xref: /openbmc/linux/arch/powerpc/kernel/setup_64.c (revision 5a244f48)
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 #ifdef DEBUG
73 #define DBG(fmt...) udbg_printf(fmt)
74 #else
75 #define DBG(fmt...)
76 #endif
77 
78 int spinning_secondaries;
79 u64 ppc64_pft_size;
80 
81 struct ppc64_caches ppc64_caches = {
82 	.l1d = {
83 		.block_size = 0x40,
84 		.log_block_size = 6,
85 	},
86 	.l1i = {
87 		.block_size = 0x40,
88 		.log_block_size = 6
89 	},
90 };
91 EXPORT_SYMBOL_GPL(ppc64_caches);
92 
93 #if defined(CONFIG_PPC_BOOK3E) && defined(CONFIG_SMP)
94 void __init setup_tlb_core_data(void)
95 {
96 	int cpu;
97 
98 	BUILD_BUG_ON(offsetof(struct tlb_core_data, lock) != 0);
99 
100 	for_each_possible_cpu(cpu) {
101 		int first = cpu_first_thread_sibling(cpu);
102 
103 		/*
104 		 * If we boot via kdump on a non-primary thread,
105 		 * make sure we point at the thread that actually
106 		 * set up this TLB.
107 		 */
108 		if (cpu_first_thread_sibling(boot_cpuid) == first)
109 			first = boot_cpuid;
110 
111 		paca[cpu].tcd_ptr = &paca[first].tcd;
112 
113 		/*
114 		 * If we have threads, we need either tlbsrx.
115 		 * or e6500 tablewalk mode, or else TLB handlers
116 		 * will be racy and could produce duplicate entries.
117 		 * Should we panic instead?
118 		 */
119 		WARN_ONCE(smt_enabled_at_boot >= 2 &&
120 			  !mmu_has_feature(MMU_FTR_USE_TLBRSRV) &&
121 			  book3e_htw_mode != PPC_HTW_E6500,
122 			  "%s: unsupported MMU configuration\n", __func__);
123 	}
124 }
125 #endif
126 
127 #ifdef CONFIG_SMP
128 
129 static char *smt_enabled_cmdline;
130 
131 /* Look for ibm,smt-enabled OF option */
132 void __init check_smt_enabled(void)
133 {
134 	struct device_node *dn;
135 	const char *smt_option;
136 
137 	/* Default to enabling all threads */
138 	smt_enabled_at_boot = threads_per_core;
139 
140 	/* Allow the command line to overrule the OF option */
141 	if (smt_enabled_cmdline) {
142 		if (!strcmp(smt_enabled_cmdline, "on"))
143 			smt_enabled_at_boot = threads_per_core;
144 		else if (!strcmp(smt_enabled_cmdline, "off"))
145 			smt_enabled_at_boot = 0;
146 		else {
147 			int smt;
148 			int rc;
149 
150 			rc = kstrtoint(smt_enabled_cmdline, 10, &smt);
151 			if (!rc)
152 				smt_enabled_at_boot =
153 					min(threads_per_core, smt);
154 		}
155 	} else {
156 		dn = of_find_node_by_path("/options");
157 		if (dn) {
158 			smt_option = of_get_property(dn, "ibm,smt-enabled",
159 						     NULL);
160 
161 			if (smt_option) {
162 				if (!strcmp(smt_option, "on"))
163 					smt_enabled_at_boot = threads_per_core;
164 				else if (!strcmp(smt_option, "off"))
165 					smt_enabled_at_boot = 0;
166 			}
167 
168 			of_node_put(dn);
169 		}
170 	}
171 }
172 
173 /* Look for smt-enabled= cmdline option */
174 static int __init early_smt_enabled(char *p)
175 {
176 	smt_enabled_cmdline = p;
177 	return 0;
178 }
179 early_param("smt-enabled", early_smt_enabled);
180 
181 #endif /* CONFIG_SMP */
182 
183 /** Fix up paca fields required for the boot cpu */
184 static void __init fixup_boot_paca(void)
185 {
186 	/* The boot cpu is started */
187 	get_paca()->cpu_start = 1;
188 	/* Allow percpu accesses to work until we setup percpu data */
189 	get_paca()->data_offset = 0;
190 }
191 
192 static void __init configure_exceptions(void)
193 {
194 	/*
195 	 * Setup the trampolines from the lowmem exception vectors
196 	 * to the kdump kernel when not using a relocatable kernel.
197 	 */
198 	setup_kdump_trampoline();
199 
200 	/* Under a PAPR hypervisor, we need hypercalls */
201 	if (firmware_has_feature(FW_FEATURE_SET_MODE)) {
202 		/* Enable AIL if possible */
203 		pseries_enable_reloc_on_exc();
204 
205 		/*
206 		 * Tell the hypervisor that we want our exceptions to
207 		 * be taken in little endian mode.
208 		 *
209 		 * We don't call this for big endian as our calling convention
210 		 * makes us always enter in BE, and the call may fail under
211 		 * some circumstances with kdump.
212 		 */
213 #ifdef __LITTLE_ENDIAN__
214 		pseries_little_endian_exceptions();
215 #endif
216 	} else {
217 		/* Set endian mode using OPAL */
218 		if (firmware_has_feature(FW_FEATURE_OPAL))
219 			opal_configure_cores();
220 
221 		/* AIL on native is done in cpu_ready_for_interrupts() */
222 	}
223 }
224 
225 static void cpu_ready_for_interrupts(void)
226 {
227 	/*
228 	 * Enable AIL if supported, and we are in hypervisor mode. This
229 	 * is called once for every processor.
230 	 *
231 	 * If we are not in hypervisor mode the job is done once for
232 	 * the whole partition in configure_exceptions().
233 	 */
234 	if (cpu_has_feature(CPU_FTR_HVMODE) &&
235 	    cpu_has_feature(CPU_FTR_ARCH_207S)) {
236 		unsigned long lpcr = mfspr(SPRN_LPCR);
237 		mtspr(SPRN_LPCR, lpcr | LPCR_AIL_3);
238 	}
239 
240 	/*
241 	 * Fixup HFSCR:TM based on CPU features. The bit is set by our
242 	 * early asm init because at that point we haven't updated our
243 	 * CPU features from firmware and device-tree. Here we have,
244 	 * so let's do it.
245 	 */
246 	if (cpu_has_feature(CPU_FTR_HVMODE) && !cpu_has_feature(CPU_FTR_TM_COMP))
247 		mtspr(SPRN_HFSCR, mfspr(SPRN_HFSCR) & ~HFSCR_TM);
248 
249 	/* Set IR and DR in PACA MSR */
250 	get_paca()->kernel_msr = MSR_KERNEL;
251 }
252 
253 /*
254  * Early initialization entry point. This is called by head.S
255  * with MMU translation disabled. We rely on the "feature" of
256  * the CPU that ignores the top 2 bits of the address in real
257  * mode so we can access kernel globals normally provided we
258  * only toy with things in the RMO region. From here, we do
259  * some early parsing of the device-tree to setup out MEMBLOCK
260  * data structures, and allocate & initialize the hash table
261  * and segment tables so we can start running with translation
262  * enabled.
263  *
264  * It is this function which will call the probe() callback of
265  * the various platform types and copy the matching one to the
266  * global ppc_md structure. Your platform can eventually do
267  * some very early initializations from the probe() routine, but
268  * this is not recommended, be very careful as, for example, the
269  * device-tree is not accessible via normal means at this point.
270  */
271 
272 void __init early_setup(unsigned long dt_ptr)
273 {
274 	static __initdata struct paca_struct boot_paca;
275 
276 	/* -------- printk is _NOT_ safe to use here ! ------- */
277 
278 	/* Try new device tree based feature discovery ... */
279 	if (!dt_cpu_ftrs_init(__va(dt_ptr)))
280 		/* Otherwise use the old style CPU table */
281 		identify_cpu(0, mfspr(SPRN_PVR));
282 
283 	/* Assume we're on cpu 0 for now. Don't write to the paca yet! */
284 	initialise_paca(&boot_paca, 0);
285 	setup_paca(&boot_paca);
286 	fixup_boot_paca();
287 
288 	/* -------- printk is now safe to use ------- */
289 
290 	/* Enable early debugging if any specified (see udbg.h) */
291 	udbg_early_init();
292 
293  	DBG(" -> early_setup(), dt_ptr: 0x%lx\n", dt_ptr);
294 
295 	/*
296 	 * Do early initialization using the flattened device
297 	 * tree, such as retrieving the physical memory map or
298 	 * calculating/retrieving the hash table size.
299 	 */
300 	early_init_devtree(__va(dt_ptr));
301 
302 	/* Now we know the logical id of our boot cpu, setup the paca. */
303 	setup_paca(&paca[boot_cpuid]);
304 	fixup_boot_paca();
305 
306 	/*
307 	 * Configure exception handlers. This include setting up trampolines
308 	 * if needed, setting exception endian mode, etc...
309 	 */
310 	configure_exceptions();
311 
312 	/* Apply all the dynamic patching */
313 	apply_feature_fixups();
314 	setup_feature_keys();
315 
316 	/* Initialize the hash table or TLB handling */
317 	early_init_mmu();
318 
319 	/*
320 	 * At this point, we can let interrupts switch to virtual mode
321 	 * (the MMU has been setup), so adjust the MSR in the PACA to
322 	 * have IR and DR set and enable AIL if it exists
323 	 */
324 	cpu_ready_for_interrupts();
325 
326 	DBG(" <- early_setup()\n");
327 
328 #ifdef CONFIG_PPC_EARLY_DEBUG_BOOTX
329 	/*
330 	 * This needs to be done *last* (after the above DBG() even)
331 	 *
332 	 * Right after we return from this function, we turn on the MMU
333 	 * which means the real-mode access trick that btext does will
334 	 * no longer work, it needs to switch to using a real MMU
335 	 * mapping. This call will ensure that it does
336 	 */
337 	btext_map();
338 #endif /* CONFIG_PPC_EARLY_DEBUG_BOOTX */
339 }
340 
341 #ifdef CONFIG_SMP
342 void early_setup_secondary(void)
343 {
344 	/* Mark interrupts disabled in PACA */
345 	get_paca()->soft_enabled = 0;
346 
347 	/* Initialize the hash table or TLB handling */
348 	early_init_mmu_secondary();
349 
350 	/*
351 	 * At this point, we can let interrupts switch to virtual mode
352 	 * (the MMU has been setup), so adjust the MSR in the PACA to
353 	 * have IR and DR set.
354 	 */
355 	cpu_ready_for_interrupts();
356 }
357 
358 #endif /* CONFIG_SMP */
359 
360 #if defined(CONFIG_SMP) || defined(CONFIG_KEXEC_CORE)
361 static bool use_spinloop(void)
362 {
363 	if (!IS_ENABLED(CONFIG_PPC_BOOK3E))
364 		return true;
365 
366 	/*
367 	 * When book3e boots from kexec, the ePAPR spin table does
368 	 * not get used.
369 	 */
370 	return of_property_read_bool(of_chosen, "linux,booted-from-kexec");
371 }
372 
373 void smp_release_cpus(void)
374 {
375 	unsigned long *ptr;
376 	int i;
377 
378 	if (!use_spinloop())
379 		return;
380 
381 	DBG(" -> smp_release_cpus()\n");
382 
383 	/* All secondary cpus are spinning on a common spinloop, release them
384 	 * all now so they can start to spin on their individual paca
385 	 * spinloops. For non SMP kernels, the secondary cpus never get out
386 	 * of the common spinloop.
387 	 */
388 
389 	ptr  = (unsigned long *)((unsigned long)&__secondary_hold_spinloop
390 			- PHYSICAL_START);
391 	*ptr = ppc_function_entry(generic_secondary_smp_init);
392 
393 	/* And wait a bit for them to catch up */
394 	for (i = 0; i < 100000; i++) {
395 		mb();
396 		HMT_low();
397 		if (spinning_secondaries == 0)
398 			break;
399 		udelay(1);
400 	}
401 	DBG("spinning_secondaries = %d\n", spinning_secondaries);
402 
403 	DBG(" <- smp_release_cpus()\n");
404 }
405 #endif /* CONFIG_SMP || CONFIG_KEXEC_CORE */
406 
407 /*
408  * Initialize some remaining members of the ppc64_caches and systemcfg
409  * structures
410  * (at least until we get rid of them completely). This is mostly some
411  * cache informations about the CPU that will be used by cache flush
412  * routines and/or provided to userland
413  */
414 
415 static void init_cache_info(struct ppc_cache_info *info, u32 size, u32 lsize,
416 			    u32 bsize, u32 sets)
417 {
418 	info->size = size;
419 	info->sets = sets;
420 	info->line_size = lsize;
421 	info->block_size = bsize;
422 	info->log_block_size = __ilog2(bsize);
423 	if (bsize)
424 		info->blocks_per_page = PAGE_SIZE / bsize;
425 	else
426 		info->blocks_per_page = 0;
427 
428 	if (sets == 0)
429 		info->assoc = 0xffff;
430 	else
431 		info->assoc = size / (sets * lsize);
432 }
433 
434 static bool __init parse_cache_info(struct device_node *np,
435 				    bool icache,
436 				    struct ppc_cache_info *info)
437 {
438 	static const char *ipropnames[] __initdata = {
439 		"i-cache-size",
440 		"i-cache-sets",
441 		"i-cache-block-size",
442 		"i-cache-line-size",
443 	};
444 	static const char *dpropnames[] __initdata = {
445 		"d-cache-size",
446 		"d-cache-sets",
447 		"d-cache-block-size",
448 		"d-cache-line-size",
449 	};
450 	const char **propnames = icache ? ipropnames : dpropnames;
451 	const __be32 *sizep, *lsizep, *bsizep, *setsp;
452 	u32 size, lsize, bsize, sets;
453 	bool success = true;
454 
455 	size = 0;
456 	sets = -1u;
457 	lsize = bsize = cur_cpu_spec->dcache_bsize;
458 	sizep = of_get_property(np, propnames[0], NULL);
459 	if (sizep != NULL)
460 		size = be32_to_cpu(*sizep);
461 	setsp = of_get_property(np, propnames[1], NULL);
462 	if (setsp != NULL)
463 		sets = be32_to_cpu(*setsp);
464 	bsizep = of_get_property(np, propnames[2], NULL);
465 	lsizep = of_get_property(np, propnames[3], NULL);
466 	if (bsizep == NULL)
467 		bsizep = lsizep;
468 	if (lsizep != NULL)
469 		lsize = be32_to_cpu(*lsizep);
470 	if (bsizep != NULL)
471 		bsize = be32_to_cpu(*bsizep);
472 	if (sizep == NULL || bsizep == NULL || lsizep == NULL)
473 		success = false;
474 
475 	/*
476 	 * OF is weird .. it represents fully associative caches
477 	 * as "1 way" which doesn't make much sense and doesn't
478 	 * leave room for direct mapped. We'll assume that 0
479 	 * in OF means direct mapped for that reason.
480 	 */
481 	if (sets == 1)
482 		sets = 0;
483 	else if (sets == 0)
484 		sets = 1;
485 
486 	init_cache_info(info, size, lsize, bsize, sets);
487 
488 	return success;
489 }
490 
491 void __init initialize_cache_info(void)
492 {
493 	struct device_node *cpu = NULL, *l2, *l3 = NULL;
494 	u32 pvr;
495 
496 	DBG(" -> initialize_cache_info()\n");
497 
498 	/*
499 	 * All shipping POWER8 machines have a firmware bug that
500 	 * puts incorrect information in the device-tree. This will
501 	 * be (hopefully) fixed for future chips but for now hard
502 	 * code the values if we are running on one of these
503 	 */
504 	pvr = PVR_VER(mfspr(SPRN_PVR));
505 	if (pvr == PVR_POWER8 || pvr == PVR_POWER8E ||
506 	    pvr == PVR_POWER8NVL) {
507 						/* size    lsize   blk  sets */
508 		init_cache_info(&ppc64_caches.l1i, 0x8000,   128,  128, 32);
509 		init_cache_info(&ppc64_caches.l1d, 0x10000,  128,  128, 64);
510 		init_cache_info(&ppc64_caches.l2,  0x80000,  128,  0,   512);
511 		init_cache_info(&ppc64_caches.l3,  0x800000, 128,  0,   8192);
512 	} else
513 		cpu = of_find_node_by_type(NULL, "cpu");
514 
515 	/*
516 	 * We're assuming *all* of the CPUs have the same
517 	 * d-cache and i-cache sizes... -Peter
518 	 */
519 	if (cpu) {
520 		if (!parse_cache_info(cpu, false, &ppc64_caches.l1d))
521 			DBG("Argh, can't find dcache properties !\n");
522 
523 		if (!parse_cache_info(cpu, true, &ppc64_caches.l1i))
524 			DBG("Argh, can't find icache properties !\n");
525 
526 		/*
527 		 * Try to find the L2 and L3 if any. Assume they are
528 		 * unified and use the D-side properties.
529 		 */
530 		l2 = of_find_next_cache_node(cpu);
531 		of_node_put(cpu);
532 		if (l2) {
533 			parse_cache_info(l2, false, &ppc64_caches.l2);
534 			l3 = of_find_next_cache_node(l2);
535 			of_node_put(l2);
536 		}
537 		if (l3) {
538 			parse_cache_info(l3, false, &ppc64_caches.l3);
539 			of_node_put(l3);
540 		}
541 	}
542 
543 	/* For use by binfmt_elf */
544 	dcache_bsize = ppc64_caches.l1d.block_size;
545 	icache_bsize = ppc64_caches.l1i.block_size;
546 
547 	cur_cpu_spec->dcache_bsize = dcache_bsize;
548 	cur_cpu_spec->icache_bsize = icache_bsize;
549 
550 	DBG(" <- initialize_cache_info()\n");
551 }
552 
553 /* This returns the limit below which memory accesses to the linear
554  * mapping are guarnateed not to cause a TLB or SLB miss. This is
555  * used to allocate interrupt or emergency stacks for which our
556  * exception entry path doesn't deal with being interrupted.
557  */
558 static __init u64 safe_stack_limit(void)
559 {
560 #ifdef CONFIG_PPC_BOOK3E
561 	/* Freescale BookE bolts the entire linear mapping */
562 	if (mmu_has_feature(MMU_FTR_TYPE_FSL_E))
563 		return linear_map_top;
564 	/* Other BookE, we assume the first GB is bolted */
565 	return 1ul << 30;
566 #else
567 	if (early_radix_enabled())
568 		return ULONG_MAX;
569 
570 	/* BookS, the first segment is bolted */
571 	if (mmu_has_feature(MMU_FTR_1T_SEGMENT))
572 		return 1UL << SID_SHIFT_1T;
573 	return 1UL << SID_SHIFT;
574 #endif
575 }
576 
577 void __init irqstack_early_init(void)
578 {
579 	u64 limit = safe_stack_limit();
580 	unsigned int i;
581 
582 	/*
583 	 * Interrupt stacks must be in the first segment since we
584 	 * cannot afford to take SLB misses on them. They are not
585 	 * accessed in realmode.
586 	 */
587 	for_each_possible_cpu(i) {
588 		softirq_ctx[i] = (struct thread_info *)
589 			__va(memblock_alloc_base(THREAD_SIZE,
590 					    THREAD_SIZE, limit));
591 		hardirq_ctx[i] = (struct thread_info *)
592 			__va(memblock_alloc_base(THREAD_SIZE,
593 					    THREAD_SIZE, limit));
594 	}
595 }
596 
597 #ifdef CONFIG_PPC_BOOK3E
598 void __init exc_lvl_early_init(void)
599 {
600 	unsigned int i;
601 	unsigned long sp;
602 
603 	for_each_possible_cpu(i) {
604 		sp = memblock_alloc(THREAD_SIZE, THREAD_SIZE);
605 		critirq_ctx[i] = (struct thread_info *)__va(sp);
606 		paca[i].crit_kstack = __va(sp + THREAD_SIZE);
607 
608 		sp = memblock_alloc(THREAD_SIZE, THREAD_SIZE);
609 		dbgirq_ctx[i] = (struct thread_info *)__va(sp);
610 		paca[i].dbg_kstack = __va(sp + THREAD_SIZE);
611 
612 		sp = memblock_alloc(THREAD_SIZE, THREAD_SIZE);
613 		mcheckirq_ctx[i] = (struct thread_info *)__va(sp);
614 		paca[i].mc_kstack = __va(sp + THREAD_SIZE);
615 	}
616 
617 	if (cpu_has_feature(CPU_FTR_DEBUG_LVL_EXC))
618 		patch_exception(0x040, exc_debug_debug_book3e);
619 }
620 #endif
621 
622 /*
623  * Emergency stacks are used for a range of things, from asynchronous
624  * NMIs (system reset, machine check) to synchronous, process context.
625  * We set preempt_count to zero, even though that isn't necessarily correct. To
626  * get the right value we'd need to copy it from the previous thread_info, but
627  * doing that might fault causing more problems.
628  * TODO: what to do with accounting?
629  */
630 static void emerg_stack_init_thread_info(struct thread_info *ti, int cpu)
631 {
632 	ti->task = NULL;
633 	ti->cpu = cpu;
634 	ti->preempt_count = 0;
635 	ti->local_flags = 0;
636 	ti->flags = 0;
637 	klp_init_thread_info(ti);
638 }
639 
640 /*
641  * Stack space used when we detect a bad kernel stack pointer, and
642  * early in SMP boots before relocation is enabled. Exclusive emergency
643  * stack for machine checks.
644  */
645 void __init emergency_stack_init(void)
646 {
647 	u64 limit;
648 	unsigned int i;
649 
650 	/*
651 	 * Emergency stacks must be under 256MB, we cannot afford to take
652 	 * SLB misses on them. The ABI also requires them to be 128-byte
653 	 * aligned.
654 	 *
655 	 * Since we use these as temporary stacks during secondary CPU
656 	 * bringup, machine check, system reset, and HMI, we need to get
657 	 * at them in real mode. This means they must also be within the RMO
658 	 * region.
659 	 *
660 	 * The IRQ stacks allocated elsewhere in this file are zeroed and
661 	 * initialized in kernel/irq.c. These are initialized here in order
662 	 * to have emergency stacks available as early as possible.
663 	 */
664 	limit = min(safe_stack_limit(), ppc64_rma_size);
665 
666 	for_each_possible_cpu(i) {
667 		struct thread_info *ti;
668 		ti = __va(memblock_alloc_base(THREAD_SIZE, THREAD_SIZE, limit));
669 		memset(ti, 0, THREAD_SIZE);
670 		emerg_stack_init_thread_info(ti, i);
671 		paca[i].emergency_sp = (void *)ti + THREAD_SIZE;
672 
673 #ifdef CONFIG_PPC_BOOK3S_64
674 		/* emergency stack for NMI exception handling. */
675 		ti = __va(memblock_alloc_base(THREAD_SIZE, THREAD_SIZE, limit));
676 		memset(ti, 0, THREAD_SIZE);
677 		emerg_stack_init_thread_info(ti, i);
678 		paca[i].nmi_emergency_sp = (void *)ti + THREAD_SIZE;
679 
680 		/* emergency stack for machine check exception handling. */
681 		ti = __va(memblock_alloc_base(THREAD_SIZE, THREAD_SIZE, limit));
682 		memset(ti, 0, THREAD_SIZE);
683 		emerg_stack_init_thread_info(ti, i);
684 		paca[i].mc_emergency_sp = (void *)ti + THREAD_SIZE;
685 #endif
686 	}
687 }
688 
689 #ifdef CONFIG_SMP
690 #define PCPU_DYN_SIZE		()
691 
692 static void * __init pcpu_fc_alloc(unsigned int cpu, size_t size, size_t align)
693 {
694 	return __alloc_bootmem_node(NODE_DATA(early_cpu_to_node(cpu)), size, align,
695 				    __pa(MAX_DMA_ADDRESS));
696 }
697 
698 static void __init pcpu_fc_free(void *ptr, size_t size)
699 {
700 	free_bootmem(__pa(ptr), size);
701 }
702 
703 static int pcpu_cpu_distance(unsigned int from, unsigned int to)
704 {
705 	if (early_cpu_to_node(from) == early_cpu_to_node(to))
706 		return LOCAL_DISTANCE;
707 	else
708 		return REMOTE_DISTANCE;
709 }
710 
711 unsigned long __per_cpu_offset[NR_CPUS] __read_mostly;
712 EXPORT_SYMBOL(__per_cpu_offset);
713 
714 void __init setup_per_cpu_areas(void)
715 {
716 	const size_t dyn_size = PERCPU_MODULE_RESERVE + PERCPU_DYNAMIC_RESERVE;
717 	size_t atom_size;
718 	unsigned long delta;
719 	unsigned int cpu;
720 	int rc;
721 
722 	/*
723 	 * Linear mapping is one of 4K, 1M and 16M.  For 4K, no need
724 	 * to group units.  For larger mappings, use 1M atom which
725 	 * should be large enough to contain a number of units.
726 	 */
727 	if (mmu_linear_psize == MMU_PAGE_4K)
728 		atom_size = PAGE_SIZE;
729 	else
730 		atom_size = 1 << 20;
731 
732 	rc = pcpu_embed_first_chunk(0, dyn_size, atom_size, pcpu_cpu_distance,
733 				    pcpu_fc_alloc, pcpu_fc_free);
734 	if (rc < 0)
735 		panic("cannot initialize percpu area (err=%d)", rc);
736 
737 	delta = (unsigned long)pcpu_base_addr - (unsigned long)__per_cpu_start;
738 	for_each_possible_cpu(cpu) {
739                 __per_cpu_offset[cpu] = delta + pcpu_unit_offsets[cpu];
740 		paca[cpu].data_offset = __per_cpu_offset[cpu];
741 	}
742 }
743 #endif
744 
745 #ifdef CONFIG_MEMORY_HOTPLUG_SPARSE
746 unsigned long memory_block_size_bytes(void)
747 {
748 	if (ppc_md.memory_block_size)
749 		return ppc_md.memory_block_size();
750 
751 	return MIN_MEMORY_BLOCK_SIZE;
752 }
753 #endif
754 
755 #if defined(CONFIG_PPC_INDIRECT_PIO) || defined(CONFIG_PPC_INDIRECT_MMIO)
756 struct ppc_pci_io ppc_pci_io;
757 EXPORT_SYMBOL(ppc_pci_io);
758 #endif
759 
760 #ifdef CONFIG_HARDLOCKUP_DETECTOR_PERF
761 u64 hw_nmi_get_sample_period(int watchdog_thresh)
762 {
763 	return ppc_proc_freq * watchdog_thresh;
764 }
765 #endif
766 
767 /*
768  * The perf based hardlockup detector breaks PMU event based branches, so
769  * disable it by default. Book3S has a soft-nmi hardlockup detector based
770  * on the decrementer interrupt, so it does not suffer from this problem.
771  *
772  * It is likely to get false positives in VM guests, so disable it there
773  * by default too.
774  */
775 static int __init disable_hardlockup_detector(void)
776 {
777 #ifdef CONFIG_HARDLOCKUP_DETECTOR_PERF
778 	hardlockup_detector_disable();
779 #else
780 	if (firmware_has_feature(FW_FEATURE_LPAR))
781 		hardlockup_detector_disable();
782 #endif
783 
784 	return 0;
785 }
786 early_initcall(disable_hardlockup_detector);
787