn) {
160 			smt_option = of_get_property(dn, "ibm,smt-enabled",
161 						     NULL);
162 
163 			if (smt_option) {
164 				if (!strcmp(smt_option, "on"))
165 					smt_enabled_at_boot = threads_per_core;
166 				else if (!strcmp(smt_option, "off"))
167 					smt_enabled_at_boot = 0;
168 			}
169 
170 			of_node_put(dn);
171 		}
172 	}
173 }
174 
175 /* Look for smt-enabled= cmdline option */
176 static int __init early_smt_enabled(char *p)
177 {
178 	smt_enabled_cmdline = p;
179 	return 0;
180 }
181 early_param("smt-enabled", early_smt_enabled);
182 
183 #endif /* CONFIG_SMP */
184 
185 /** Fix up paca fields required for the boot cpu */
186 static void __init fixup_boot_paca(void)
187 {
188 	/* The boot cpu is started */
189 	get_paca()->cpu_start = 1;
190 	/* Allow percpu accesses to work until we setup percpu data */
191 	get_paca()->data_offset = 0;
192 }
193 
194 static void __init configure_exceptions(void)
195 {
196 	/*
197 	 * Setup the trampolines from the lowmem exception vectors
198 	 * to the kdump kernel when not using a relocatable kernel.
199 	 */
200 	setup_kdump_trampoline();
201 
202 	/* Under a PAPR hypervisor, we need hypercalls */
203 	if (firmware_has_feature(FW_FEATURE_SET_MODE)) {
204 		/* Enable AIL if possible */
205 		pseries_enable_reloc_on_exc();
206 
207 		/*
208 		 * Tell the hypervisor that we want our exceptions to
209 		 * be taken in little endian mode.
210 		 *
211 		 * We don't call this for big endian as our calling convention
212 		 * makes us always enter in BE, and the call may fail under
213 		 * some circumstances with kdump.
214 		 */
215 #ifdef __LITTLE_ENDIAN__
216 		pseries_little_endian_exceptions();
217 #endif
218 	} else {
219 		/* Set endian mode using OPAL */
220 		if (firmware_has_feature(FW_FEATURE_OPAL))
221 			opal_configure_cores();
222 
223 		/* AIL on native is done in cpu_ready_for_interrupts() */
224 	}
225 }
226 
227 static void cpu_ready_for_interrupts(void)
228 {
229 	/*
230 	 * Enable AIL if supported, and we are in hypervisor mode. This
231 	 * is called once for every processor.
232 	 *
233 	 * If we are not in hypervisor mode the job is done once for
234 	 * the whole partition in configure_exceptions().
235 	 */
236 	if (cpu_has_feature(CPU_FTR_HVMODE) &&
237 	    cpu_has_feature(CPU_FTR_ARCH_207S)) {
238 		unsigned long lpcr = mfspr(SPRN_LPCR);
239 		mtspr(SPRN_LPCR, lpcr | LPCR_AIL_3);
240 	}
241 
242 	/*
243 	 * Fixup HFSCR:TM based on CPU features. The bit is set by our
244 	 * early asm init because at that point we haven't updated our
245 	 * CPU features from firmware and device-tree. Here we have,
246 	 * so let's do it.
247 	 */
248 	if (cpu_has_feature(CPU_FTR_HVMODE) && !cpu_has_feature(CPU_FTR_TM_COMP))
249 		mtspr(SPRN_HFSCR, mfspr(SPRN_HFSCR) & ~HFSCR_TM);
250 
251 	/* Set IR and DR in PACA MSR */
252 	get_paca()->kernel_msr = MSR_KERNEL;
253 }
254 
255 /*
256  * Early initialization entry point. This is called by head.S
257  * with MMU translation disabled. We rely on the "feature" of
258  * the CPU that ignores the top 2 bits of the address in real
259  * mode so we can access kernel globals normally provided we
260  * only toy with things in the RMO region. From here, we do
261  * some early parsing of the device-tree to setup out MEMBLOCK
262  * data structures, and allocate & initialize the hash table
263  * and segment tables so we can start running with translation
264  * enabled.
265  *
266  * It is this function which will call the probe() callback of
267  * the various platform types and copy the matching one to the
268  * global ppc_md structure. Your platform can eventually do
269  * some very early initializations from the probe() routine, but
270  * this is not recommended, be very careful as, for example, the
271  * device-tree is not accessible via normal means at this point.
272  */
273 
274 void __init early_setup(unsigned long dt_ptr)
275 {
276 	static __initdata struct paca_struct boot_paca;
277 
278 	/* -------- printk is _NOT_ safe to use here ! ------- */
279 
280 	/* Try new device tree based feature discovery ... */
281 	if (!dt_cpu_ftrs_init(__va(dt_ptr)))
282 		/* Otherwise use the old style CPU table */
283 		identify_cpu(0, mfspr(SPRN_PVR));
284 
285 	/* Assume we're on cpu 0 for now. Don't write to the paca yet! */
286 	initialise_paca(&boot_paca, 0);
287 	setup_paca(&boot_paca);
288 	fixup_boot_paca();
289 
290 	/* -------- printk is now safe to use ------- */
291 
292 	/* Enable early debugging if any specified (see udbg.h) */
293 	udbg_early_init();
294 
295  	DBG(" -> early_setup(), dt_ptr: 0x%lx\n", dt_ptr);
296 
297 	/*
298 	 * Do early initialization using the flattened device
299 	 * tree, such as retrieving the physical memory map or
300 	 * calculating/retrieving the hash table size.
301 	 */
302 	early_init_devtree(__va(dt_ptr));
303 
304 	/* Now we know the logical id of our boot cpu, setup the paca. */
305 	setup_paca(&paca[boot_cpuid]);
306 	fixup_boot_paca();
307 
308 	/*
309 	 * Configure exception handlers. This include setting up trampolines
310 	 * if needed, setting exception endian mode, etc...
311 	 */
312 	configure_exceptions();
313 
314 	/* Apply all the dynamic patching */
315 	apply_feature_fixups();
316 	setup_feature_keys();
317 
318 	/* Initialize the hash table or TLB handling */
319 	early_init_mmu();
320 
321 	/*
322 	 * After firmware and early platform setup code has set things up,
323 	 * we note the SPR values for configurable control/performance
324 	 * registers, and use those as initial defaults.
325 	 */
326 	record_spr_defaults();
327 
328 	/*
329 	 * At this point, we can let interrupts switch to virtual mode
330 	 * (the MMU has been setup), so adjust the MSR in the PACA to
331 	 * have IR and DR set and enable AIL if it exists
332 	 */
333 	cpu_ready_for_interrupts();
334 
335 	DBG(" <- early_setup()\n");
336 
337 #ifdef CONFIG_PPC_EARLY_DEBUG_BOOTX
338 	/*
339 	 * This needs to be done *last* (after the above DBG() even)
340 	 *
341 	 * Right after we return from this function, we turn on the MMU
342 	 * which means the real-mode access trick that btext does will
343 	 * no longer work, it needs to switch to using a real MMU
344 	 * mapping. This call will ensure that it does
345 	 */
346 	btext_map();
347 #endif /* CONFIG_PPC_EARLY_DEBUG_BOOTX */
348 }
349 
350 #ifdef CONFIG_SMP
351 void early_setup_secondary(void)
352 {
353 	/* Mark interrupts disabled in PACA */
354 	get_paca()->soft_enabled = 0;
355 
356 	/* Initialize the hash table or TLB handling */
357 	early_init_mmu_secondary();
358 
359 	/*
360 	 * At this point, we can let interrupts switch to virtual mode
361 	 * (the MMU has been setup), so adjust the MSR in the PACA to
362 	 * have IR and DR set.
363 	 */
364 	cpu_ready_for_interrupts();
365 }
366 
367 #endif /* CONFIG_SMP */
368 
369 #if defined(CONFIG_SMP) || defined(CONFIG_KEXEC_CORE)
370 static bool use_spinloop(void)
371 {
372 	if (IS_ENABLED(CONFIG_PPC_BOOK3S)) {
373 		/*
374 		 * See comments in head_64.S -- not all platforms insert
375 		 * secondaries at __secondary_hold and wait at the spin
376 		 * loop.
377 		 */
378 		if (firmware_has_feature(FW_FEATURE_OPAL))
379 			return false;
380 		return true;
381 	}
382 
383 	/*
384 	 * When book3e boots from kexec, the ePAPR spin table does
385 	 * not get used.
386 	 */
387 	return of_property_read_bool(of_chosen, "linux,booted-from-kexec");
388 }
389 
390 void smp_release_cpus(void)
391 {
392 	unsigned long *ptr;
393 	int i;
394 
395 	if (!use_spinloop())
396 		return;
397 
398 	DBG(" -> smp_release_cpus()\n");
399 
400 	/* All secondary cpus are spinning on a common spinloop, release them
401 	 * all now so they can start to spin on their individual paca
402 	 * spinloops. For non SMP kernels, the secondary cpus never get out
403 	 * of the common spinloop.
404 	 */
405 
406 	ptr  = (unsigned long *)((unsigned long)&__secondary_hold_spinloop
407 			- PHYSICAL_START);
408 	*ptr = ppc_function_entry(generic_secondary_smp_init);
409 
410 	/* And wait a bit for them to catch up */
411 	for (i = 0; i < 100000; i++) {
412 		mb();
413 		HMT_low();
414 		if (spinning_secondaries == 0)
415 			break;
416 		udelay(1);
417 	}
418 	DBG("spinning_secondaries = %d\n", spinning_secondaries);
419 
420 	DBG(" <- smp_release_cpus()\n");
421 }
422 #endif /* CONFIG_SMP || CONFIG_KEXEC_CORE */
423 
424 /*
425  * Initialize some remaining members of the ppc64_caches and systemcfg
426  * structures
427  * (at least until we get rid of them completely). This is mostly some
428  * cache informations about the CPU that will be used by cache flush
429  * routines and/or provided to userland
430  */
431 
432 static void init_cache_info(struct ppc_cache_info *info, u32 size, u32 lsize,
433 			    u32 bsize, u32 sets)
434 {
435 	info->size = size;
436 	info->sets = sets;
437 	info->line_size = lsize;
438 	info->block_size = bsize;
439 	info->log_block_size = __ilog2(bsize);
440 	if (bsize)
441 		info->blocks_per_page = PAGE_SIZE / bsize;
442 	else
443 		info->blocks_per_page = 0;
444 
445 	if (sets == 0)
446 		info->assoc = 0xffff;
447 	else
448 		info->assoc = size / (sets * lsize);
449 }
450 
451 static bool __init parse_cache_info(struct device_node *np,
452 				    bool icache,
453 				    struct ppc_cache_info *info)
454 {
455 	static const char *ipropnames[] __initdata = {
456 		"i-cache-size",
457 		"i-cache-sets",
458 		"i-cache-block-size",
459 		"i-cache-line-size",
460 	};
461 	static const char *dpropnames[] __initdata = {
462 		"d-cache-size",
463 		"d-cache-sets",
464 		"d-cache-block-size",
465 		"d-cache-line-size",
466 	};
467 	const char **propnames = icache ? ipropnames : dpropnames;
468 	const __be32 *sizep, *lsizep, *bsizep, *setsp;
469 	u32 size, lsize, bsize, sets;
470 	bool success = true;
471 
472 	size = 0;
473 	sets = -1u;
474 	lsize = bsize = cur_cpu_spec->dcache_bsize;
475 	sizep = of_get_property(np, propnames[0], NULL);
476 	if (sizep != NULL)
477 		size = be32_to_cpu(*sizep);
478 	setsp = of_get_property(np, propnames[1], NULL);
479 	if (setsp != NULL)
480 		sets = be32_to_cpu(*setsp);
481 	bsizep = of_get_property(np, propnames[2], NULL);
482 	lsizep = of_get_property(np, propnames[3], NULL);
483 	if (bsizep == NULL)
484 		bsizep = lsizep;
485 	if (lsizep != NULL)
486 		lsize = be32_to_cpu(*lsizep);
487 	if (bsizep != NULL)
488 		bsize = be32_to_cpu(*bsizep);
489 	if (sizep == NULL || bsizep == NULL || lsizep == NULL)
490 		success = false;
491 
492 	/*
493 	 * OF is weird .. it represents fully associative caches
494 	 * as "1 way" which doesn't make much sense and doesn't
495 	 * leave room for direct mapped. We'll assume that 0
496 	 * in OF means direct mapped for that reason.
497 	 */
498 	if (sets == 1)
499 		sets = 0;
500 	else if (sets == 0)
501 		sets = 1;
502 
503 	init_cache_info(info, size, lsize, bsize, sets);
504 
505 	return success;
506 }
507 
508 void __init initialize_cache_info(void)
509 {
510 	struct device_node *cpu = NULL, *l2, *l3 = NULL;
511 	u32 pvr;
512 
513 	DBG(" -> initialize_cache_info()\n");
514 
515 	/*
516 	 * All shipping POWER8 machines have a firmware bug that
517 	 * puts incorrect information in the device-tree. This will
518 	 * be (hopefully) fixed for future chips but for now hard
519 	 * code the values if we are running on one of these
520 	 */
521 	pvr = PVR_VER(mfspr(SPRN_PVR));
522 	if (pvr == PVR_POWER8 || pvr == PVR_POWER8E ||
523 	    pvr == PVR_POWER8NVL) {
524 						/* size    lsize   blk  sets */
525 		init_cache_info(&ppc64_caches.l1i, 0x8000,   128,  128, 32);
526 		init_cache_info(&ppc64_caches.l1d, 0x10000,  128,  128, 64);
527 		init_cache_info(&ppc64_caches.l2,  0x80000,  128,  0,   512);
528 		init_cache_info(&ppc64_caches.l3,  0x800000, 128,  0,   8192);
529 	} else
530 		cpu = of_find_node_by_type(NULL, "cpu");
531 
532 	/*
533 	 * We're assuming *all* of the CPUs have the same
534 	 * d-cache and i-cache sizes... -Peter
535 	 */
536 	if (cpu) {
537 		if (!parse_cache_info(cpu, false, &ppc64_caches.l1d))
538 			DBG("Argh, can't find dcache properties !\n");
539 
540 		if (!parse_cache_info(cpu, true, &ppc64_caches.l1i))
541 			DBG("Argh, can't find icache properties !\n");
542 
543 		/*
544 		 * Try to find the L2 and L3 if any. Assume they are
545 		 * unified and use the D-side properties.
546 		 */
547 		l2 = of_find_next_cache_node(cpu);
548 		of_node_put(cpu);
549 		if (l2) {
550 			parse_cache_info(l2, false, &ppc64_caches.l2);
551 			l3 = of_find_next_cache_node(l2);
552 			of_node_put(l2);
553 		}
554 		if (l3) {
555 			parse_cache_info(l3, false, &ppc64_caches.l3);
556 			of_node_put(l3);
557 		}
558 	}
559 
560 	/* For use by binfmt_elf */
561 	dcache_bsize = ppc64_caches.l1d.block_size;
562 	icache_bsize = ppc64_caches.l1i.block_size;
563 
564 	cur_cpu_spec->dcache_bsize = dcache_bsize;
565 	cur_cpu_spec->icache_bsize = icache_bsize;
566 
567 	DBG(" <- initialize_cache_info()\n");
568 }
569 
570 /* This returns the limit below which memory accesses to the linear
571  * mapping are guarnateed not to cause a TLB or SLB miss. This is
572  * used to allocate interrupt or emergency stacks for which our
573  * exception entry path doesn't deal with being interrupted.
574  */
575 static __init u64 safe_stack_limit(void)
576 {
577 #ifdef CONFIG_PPC_BOOK3E
578 	/* Freescale BookE bolts the entire linear mapping */
579 	if (mmu_has_feature(MMU_FTR_TYPE_FSL_E))
580 		return linear_map_top;
581 	/* Other BookE, we assume the first GB is bolted */
582 	return 1ul << 30;
583 #else
584 	if (early_radix_enabled())
585 		return ULONG_MAX;
586 
587 	/* BookS, the first segment is bolted */
588 	if (mmu_has_feature(MMU_FTR_1T_SEGMENT))
589 		return 1UL << SID_SHIFT_1T;
590 	return 1UL << SID_SHIFT;
591 #endif
592 }
593 
594 void __init irqstack_early_init(void)
595 {
596 	u64 limit = safe_stack_limit();
597 	unsigned int i;
598 
599 	/*
600 	 * Interrupt stacks must be in the first segment since we
601 	 * cannot afford to take SLB misses on them. They are not
602 	 * accessed in realmode.
603 	 */
604 	for_each_possible_cpu(i) {
605 		softirq_ctx[i] = (struct thread_info *)
606 			__va(memblock_alloc_base(THREAD_SIZE,
607 					    THREAD_SIZE, limit));
608 		hardirq_ctx[i] = (struct thread_info *)
609 			__va(memblock_alloc_base(THREAD_SIZE,
610 					    THREAD_SIZE, limit));
611 	}
612 }
613 
614 #ifdef CONFIG_PPC_BOOK3E
615 void __init exc_lvl_early_init(void)
616 {
617 	unsigned int i;
618 	unsigned long sp;
619 
620 	for_each_possible_cpu(i) {
621 		sp = memblock_alloc(THREAD_SIZE, THREAD_SIZE);
622 		critirq_ctx[i] = (struct thread_info *)__va(sp);
623 		paca[i].crit_kstack = __va(sp + THREAD_SIZE);
624 
625 		sp = memblock_alloc(THREAD_SIZE, THREAD_SIZE);
626 		dbgirq_ctx[i] = (struct thread_info *)__va(sp);
627 		paca[i].dbg_kstack = __va(sp + THREAD_SIZE);
628 
629 		sp = memblock_alloc(THREAD_SIZE, THREAD_SIZE);
630 		mcheckirq_ctx[i] = (struct thread_info *)__va(sp);
631 		paca[i].mc_kstack = __va(sp + THREAD_SIZE);
632 	}
633 
634 	if (cpu_has_feature(CPU_FTR_DEBUG_LVL_EXC))
635 		patch_exception(0x040, exc_debug_debug_book3e);
636 }
637 #endif
638 
639 /*
640  * Emergency stacks are used for a range of things, from asynchronous
641  * NMIs (system reset, machine check) to synchronous, process context.
642  * We set preempt_count to zero, even though that isn't necessarily correct. To
643  * get the right value we'd need to copy it from the previous thread_info, but
644  * doing that might fault causing more problems.
645  * TODO: what to do with accounting?
646  */
647 static void emerg_stack_init_thread_info(struct thread_info *ti, int cpu)
648 {
649 	ti->task = NULL;
650 	ti->cpu = cpu;
651 	ti->preempt_count = 0;
652 	ti->local_flags = 0;
653 	ti->flags = 0;
654 	klp_init_thread_info(ti);
655 }
656 
657 /*
658  * Stack space used when we detect a bad kernel stack pointer, and
659  * early in SMP boots before relocation is enabled. Exclusive emergency
660  * stack for machine checks.
661  */
662 void __init emergency_stack_init(void)
663 {
664 	u64 limit;
665 	unsigned int i;
666 
667 	/*
668 	 * Emergency stacks must be under 256MB, we cannot afford to take
669 	 * SLB misses on them. The ABI also requires them to be 128-byte
670 	 * aligned.
671 	 *
672 	 * Since we use these as temporary stacks during secondary CPU
673 	 * bringup, machine check, system reset, and HMI, we need to get
674 	 * at them in real mode. This means they must also be within the RMO
675 	 * region.
676 	 *
677 	 * The IRQ stacks allocated elsewhere in this file are zeroed and
678 	 * initialized in kernel/irq.c. These are initialized here in order
679 	 * to have emergency stacks available as early as possible.
680 	 */
681 	limit = min(safe_stack_limit(), ppc64_rma_size);
682 
683 	for_each_possible_cpu(i) {
684 		struct thread_info *ti;
685 		ti = __va(memblock_alloc_base(THREAD_SIZE, THREAD_SIZE, limit));
686 		memset(ti, 0, THREAD_SIZE);
687 		emerg_stack_init_thread_info(ti, i);
688 		paca[i].emergency_sp = (void *)ti + THREAD_SIZE;
689 
690 #ifdef CONFIG_PPC_BOOK3S_64
691 		/* emergency stack for NMI exception handling. */
692 		ti = __va(memblock_alloc_base(THREAD_SIZE, THREAD_SIZE, limit));
693 		memset(ti, 0, THREAD_SIZE);
694 		emerg_stack_init_thread_info(ti, i);
695 		paca[i].nmi_emergency_sp = (void *)ti + THREAD_SIZE;
696 
697 		/* emergency stack for machine check exception handling. */
698 		ti = __va(memblock_alloc_base(THREAD_SIZE, THREAD_SIZE, limit));
699 		memset(ti, 0, THREAD_SIZE);
700 		emerg_stack_init_thread_info(ti, i);
701 		paca[i].mc_emergency_sp = (void *)ti + THREAD_SIZE;
702 #endif
703 	}
704 }
705 
706 #ifdef CONFIG_SMP
707 #define PCPU_DYN_SIZE		()
708 
709 static void * __init pcpu_fc_alloc(unsigned int cpu, size_t size, size_t align)
710 {
711 	return __alloc_bootmem_node(NODE_DATA(early_cpu_to_node(cpu)), size, align,
712 				    __pa(MAX_DMA_ADDRESS));
713 }
714 
715 static void __init pcpu_fc_free(void *ptr, size_t size)
716 {
717 	free_bootmem(__pa(ptr), size);
718 }
719 
720 static int pcpu_cpu_distance(unsigned int from, unsigned int to)
721 {
722 	if (early_cpu_to_node(from) == early_cpu_to_node(to))
723 		return LOCAL_DISTANCE;
724 	else
725 		return REMOTE_DISTANCE;
726 }
727 
728 unsigned long __per_cpu_offset[NR_CPUS] __read_mostly;
729 EXPORT_SYMBOL(__per_cpu_offset);
730 
731 void __init setup_per_cpu_areas(void)
732 {
733 	const size_t dyn_size = PERCPU_MODULE_RESERVE + PERCPU_DYNAMIC_RESERVE;
734 	size_t atom_size;
735 	unsigned long delta;
736 	unsigned int cpu;
737 	int rc;
738 
739 	/*
740 	 * Linear mapping is one of 4K, 1M and 16M.  For 4K, no need
741 	 * to group units.  For larger mappings, use 1M atom which
742 	 * should be large enough to contain a number of units.
743 	 */
744 	if (mmu_linear_psize == MMU_PAGE_4K)
745 		atom_size = PAGE_SIZE;
746 	else
747 		atom_size = 1 << 20;
748 
749 	rc = pcpu_embed_first_chunk(0, dyn_size, atom_size, pcpu_cpu_distance,
750 				    pcpu_fc_alloc, pcpu_fc_free);
751 	if (rc < 0)
752 		panic("cannot initialize percpu area (err=%d)", rc);
753 
754 	delta = (unsigned long)pcpu_base_addr - (unsigned long)__per_cpu_start;
755 	for_each_possible_cpu(cpu) {
756                 __per_cpu_offset[cpu] = delta + pcpu_unit_offsets[cpu];
757 		paca[cpu].data_offset = __per_cpu_offset[cpu];
758 	}
759 }
760 #endif
761 
762 #ifdef CONFIG_MEMORY_HOTPLUG_SPARSE
763 unsigned long memory_block_size_bytes(void)
764 {
765 	if (ppc_md.memory_block_size)
766 		return ppc_md.memory_block_size();
767 
768 	return MIN_MEMORY_BLOCK_SIZE;
769 }
770 #endif
771 
772 #if defined(CONFIG_PPC_INDIRECT_PIO) || defined(CONFIG_PPC_INDIRECT_MMIO)
773 struct ppc_pci_io ppc_pci_io;
774 EXPORT_SYMBOL(ppc_pci_io);
775 #endif
776 
777 #ifdef CONFIG_HARDLOCKUP_DETECTOR_PERF
778 u64 hw_nmi_get_sample_period(int watchdog_thresh)
779 {
780 	return ppc_proc_freq * watchdog_thresh;
781 }
782 #endif
783 
784 /*
785  * The perf based hardlockup detector breaks PMU event based branches, so
786  * disable it by default. Book3S has a soft-nmi hardlockup detector based
787  * on the decrementer interrupt, so it does not suffer from this problem.
788  *
789  * It is likely to get false positives in VM guests, so disable it there
790  * by default too.
791  */
792 static int __init disable_hardlockup_detector(void)
793 {
794 #ifdef CONFIG_HARDLOCKUP_DETECTOR_PERF
795 	hardlockup_detector_disable();
796 #else
797 	if (firmware_has_feature(FW_FEATURE_LPAR))
798 		hardlockup_detector_disable();
799 #endif
800 
801 	return 0;
802 }
803 early_initcall(disable_hardlockup_detector);
804