xref: /openbmc/linux/arch/powerpc/kernel/setup_64.c (revision 9cf0666f)
1 // SPDX-License-Identifier: GPL-2.0-or-later
2 /*
3  *
4  * Common boot and setup code.
5  *
6  * Copyright (C) 2001 PPC64 Team, IBM Corp
7  */
8 
9 #include <linux/export.h>
10 #include <linux/string.h>
11 #include <linux/sched.h>
12 #include <linux/init.h>
13 #include <linux/kernel.h>
14 #include <linux/reboot.h>
15 #include <linux/delay.h>
16 #include <linux/initrd.h>
17 #include <linux/seq_file.h>
18 #include <linux/ioport.h>
19 #include <linux/console.h>
20 #include <linux/utsname.h>
21 #include <linux/tty.h>
22 #include <linux/root_dev.h>
23 #include <linux/notifier.h>
24 #include <linux/cpu.h>
25 #include <linux/unistd.h>
26 #include <linux/serial.h>
27 #include <linux/serial_8250.h>
28 #include <linux/memblock.h>
29 #include <linux/pci.h>
30 #include <linux/lockdep.h>
31 #include <linux/memory.h>
32 #include <linux/nmi.h>
33 #include <linux/pgtable.h>
34 
35 #include <asm/kvm_guest.h>
36 #include <asm/io.h>
37 #include <asm/kdump.h>
38 #include <asm/prom.h>
39 #include <asm/processor.h>
40 #include <asm/smp.h>
41 #include <asm/elf.h>
42 #include <asm/machdep.h>
43 #include <asm/paca.h>
44 #include <asm/time.h>
45 #include <asm/cputable.h>
46 #include <asm/dt_cpu_ftrs.h>
47 #include <asm/sections.h>
48 #include <asm/btext.h>
49 #include <asm/nvram.h>
50 #include <asm/setup.h>
51 #include <asm/rtas.h>
52 #include <asm/iommu.h>
53 #include <asm/serial.h>
54 #include <asm/cache.h>
55 #include <asm/page.h>
56 #include <asm/mmu.h>
57 #include <asm/firmware.h>
58 #include <asm/xmon.h>
59 #include <asm/udbg.h>
60 #include <asm/kexec.h>
61 #include <asm/code-patching.h>
62 #include <asm/livepatch.h>
63 #include <asm/opal.h>
64 #include <asm/cputhreads.h>
65 #include <asm/hw_irq.h>
66 #include <asm/feature-fixups.h>
67 #include <asm/kup.h>
68 #include <asm/early_ioremap.h>
69 #include <asm/pgalloc.h>
70 #include <asm/asm-prototypes.h>
71 
72 #include "setup.h"
73 
74 int spinning_secondaries;
75 u64 ppc64_pft_size;
76 
77 struct ppc64_caches ppc64_caches = {
78 	.l1d = {
79 		.block_size = 0x40,
80 		.log_block_size = 6,
81 	},
82 	.l1i = {
83 		.block_size = 0x40,
84 		.log_block_size = 6
85 	},
86 };
87 EXPORT_SYMBOL_GPL(ppc64_caches);
88 
89 #if defined(CONFIG_PPC_BOOK3E) && defined(CONFIG_SMP)
90 void __init setup_tlb_core_data(void)
91 {
92 	int cpu;
93 
94 	BUILD_BUG_ON(offsetof(struct tlb_core_data, lock) != 0);
95 
96 	for_each_possible_cpu(cpu) {
97 		int first = cpu_first_thread_sibling(cpu);
98 
99 		/*
100 		 * If we boot via kdump on a non-primary thread,
101 		 * make sure we point at the thread that actually
102 		 * set up this TLB.
103 		 */
104 		if (cpu_first_thread_sibling(boot_cpuid) == first)
105 			first = boot_cpuid;
106 
107 		paca_ptrs[cpu]->tcd_ptr = &paca_ptrs[first]->tcd;
108 
109 		/*
110 		 * If we have threads, we need either tlbsrx.
111 		 * or e6500 tablewalk mode, or else TLB handlers
112 		 * will be racy and could produce duplicate entries.
113 		 * Should we panic instead?
114 		 */
115 		WARN_ONCE(smt_enabled_at_boot >= 2 &&
116 			  !mmu_has_feature(MMU_FTR_USE_TLBRSRV) &&
117 			  book3e_htw_mode != PPC_HTW_E6500,
118 			  "%s: unsupported MMU configuration\n", __func__);
119 	}
120 }
121 #endif
122 
123 #ifdef CONFIG_SMP
124 
125 static char *smt_enabled_cmdline;
126 
127 /* Look for ibm,smt-enabled OF option */
128 void __init check_smt_enabled(void)
129 {
130 	struct device_node *dn;
131 	const char *smt_option;
132 
133 	/* Default to enabling all threads */
134 	smt_enabled_at_boot = threads_per_core;
135 
136 	/* Allow the command line to overrule the OF option */
137 	if (smt_enabled_cmdline) {
138 		if (!strcmp(smt_enabled_cmdline, "on"))
139 			smt_enabled_at_boot = threads_per_core;
140 		else if (!strcmp(smt_enabled_cmdline, "off"))
141 			smt_enabled_at_boot = 0;
142 		else {
143 			int smt;
144 			int rc;
145 
146 			rc = kstrtoint(smt_enabled_cmdline, 10, &smt);
147 			if (!rc)
148 				smt_enabled_at_boot =
149 					min(threads_per_core, smt);
150 		}
151 	} else {
152 		dn = of_find_node_by_path("/options");
153 		if (dn) {
154 			smt_option = of_get_property(dn, "ibm,smt-enabled",
155 						     NULL);
156 
157 			if (smt_option) {
158 				if (!strcmp(smt_option, "on"))
159 					smt_enabled_at_boot = threads_per_core;
160 				else if (!strcmp(smt_option, "off"))
161 					smt_enabled_at_boot = 0;
162 			}
163 
164 			of_node_put(dn);
165 		}
166 	}
167 }
168 
169 /* Look for smt-enabled= cmdline option */
170 static int __init early_smt_enabled(char *p)
171 {
172 	smt_enabled_cmdline = p;
173 	return 0;
174 }
175 early_param("smt-enabled", early_smt_enabled);
176 
177 #endif /* CONFIG_SMP */
178 
179 /** Fix up paca fields required for the boot cpu */
180 static void __init fixup_boot_paca(void)
181 {
182 	/* The boot cpu is started */
183 	get_paca()->cpu_start = 1;
184 	/* Allow percpu accesses to work until we setup percpu data */
185 	get_paca()->data_offset = 0;
186 	/* Mark interrupts disabled in PACA */
187 	irq_soft_mask_set(IRQS_DISABLED);
188 }
189 
190 static void __init configure_exceptions(void)
191 {
192 	/*
193 	 * Setup the trampolines from the lowmem exception vectors
194 	 * to the kdump kernel when not using a relocatable kernel.
195 	 */
196 	setup_kdump_trampoline();
197 
198 	/* Under a PAPR hypervisor, we need hypercalls */
199 	if (firmware_has_feature(FW_FEATURE_SET_MODE)) {
200 		/* Enable AIL if possible */
201 		if (!pseries_enable_reloc_on_exc()) {
202 			init_task.thread.fscr &= ~FSCR_SCV;
203 			cur_cpu_spec->cpu_user_features2 &= ~PPC_FEATURE2_SCV;
204 		}
205 
206 		/*
207 		 * Tell the hypervisor that we want our exceptions to
208 		 * be taken in little endian mode.
209 		 *
210 		 * We don't call this for big endian as our calling convention
211 		 * makes us always enter in BE, and the call may fail under
212 		 * some circumstances with kdump.
213 		 */
214 #ifdef __LITTLE_ENDIAN__
215 		pseries_little_endian_exceptions();
216 #endif
217 	} else {
218 		/* Set endian mode using OPAL */
219 		if (firmware_has_feature(FW_FEATURE_OPAL))
220 			opal_configure_cores();
221 
222 		/* AIL on native is done in cpu_ready_for_interrupts() */
223 	}
224 }
225 
226 static void cpu_ready_for_interrupts(void)
227 {
228 	/*
229 	 * Enable AIL if supported, and we are in hypervisor mode. This
230 	 * is called once for every processor.
231 	 *
232 	 * If we are not in hypervisor mode the job is done once for
233 	 * the whole partition in configure_exceptions().
234 	 */
235 	if (cpu_has_feature(CPU_FTR_HVMODE)) {
236 		unsigned long lpcr = mfspr(SPRN_LPCR);
237 		unsigned long new_lpcr = lpcr;
238 
239 		if (cpu_has_feature(CPU_FTR_ARCH_31)) {
240 			/* P10 DD1 does not have HAIL */
241 			if (pvr_version_is(PVR_POWER10) &&
242 					(mfspr(SPRN_PVR) & 0xf00) == 0x100)
243 				new_lpcr |= LPCR_AIL_3;
244 			else
245 				new_lpcr |= LPCR_HAIL;
246 		} else if (cpu_has_feature(CPU_FTR_ARCH_207S)) {
247 			new_lpcr |= LPCR_AIL_3;
248 		}
249 
250 		if (new_lpcr != lpcr)
251 			mtspr(SPRN_LPCR, new_lpcr);
252 	}
253 
254 	/*
255 	 * Set HFSCR:TM based on CPU features:
256 	 * In the special case of TM no suspend (P9N DD2.1), Linux is
257 	 * told TM is off via the dt-ftrs but told to (partially) use
258 	 * it via OPAL_REINIT_CPUS_TM_SUSPEND_DISABLED. So HFSCR[TM]
259 	 * will be off from dt-ftrs but we need to turn it on for the
260 	 * no suspend case.
261 	 */
262 	if (cpu_has_feature(CPU_FTR_HVMODE)) {
263 		if (cpu_has_feature(CPU_FTR_TM_COMP))
264 			mtspr(SPRN_HFSCR, mfspr(SPRN_HFSCR) | HFSCR_TM);
265 		else
266 			mtspr(SPRN_HFSCR, mfspr(SPRN_HFSCR) & ~HFSCR_TM);
267 	}
268 
269 	/* Set IR and DR in PACA MSR */
270 	get_paca()->kernel_msr = MSR_KERNEL;
271 }
272 
273 unsigned long spr_default_dscr = 0;
274 
275 static void __init record_spr_defaults(void)
276 {
277 	if (early_cpu_has_feature(CPU_FTR_DSCR))
278 		spr_default_dscr = mfspr(SPRN_DSCR);
279 }
280 
281 /*
282  * Early initialization entry point. This is called by head.S
283  * with MMU translation disabled. We rely on the "feature" of
284  * the CPU that ignores the top 2 bits of the address in real
285  * mode so we can access kernel globals normally provided we
286  * only toy with things in the RMO region. From here, we do
287  * some early parsing of the device-tree to setup out MEMBLOCK
288  * data structures, and allocate & initialize the hash table
289  * and segment tables so we can start running with translation
290  * enabled.
291  *
292  * It is this function which will call the probe() callback of
293  * the various platform types and copy the matching one to the
294  * global ppc_md structure. Your platform can eventually do
295  * some very early initializations from the probe() routine, but
296  * this is not recommended, be very careful as, for example, the
297  * device-tree is not accessible via normal means at this point.
298  */
299 
300 void __init early_setup(unsigned long dt_ptr)
301 {
302 	static __initdata struct paca_struct boot_paca;
303 
304 	/* -------- printk is _NOT_ safe to use here ! ------- */
305 
306 	/*
307 	 * Assume we're on cpu 0 for now.
308 	 *
309 	 * We need to load a PACA very early for a few reasons.
310 	 *
311 	 * The stack protector canary is stored in the paca, so as soon as we
312 	 * call any stack protected code we need r13 pointing somewhere valid.
313 	 *
314 	 * If we are using kcov it will call in_task() in its instrumentation,
315 	 * which relies on the current task from the PACA.
316 	 *
317 	 * dt_cpu_ftrs_init() calls into generic OF/fdt code, as well as
318 	 * printk(), which can trigger both stack protector and kcov.
319 	 *
320 	 * percpu variables and spin locks also use the paca.
321 	 *
322 	 * So set up a temporary paca. It will be replaced below once we know
323 	 * what CPU we are on.
324 	 */
325 	initialise_paca(&boot_paca, 0);
326 	setup_paca(&boot_paca);
327 	fixup_boot_paca();
328 
329 	/* -------- printk is now safe to use ------- */
330 
331 	/* Try new device tree based feature discovery ... */
332 	if (!dt_cpu_ftrs_init(__va(dt_ptr)))
333 		/* Otherwise use the old style CPU table */
334 		identify_cpu(0, mfspr(SPRN_PVR));
335 
336 	/* Enable early debugging if any specified (see udbg.h) */
337 	udbg_early_init();
338 
339 	udbg_printf(" -> %s(), dt_ptr: 0x%lx\n", __func__, dt_ptr);
340 
341 	/*
342 	 * Do early initialization using the flattened device
343 	 * tree, such as retrieving the physical memory map or
344 	 * calculating/retrieving the hash table size.
345 	 */
346 	early_init_devtree(__va(dt_ptr));
347 
348 	/* Now we know the logical id of our boot cpu, setup the paca. */
349 	if (boot_cpuid != 0) {
350 		/* Poison paca_ptrs[0] again if it's not the boot cpu */
351 		memset(&paca_ptrs[0], 0x88, sizeof(paca_ptrs[0]));
352 	}
353 	setup_paca(paca_ptrs[boot_cpuid]);
354 	fixup_boot_paca();
355 
356 	/*
357 	 * Configure exception handlers. This include setting up trampolines
358 	 * if needed, setting exception endian mode, etc...
359 	 */
360 	configure_exceptions();
361 
362 	/*
363 	 * Configure Kernel Userspace Protection. This needs to happen before
364 	 * feature fixups for platforms that implement this using features.
365 	 */
366 	setup_kup();
367 
368 	/* Apply all the dynamic patching */
369 	apply_feature_fixups();
370 	setup_feature_keys();
371 
372 	/* Initialize the hash table or TLB handling */
373 	early_init_mmu();
374 
375 	early_ioremap_setup();
376 
377 	/*
378 	 * After firmware and early platform setup code has set things up,
379 	 * we note the SPR values for configurable control/performance
380 	 * registers, and use those as initial defaults.
381 	 */
382 	record_spr_defaults();
383 
384 	/*
385 	 * At this point, we can let interrupts switch to virtual mode
386 	 * (the MMU has been setup), so adjust the MSR in the PACA to
387 	 * have IR and DR set and enable AIL if it exists
388 	 */
389 	cpu_ready_for_interrupts();
390 
391 	/*
392 	 * We enable ftrace here, but since we only support DYNAMIC_FTRACE, it
393 	 * will only actually get enabled on the boot cpu much later once
394 	 * ftrace itself has been initialized.
395 	 */
396 	this_cpu_enable_ftrace();
397 
398 	udbg_printf(" <- %s()\n", __func__);
399 
400 #ifdef CONFIG_PPC_EARLY_DEBUG_BOOTX
401 	/*
402 	 * This needs to be done *last* (after the above udbg_printf() even)
403 	 *
404 	 * Right after we return from this function, we turn on the MMU
405 	 * which means the real-mode access trick that btext does will
406 	 * no longer work, it needs to switch to using a real MMU
407 	 * mapping. This call will ensure that it does
408 	 */
409 	btext_map();
410 #endif /* CONFIG_PPC_EARLY_DEBUG_BOOTX */
411 }
412 
413 #ifdef CONFIG_SMP
414 void early_setup_secondary(void)
415 {
416 	/* Mark interrupts disabled in PACA */
417 	irq_soft_mask_set(IRQS_DISABLED);
418 
419 	/* Initialize the hash table or TLB handling */
420 	early_init_mmu_secondary();
421 
422 	/* Perform any KUP setup that is per-cpu */
423 	setup_kup();
424 
425 	/*
426 	 * At this point, we can let interrupts switch to virtual mode
427 	 * (the MMU has been setup), so adjust the MSR in the PACA to
428 	 * have IR and DR set.
429 	 */
430 	cpu_ready_for_interrupts();
431 }
432 
433 #endif /* CONFIG_SMP */
434 
435 void panic_smp_self_stop(void)
436 {
437 	hard_irq_disable();
438 	spin_begin();
439 	while (1)
440 		spin_cpu_relax();
441 }
442 
443 #if defined(CONFIG_SMP) || defined(CONFIG_KEXEC_CORE)
444 static bool use_spinloop(void)
445 {
446 	if (IS_ENABLED(CONFIG_PPC_BOOK3S)) {
447 		/*
448 		 * See comments in head_64.S -- not all platforms insert
449 		 * secondaries at __secondary_hold and wait at the spin
450 		 * loop.
451 		 */
452 		if (firmware_has_feature(FW_FEATURE_OPAL))
453 			return false;
454 		return true;
455 	}
456 
457 	/*
458 	 * When book3e boots from kexec, the ePAPR spin table does
459 	 * not get used.
460 	 */
461 	return of_property_read_bool(of_chosen, "linux,booted-from-kexec");
462 }
463 
464 void smp_release_cpus(void)
465 {
466 	unsigned long *ptr;
467 	int i;
468 
469 	if (!use_spinloop())
470 		return;
471 
472 	/* All secondary cpus are spinning on a common spinloop, release them
473 	 * all now so they can start to spin on their individual paca
474 	 * spinloops. For non SMP kernels, the secondary cpus never get out
475 	 * of the common spinloop.
476 	 */
477 
478 	ptr  = (unsigned long *)((unsigned long)&__secondary_hold_spinloop
479 			- PHYSICAL_START);
480 	*ptr = ppc_function_entry(generic_secondary_smp_init);
481 
482 	/* And wait a bit for them to catch up */
483 	for (i = 0; i < 100000; i++) {
484 		mb();
485 		HMT_low();
486 		if (spinning_secondaries == 0)
487 			break;
488 		udelay(1);
489 	}
490 	pr_debug("spinning_secondaries = %d\n", spinning_secondaries);
491 }
492 #endif /* CONFIG_SMP || CONFIG_KEXEC_CORE */
493 
494 /*
495  * Initialize some remaining members of the ppc64_caches and systemcfg
496  * structures
497  * (at least until we get rid of them completely). This is mostly some
498  * cache informations about the CPU that will be used by cache flush
499  * routines and/or provided to userland
500  */
501 
502 static void init_cache_info(struct ppc_cache_info *info, u32 size, u32 lsize,
503 			    u32 bsize, u32 sets)
504 {
505 	info->size = size;
506 	info->sets = sets;
507 	info->line_size = lsize;
508 	info->block_size = bsize;
509 	info->log_block_size = __ilog2(bsize);
510 	if (bsize)
511 		info->blocks_per_page = PAGE_SIZE / bsize;
512 	else
513 		info->blocks_per_page = 0;
514 
515 	if (sets == 0)
516 		info->assoc = 0xffff;
517 	else
518 		info->assoc = size / (sets * lsize);
519 }
520 
521 static bool __init parse_cache_info(struct device_node *np,
522 				    bool icache,
523 				    struct ppc_cache_info *info)
524 {
525 	static const char *ipropnames[] __initdata = {
526 		"i-cache-size",
527 		"i-cache-sets",
528 		"i-cache-block-size",
529 		"i-cache-line-size",
530 	};
531 	static const char *dpropnames[] __initdata = {
532 		"d-cache-size",
533 		"d-cache-sets",
534 		"d-cache-block-size",
535 		"d-cache-line-size",
536 	};
537 	const char **propnames = icache ? ipropnames : dpropnames;
538 	const __be32 *sizep, *lsizep, *bsizep, *setsp;
539 	u32 size, lsize, bsize, sets;
540 	bool success = true;
541 
542 	size = 0;
543 	sets = -1u;
544 	lsize = bsize = cur_cpu_spec->dcache_bsize;
545 	sizep = of_get_property(np, propnames[0], NULL);
546 	if (sizep != NULL)
547 		size = be32_to_cpu(*sizep);
548 	setsp = of_get_property(np, propnames[1], NULL);
549 	if (setsp != NULL)
550 		sets = be32_to_cpu(*setsp);
551 	bsizep = of_get_property(np, propnames[2], NULL);
552 	lsizep = of_get_property(np, propnames[3], NULL);
553 	if (bsizep == NULL)
554 		bsizep = lsizep;
555 	if (lsizep == NULL)
556 		lsizep = bsizep;
557 	if (lsizep != NULL)
558 		lsize = be32_to_cpu(*lsizep);
559 	if (bsizep != NULL)
560 		bsize = be32_to_cpu(*bsizep);
561 	if (sizep == NULL || bsizep == NULL || lsizep == NULL)
562 		success = false;
563 
564 	/*
565 	 * OF is weird .. it represents fully associative caches
566 	 * as "1 way" which doesn't make much sense and doesn't
567 	 * leave room for direct mapped. We'll assume that 0
568 	 * in OF means direct mapped for that reason.
569 	 */
570 	if (sets == 1)
571 		sets = 0;
572 	else if (sets == 0)
573 		sets = 1;
574 
575 	init_cache_info(info, size, lsize, bsize, sets);
576 
577 	return success;
578 }
579 
580 void __init initialize_cache_info(void)
581 {
582 	struct device_node *cpu = NULL, *l2, *l3 = NULL;
583 	u32 pvr;
584 
585 	/*
586 	 * All shipping POWER8 machines have a firmware bug that
587 	 * puts incorrect information in the device-tree. This will
588 	 * be (hopefully) fixed for future chips but for now hard
589 	 * code the values if we are running on one of these
590 	 */
591 	pvr = PVR_VER(mfspr(SPRN_PVR));
592 	if (pvr == PVR_POWER8 || pvr == PVR_POWER8E ||
593 	    pvr == PVR_POWER8NVL) {
594 						/* size    lsize   blk  sets */
595 		init_cache_info(&ppc64_caches.l1i, 0x8000,   128,  128, 32);
596 		init_cache_info(&ppc64_caches.l1d, 0x10000,  128,  128, 64);
597 		init_cache_info(&ppc64_caches.l2,  0x80000,  128,  0,   512);
598 		init_cache_info(&ppc64_caches.l3,  0x800000, 128,  0,   8192);
599 	} else
600 		cpu = of_find_node_by_type(NULL, "cpu");
601 
602 	/*
603 	 * We're assuming *all* of the CPUs have the same
604 	 * d-cache and i-cache sizes... -Peter
605 	 */
606 	if (cpu) {
607 		if (!parse_cache_info(cpu, false, &ppc64_caches.l1d))
608 			pr_warn("Argh, can't find dcache properties !\n");
609 
610 		if (!parse_cache_info(cpu, true, &ppc64_caches.l1i))
611 			pr_warn("Argh, can't find icache properties !\n");
612 
613 		/*
614 		 * Try to find the L2 and L3 if any. Assume they are
615 		 * unified and use the D-side properties.
616 		 */
617 		l2 = of_find_next_cache_node(cpu);
618 		of_node_put(cpu);
619 		if (l2) {
620 			parse_cache_info(l2, false, &ppc64_caches.l2);
621 			l3 = of_find_next_cache_node(l2);
622 			of_node_put(l2);
623 		}
624 		if (l3) {
625 			parse_cache_info(l3, false, &ppc64_caches.l3);
626 			of_node_put(l3);
627 		}
628 	}
629 
630 	/* For use by binfmt_elf */
631 	dcache_bsize = ppc64_caches.l1d.block_size;
632 	icache_bsize = ppc64_caches.l1i.block_size;
633 
634 	cur_cpu_spec->dcache_bsize = dcache_bsize;
635 	cur_cpu_spec->icache_bsize = icache_bsize;
636 }
637 
638 /*
639  * This returns the limit below which memory accesses to the linear
640  * mapping are guarnateed not to cause an architectural exception (e.g.,
641  * TLB or SLB miss fault).
642  *
643  * This is used to allocate PACAs and various interrupt stacks that
644  * that are accessed early in interrupt handlers that must not cause
645  * re-entrant interrupts.
646  */
647 __init u64 ppc64_bolted_size(void)
648 {
649 #ifdef CONFIG_PPC_BOOK3E
650 	/* Freescale BookE bolts the entire linear mapping */
651 	/* XXX: BookE ppc64_rma_limit setup seems to disagree? */
652 	if (early_mmu_has_feature(MMU_FTR_TYPE_FSL_E))
653 		return linear_map_top;
654 	/* Other BookE, we assume the first GB is bolted */
655 	return 1ul << 30;
656 #else
657 	/* BookS radix, does not take faults on linear mapping */
658 	if (early_radix_enabled())
659 		return ULONG_MAX;
660 
661 	/* BookS hash, the first segment is bolted */
662 	if (early_mmu_has_feature(MMU_FTR_1T_SEGMENT))
663 		return 1UL << SID_SHIFT_1T;
664 	return 1UL << SID_SHIFT;
665 #endif
666 }
667 
668 static void *__init alloc_stack(unsigned long limit, int cpu)
669 {
670 	void *ptr;
671 
672 	BUILD_BUG_ON(STACK_INT_FRAME_SIZE % 16);
673 
674 	ptr = memblock_alloc_try_nid(THREAD_SIZE, THREAD_ALIGN,
675 				     MEMBLOCK_LOW_LIMIT, limit,
676 				     early_cpu_to_node(cpu));
677 	if (!ptr)
678 		panic("cannot allocate stacks");
679 
680 	return ptr;
681 }
682 
683 void __init irqstack_early_init(void)
684 {
685 	u64 limit = ppc64_bolted_size();
686 	unsigned int i;
687 
688 	/*
689 	 * Interrupt stacks must be in the first segment since we
690 	 * cannot afford to take SLB misses on them. They are not
691 	 * accessed in realmode.
692 	 */
693 	for_each_possible_cpu(i) {
694 		softirq_ctx[i] = alloc_stack(limit, i);
695 		hardirq_ctx[i] = alloc_stack(limit, i);
696 	}
697 }
698 
699 #ifdef CONFIG_PPC_BOOK3E
700 void __init exc_lvl_early_init(void)
701 {
702 	unsigned int i;
703 
704 	for_each_possible_cpu(i) {
705 		void *sp;
706 
707 		sp = alloc_stack(ULONG_MAX, i);
708 		critirq_ctx[i] = sp;
709 		paca_ptrs[i]->crit_kstack = sp + THREAD_SIZE;
710 
711 		sp = alloc_stack(ULONG_MAX, i);
712 		dbgirq_ctx[i] = sp;
713 		paca_ptrs[i]->dbg_kstack = sp + THREAD_SIZE;
714 
715 		sp = alloc_stack(ULONG_MAX, i);
716 		mcheckirq_ctx[i] = sp;
717 		paca_ptrs[i]->mc_kstack = sp + THREAD_SIZE;
718 	}
719 
720 	if (cpu_has_feature(CPU_FTR_DEBUG_LVL_EXC))
721 		patch_exception(0x040, exc_debug_debug_book3e);
722 }
723 #endif
724 
725 /*
726  * Stack space used when we detect a bad kernel stack pointer, and
727  * early in SMP boots before relocation is enabled. Exclusive emergency
728  * stack for machine checks.
729  */
730 void __init emergency_stack_init(void)
731 {
732 	u64 limit, mce_limit;
733 	unsigned int i;
734 
735 	/*
736 	 * Emergency stacks must be under 256MB, we cannot afford to take
737 	 * SLB misses on them. The ABI also requires them to be 128-byte
738 	 * aligned.
739 	 *
740 	 * Since we use these as temporary stacks during secondary CPU
741 	 * bringup, machine check, system reset, and HMI, we need to get
742 	 * at them in real mode. This means they must also be within the RMO
743 	 * region.
744 	 *
745 	 * The IRQ stacks allocated elsewhere in this file are zeroed and
746 	 * initialized in kernel/irq.c. These are initialized here in order
747 	 * to have emergency stacks available as early as possible.
748 	 */
749 	limit = mce_limit = min(ppc64_bolted_size(), ppc64_rma_size);
750 
751 	/*
752 	 * Machine check on pseries calls rtas, but can't use the static
753 	 * rtas_args due to a machine check hitting while the lock is held.
754 	 * rtas args have to be under 4GB, so the machine check stack is
755 	 * limited to 4GB so args can be put on stack.
756 	 */
757 	if (firmware_has_feature(FW_FEATURE_LPAR) && mce_limit > SZ_4G)
758 		mce_limit = SZ_4G;
759 
760 	for_each_possible_cpu(i) {
761 		paca_ptrs[i]->emergency_sp = alloc_stack(limit, i) + THREAD_SIZE;
762 
763 #ifdef CONFIG_PPC_BOOK3S_64
764 		/* emergency stack for NMI exception handling. */
765 		paca_ptrs[i]->nmi_emergency_sp = alloc_stack(limit, i) + THREAD_SIZE;
766 
767 		/* emergency stack for machine check exception handling. */
768 		paca_ptrs[i]->mc_emergency_sp = alloc_stack(mce_limit, i) + THREAD_SIZE;
769 #endif
770 	}
771 }
772 
773 #ifdef CONFIG_SMP
774 /**
775  * pcpu_alloc_bootmem - NUMA friendly alloc_bootmem wrapper for percpu
776  * @cpu: cpu to allocate for
777  * @size: size allocation in bytes
778  * @align: alignment
779  *
780  * Allocate @size bytes aligned at @align for cpu @cpu.  This wrapper
781  * does the right thing for NUMA regardless of the current
782  * configuration.
783  *
784  * RETURNS:
785  * Pointer to the allocated area on success, NULL on failure.
786  */
787 static void * __init pcpu_alloc_bootmem(unsigned int cpu, size_t size,
788 					size_t align)
789 {
790 	const unsigned long goal = __pa(MAX_DMA_ADDRESS);
791 #ifdef CONFIG_NUMA
792 	int node = early_cpu_to_node(cpu);
793 	void *ptr;
794 
795 	if (!node_online(node) || !NODE_DATA(node)) {
796 		ptr = memblock_alloc_from(size, align, goal);
797 		pr_info("cpu %d has no node %d or node-local memory\n",
798 			cpu, node);
799 		pr_debug("per cpu data for cpu%d %lu bytes at %016lx\n",
800 			 cpu, size, __pa(ptr));
801 	} else {
802 		ptr = memblock_alloc_try_nid(size, align, goal,
803 					     MEMBLOCK_ALLOC_ACCESSIBLE, node);
804 		pr_debug("per cpu data for cpu%d %lu bytes on node%d at "
805 			 "%016lx\n", cpu, size, node, __pa(ptr));
806 	}
807 	return ptr;
808 #else
809 	return memblock_alloc_from(size, align, goal);
810 #endif
811 }
812 
813 static void __init pcpu_free_bootmem(void *ptr, size_t size)
814 {
815 	memblock_free(ptr, size);
816 }
817 
818 static int pcpu_cpu_distance(unsigned int from, unsigned int to)
819 {
820 	if (early_cpu_to_node(from) == early_cpu_to_node(to))
821 		return LOCAL_DISTANCE;
822 	else
823 		return REMOTE_DISTANCE;
824 }
825 
826 unsigned long __per_cpu_offset[NR_CPUS] __read_mostly;
827 EXPORT_SYMBOL(__per_cpu_offset);
828 
829 static void __init pcpu_populate_pte(unsigned long addr)
830 {
831 	pgd_t *pgd = pgd_offset_k(addr);
832 	p4d_t *p4d;
833 	pud_t *pud;
834 	pmd_t *pmd;
835 
836 	p4d = p4d_offset(pgd, addr);
837 	if (p4d_none(*p4d)) {
838 		pud_t *new;
839 
840 		new = memblock_alloc(PUD_TABLE_SIZE, PUD_TABLE_SIZE);
841 		if (!new)
842 			goto err_alloc;
843 		p4d_populate(&init_mm, p4d, new);
844 	}
845 
846 	pud = pud_offset(p4d, addr);
847 	if (pud_none(*pud)) {
848 		pmd_t *new;
849 
850 		new = memblock_alloc(PMD_TABLE_SIZE, PMD_TABLE_SIZE);
851 		if (!new)
852 			goto err_alloc;
853 		pud_populate(&init_mm, pud, new);
854 	}
855 
856 	pmd = pmd_offset(pud, addr);
857 	if (!pmd_present(*pmd)) {
858 		pte_t *new;
859 
860 		new = memblock_alloc(PTE_TABLE_SIZE, PTE_TABLE_SIZE);
861 		if (!new)
862 			goto err_alloc;
863 		pmd_populate_kernel(&init_mm, pmd, new);
864 	}
865 
866 	return;
867 
868 err_alloc:
869 	panic("%s: Failed to allocate %lu bytes align=%lx from=%lx\n",
870 	      __func__, PAGE_SIZE, PAGE_SIZE, PAGE_SIZE);
871 }
872 
873 
874 void __init setup_per_cpu_areas(void)
875 {
876 	const size_t dyn_size = PERCPU_MODULE_RESERVE + PERCPU_DYNAMIC_RESERVE;
877 	size_t atom_size;
878 	unsigned long delta;
879 	unsigned int cpu;
880 	int rc = -EINVAL;
881 
882 	/*
883 	 * Linear mapping is one of 4K, 1M and 16M.  For 4K, no need
884 	 * to group units.  For larger mappings, use 1M atom which
885 	 * should be large enough to contain a number of units.
886 	 */
887 	if (mmu_linear_psize == MMU_PAGE_4K)
888 		atom_size = PAGE_SIZE;
889 	else
890 		atom_size = 1 << 20;
891 
892 	if (pcpu_chosen_fc != PCPU_FC_PAGE) {
893 		rc = pcpu_embed_first_chunk(0, dyn_size, atom_size, pcpu_cpu_distance,
894 					    pcpu_alloc_bootmem, pcpu_free_bootmem);
895 		if (rc)
896 			pr_warn("PERCPU: %s allocator failed (%d), "
897 				"falling back to page size\n",
898 				pcpu_fc_names[pcpu_chosen_fc], rc);
899 	}
900 
901 	if (rc < 0)
902 		rc = pcpu_page_first_chunk(0, pcpu_alloc_bootmem, pcpu_free_bootmem,
903 					   pcpu_populate_pte);
904 	if (rc < 0)
905 		panic("cannot initialize percpu area (err=%d)", rc);
906 
907 	delta = (unsigned long)pcpu_base_addr - (unsigned long)__per_cpu_start;
908 	for_each_possible_cpu(cpu) {
909                 __per_cpu_offset[cpu] = delta + pcpu_unit_offsets[cpu];
910 		paca_ptrs[cpu]->data_offset = __per_cpu_offset[cpu];
911 	}
912 }
913 #endif
914 
915 #ifdef CONFIG_MEMORY_HOTPLUG
916 unsigned long memory_block_size_bytes(void)
917 {
918 	if (ppc_md.memory_block_size)
919 		return ppc_md.memory_block_size();
920 
921 	return MIN_MEMORY_BLOCK_SIZE;
922 }
923 #endif
924 
925 #if defined(CONFIG_PPC_INDIRECT_PIO) || defined(CONFIG_PPC_INDIRECT_MMIO)
926 struct ppc_pci_io ppc_pci_io;
927 EXPORT_SYMBOL(ppc_pci_io);
928 #endif
929 
930 #ifdef CONFIG_HARDLOCKUP_DETECTOR_PERF
931 u64 hw_nmi_get_sample_period(int watchdog_thresh)
932 {
933 	return ppc_proc_freq * watchdog_thresh;
934 }
935 #endif
936 
937 /*
938  * The perf based hardlockup detector breaks PMU event based branches, so
939  * disable it by default. Book3S has a soft-nmi hardlockup detector based
940  * on the decrementer interrupt, so it does not suffer from this problem.
941  *
942  * It is likely to get false positives in KVM guests, so disable it there
943  * by default too. PowerVM will not stop or arbitrarily oversubscribe
944  * CPUs, but give a minimum regular allotment even with SPLPAR, so enable
945  * the detector for non-KVM guests, assume PowerVM.
946  */
947 static int __init disable_hardlockup_detector(void)
948 {
949 #ifdef CONFIG_HARDLOCKUP_DETECTOR_PERF
950 	hardlockup_detector_disable();
951 #else
952 	if (firmware_has_feature(FW_FEATURE_LPAR)) {
953 		if (is_kvm_guest())
954 			hardlockup_detector_disable();
955 	}
956 #endif
957 
958 	return 0;
959 }
960 early_initcall(disable_hardlockup_detector);
961