xref: /openbmc/linux/arch/powerpc/kernel/setup_64.c (revision 0661cb2a)
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/debugfs.h>
36 #include <asm/kvm_guest.h>
37 #include <asm/io.h>
38 #include <asm/kdump.h>
39 #include <asm/prom.h>
40 #include <asm/processor.h>
41 #include <asm/smp.h>
42 #include <asm/elf.h>
43 #include <asm/machdep.h>
44 #include <asm/paca.h>
45 #include <asm/time.h>
46 #include <asm/cputable.h>
47 #include <asm/dt_cpu_ftrs.h>
48 #include <asm/sections.h>
49 #include <asm/btext.h>
50 #include <asm/nvram.h>
51 #include <asm/setup.h>
52 #include <asm/rtas.h>
53 #include <asm/iommu.h>
54 #include <asm/serial.h>
55 #include <asm/cache.h>
56 #include <asm/page.h>
57 #include <asm/mmu.h>
58 #include <asm/firmware.h>
59 #include <asm/xmon.h>
60 #include <asm/udbg.h>
61 #include <asm/kexec.h>
62 #include <asm/code-patching.h>
63 #include <asm/livepatch.h>
64 #include <asm/opal.h>
65 #include <asm/cputhreads.h>
66 #include <asm/hw_irq.h>
67 #include <asm/feature-fixups.h>
68 #include <asm/kup.h>
69 #include <asm/early_ioremap.h>
70 #include <asm/pgalloc.h>
71 #include <asm/asm-prototypes.h>
72 
73 #include "setup.h"
74 
75 int spinning_secondaries;
76 u64 ppc64_pft_size;
77 
78 struct ppc64_caches ppc64_caches = {
79 	.l1d = {
80 		.block_size = 0x40,
81 		.log_block_size = 6,
82 	},
83 	.l1i = {
84 		.block_size = 0x40,
85 		.log_block_size = 6
86 	},
87 };
88 EXPORT_SYMBOL_GPL(ppc64_caches);
89 
90 #if defined(CONFIG_PPC_BOOK3E) && defined(CONFIG_SMP)
91 void __init setup_tlb_core_data(void)
92 {
93 	int cpu;
94 
95 	BUILD_BUG_ON(offsetof(struct tlb_core_data, lock) != 0);
96 
97 	for_each_possible_cpu(cpu) {
98 		int first = cpu_first_thread_sibling(cpu);
99 
100 		/*
101 		 * If we boot via kdump on a non-primary thread,
102 		 * make sure we point at the thread that actually
103 		 * set up this TLB.
104 		 */
105 		if (cpu_first_thread_sibling(boot_cpuid) == first)
106 			first = boot_cpuid;
107 
108 		paca_ptrs[cpu]->tcd_ptr = &paca_ptrs[first]->tcd;
109 
110 		/*
111 		 * If we have threads, we need either tlbsrx.
112 		 * or e6500 tablewalk mode, or else TLB handlers
113 		 * will be racy and could produce duplicate entries.
114 		 * Should we panic instead?
115 		 */
116 		WARN_ONCE(smt_enabled_at_boot >= 2 &&
117 			  !mmu_has_feature(MMU_FTR_USE_TLBRSRV) &&
118 			  book3e_htw_mode != PPC_HTW_E6500,
119 			  "%s: unsupported MMU configuration\n", __func__);
120 	}
121 }
122 #endif
123 
124 #ifdef CONFIG_SMP
125 
126 static char *smt_enabled_cmdline;
127 
128 /* Look for ibm,smt-enabled OF option */
129 void __init check_smt_enabled(void)
130 {
131 	struct device_node *dn;
132 	const char *smt_option;
133 
134 	/* Default to enabling all threads */
135 	smt_enabled_at_boot = threads_per_core;
136 
137 	/* Allow the command line to overrule the OF option */
138 	if (smt_enabled_cmdline) {
139 		if (!strcmp(smt_enabled_cmdline, "on"))
140 			smt_enabled_at_boot = threads_per_core;
141 		else if (!strcmp(smt_enabled_cmdline, "off"))
142 			smt_enabled_at_boot = 0;
143 		else {
144 			int smt;
145 			int rc;
146 
147 			rc = kstrtoint(smt_enabled_cmdline, 10, &smt);
148 			if (!rc)
149 				smt_enabled_at_boot =
150 					min(threads_per_core, smt);
151 		}
152 	} else {
153 		dn = of_find_node_by_path("/options");
154 		if (dn) {
155 			smt_option = of_get_property(dn, "ibm,smt-enabled",
156 						     NULL);
157 
158 			if (smt_option) {
159 				if (!strcmp(smt_option, "on"))
160 					smt_enabled_at_boot = threads_per_core;
161 				else if (!strcmp(smt_option, "off"))
162 					smt_enabled_at_boot = 0;
163 			}
164 
165 			of_node_put(dn);
166 		}
167 	}
168 }
169 
170 /* Look for smt-enabled= cmdline option */
171 static int __init early_smt_enabled(char *p)
172 {
173 	smt_enabled_cmdline = p;
174 	return 0;
175 }
176 early_param("smt-enabled", early_smt_enabled);
177 
178 #endif /* CONFIG_SMP */
179 
180 /** Fix up paca fields required for the boot cpu */
181 static void __init fixup_boot_paca(void)
182 {
183 	/* The boot cpu is started */
184 	get_paca()->cpu_start = 1;
185 	/* Allow percpu accesses to work until we setup percpu data */
186 	get_paca()->data_offset = 0;
187 	/* Mark interrupts disabled in PACA */
188 	irq_soft_mask_set(IRQS_DISABLED);
189 }
190 
191 static void __init configure_exceptions(void)
192 {
193 	/*
194 	 * Setup the trampolines from the lowmem exception vectors
195 	 * to the kdump kernel when not using a relocatable kernel.
196 	 */
197 	setup_kdump_trampoline();
198 
199 	/* Under a PAPR hypervisor, we need hypercalls */
200 	if (firmware_has_feature(FW_FEATURE_SET_MODE)) {
201 		/* Enable AIL if possible */
202 		if (!pseries_enable_reloc_on_exc()) {
203 			init_task.thread.fscr &= ~FSCR_SCV;
204 			cur_cpu_spec->cpu_user_features2 &= ~PPC_FEATURE2_SCV;
205 		}
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 		unsigned long lpcr = mfspr(SPRN_LPCR);
238 		unsigned long new_lpcr = lpcr;
239 
240 		if (cpu_has_feature(CPU_FTR_ARCH_31)) {
241 			/* P10 DD1 does not have HAIL */
242 			if (pvr_version_is(PVR_POWER10) &&
243 					(mfspr(SPRN_PVR) & 0xf00) == 0x100)
244 				new_lpcr |= LPCR_AIL_3;
245 			else
246 				new_lpcr |= LPCR_HAIL;
247 		} else if (cpu_has_feature(CPU_FTR_ARCH_207S)) {
248 			new_lpcr |= LPCR_AIL_3;
249 		}
250 
251 		if (new_lpcr != lpcr)
252 			mtspr(SPRN_LPCR, new_lpcr);
253 	}
254 
255 	/*
256 	 * Set HFSCR:TM based on CPU features:
257 	 * In the special case of TM no suspend (P9N DD2.1), Linux is
258 	 * told TM is off via the dt-ftrs but told to (partially) use
259 	 * it via OPAL_REINIT_CPUS_TM_SUSPEND_DISABLED. So HFSCR[TM]
260 	 * will be off from dt-ftrs but we need to turn it on for the
261 	 * no suspend case.
262 	 */
263 	if (cpu_has_feature(CPU_FTR_HVMODE)) {
264 		if (cpu_has_feature(CPU_FTR_TM_COMP))
265 			mtspr(SPRN_HFSCR, mfspr(SPRN_HFSCR) | HFSCR_TM);
266 		else
267 			mtspr(SPRN_HFSCR, mfspr(SPRN_HFSCR) & ~HFSCR_TM);
268 	}
269 
270 	/* Set IR and DR in PACA MSR */
271 	get_paca()->kernel_msr = MSR_KERNEL;
272 }
273 
274 unsigned long spr_default_dscr = 0;
275 
276 static void __init record_spr_defaults(void)
277 {
278 	if (early_cpu_has_feature(CPU_FTR_DSCR))
279 		spr_default_dscr = mfspr(SPRN_DSCR);
280 }
281 
282 /*
283  * Early initialization entry point. This is called by head.S
284  * with MMU translation disabled. We rely on the "feature" of
285  * the CPU that ignores the top 2 bits of the address in real
286  * mode so we can access kernel globals normally provided we
287  * only toy with things in the RMO region. From here, we do
288  * some early parsing of the device-tree to setup out MEMBLOCK
289  * data structures, and allocate & initialize the hash table
290  * and segment tables so we can start running with translation
291  * enabled.
292  *
293  * It is this function which will call the probe() callback of
294  * the various platform types and copy the matching one to the
295  * global ppc_md structure. Your platform can eventually do
296  * some very early initializations from the probe() routine, but
297  * this is not recommended, be very careful as, for example, the
298  * device-tree is not accessible via normal means at this point.
299  */
300 
301 void __init early_setup(unsigned long dt_ptr)
302 {
303 	static __initdata struct paca_struct boot_paca;
304 
305 	/* -------- printk is _NOT_ safe to use here ! ------- */
306 
307 	/*
308 	 * Assume we're on cpu 0 for now.
309 	 *
310 	 * We need to load a PACA very early for a few reasons.
311 	 *
312 	 * The stack protector canary is stored in the paca, so as soon as we
313 	 * call any stack protected code we need r13 pointing somewhere valid.
314 	 *
315 	 * If we are using kcov it will call in_task() in its instrumentation,
316 	 * which relies on the current task from the PACA.
317 	 *
318 	 * dt_cpu_ftrs_init() calls into generic OF/fdt code, as well as
319 	 * printk(), which can trigger both stack protector and kcov.
320 	 *
321 	 * percpu variables and spin locks also use the paca.
322 	 *
323 	 * So set up a temporary paca. It will be replaced below once we know
324 	 * what CPU we are on.
325 	 */
326 	initialise_paca(&boot_paca, 0);
327 	setup_paca(&boot_paca);
328 	fixup_boot_paca();
329 
330 	/* -------- printk is now safe to use ------- */
331 
332 	/* Try new device tree based feature discovery ... */
333 	if (!dt_cpu_ftrs_init(__va(dt_ptr)))
334 		/* Otherwise use the old style CPU table */
335 		identify_cpu(0, mfspr(SPRN_PVR));
336 
337 	/* Enable early debugging if any specified (see udbg.h) */
338 	udbg_early_init();
339 
340 	udbg_printf(" -> %s(), dt_ptr: 0x%lx\n", __func__, dt_ptr);
341 
342 	/*
343 	 * Do early initialization using the flattened device
344 	 * tree, such as retrieving the physical memory map or
345 	 * calculating/retrieving the hash table size.
346 	 */
347 	early_init_devtree(__va(dt_ptr));
348 
349 	/* Now we know the logical id of our boot cpu, setup the paca. */
350 	if (boot_cpuid != 0) {
351 		/* Poison paca_ptrs[0] again if it's not the boot cpu */
352 		memset(&paca_ptrs[0], 0x88, sizeof(paca_ptrs[0]));
353 	}
354 	setup_paca(paca_ptrs[boot_cpuid]);
355 	fixup_boot_paca();
356 
357 	/*
358 	 * Configure exception handlers. This include setting up trampolines
359 	 * if needed, setting exception endian mode, etc...
360 	 */
361 	configure_exceptions();
362 
363 	/*
364 	 * Configure Kernel Userspace Protection. This needs to happen before
365 	 * feature fixups for platforms that implement this using features.
366 	 */
367 	setup_kup();
368 
369 	/* Apply all the dynamic patching */
370 	apply_feature_fixups();
371 	setup_feature_keys();
372 
373 	/* Initialize the hash table or TLB handling */
374 	early_init_mmu();
375 
376 	early_ioremap_setup();
377 
378 	/*
379 	 * After firmware and early platform setup code has set things up,
380 	 * we note the SPR values for configurable control/performance
381 	 * registers, and use those as initial defaults.
382 	 */
383 	record_spr_defaults();
384 
385 	/*
386 	 * At this point, we can let interrupts switch to virtual mode
387 	 * (the MMU has been setup), so adjust the MSR in the PACA to
388 	 * have IR and DR set and enable AIL if it exists
389 	 */
390 	cpu_ready_for_interrupts();
391 
392 	/*
393 	 * We enable ftrace here, but since we only support DYNAMIC_FTRACE, it
394 	 * will only actually get enabled on the boot cpu much later once
395 	 * ftrace itself has been initialized.
396 	 */
397 	this_cpu_enable_ftrace();
398 
399 	udbg_printf(" <- %s()\n", __func__);
400 
401 #ifdef CONFIG_PPC_EARLY_DEBUG_BOOTX
402 	/*
403 	 * This needs to be done *last* (after the above udbg_printf() even)
404 	 *
405 	 * Right after we return from this function, we turn on the MMU
406 	 * which means the real-mode access trick that btext does will
407 	 * no longer work, it needs to switch to using a real MMU
408 	 * mapping. This call will ensure that it does
409 	 */
410 	btext_map();
411 #endif /* CONFIG_PPC_EARLY_DEBUG_BOOTX */
412 }
413 
414 #ifdef CONFIG_SMP
415 void early_setup_secondary(void)
416 {
417 	/* Mark interrupts disabled in PACA */
418 	irq_soft_mask_set(IRQS_DISABLED);
419 
420 	/* Initialize the hash table or TLB handling */
421 	early_init_mmu_secondary();
422 
423 	/* Perform any KUP setup that is per-cpu */
424 	setup_kup();
425 
426 	/*
427 	 * At this point, we can let interrupts switch to virtual mode
428 	 * (the MMU has been setup), so adjust the MSR in the PACA to
429 	 * have IR and DR set.
430 	 */
431 	cpu_ready_for_interrupts();
432 }
433 
434 #endif /* CONFIG_SMP */
435 
436 void panic_smp_self_stop(void)
437 {
438 	hard_irq_disable();
439 	spin_begin();
440 	while (1)
441 		spin_cpu_relax();
442 }
443 
444 #if defined(CONFIG_SMP) || defined(CONFIG_KEXEC_CORE)
445 static bool use_spinloop(void)
446 {
447 	if (IS_ENABLED(CONFIG_PPC_BOOK3S)) {
448 		/*
449 		 * See comments in head_64.S -- not all platforms insert
450 		 * secondaries at __secondary_hold and wait at the spin
451 		 * loop.
452 		 */
453 		if (firmware_has_feature(FW_FEATURE_OPAL))
454 			return false;
455 		return true;
456 	}
457 
458 	/*
459 	 * When book3e boots from kexec, the ePAPR spin table does
460 	 * not get used.
461 	 */
462 	return of_property_read_bool(of_chosen, "linux,booted-from-kexec");
463 }
464 
465 void smp_release_cpus(void)
466 {
467 	unsigned long *ptr;
468 	int i;
469 
470 	if (!use_spinloop())
471 		return;
472 
473 	/* All secondary cpus are spinning on a common spinloop, release them
474 	 * all now so they can start to spin on their individual paca
475 	 * spinloops. For non SMP kernels, the secondary cpus never get out
476 	 * of the common spinloop.
477 	 */
478 
479 	ptr  = (unsigned long *)((unsigned long)&__secondary_hold_spinloop
480 			- PHYSICAL_START);
481 	*ptr = ppc_function_entry(generic_secondary_smp_init);
482 
483 	/* And wait a bit for them to catch up */
484 	for (i = 0; i < 100000; i++) {
485 		mb();
486 		HMT_low();
487 		if (spinning_secondaries == 0)
488 			break;
489 		udelay(1);
490 	}
491 	pr_debug("spinning_secondaries = %d\n", spinning_secondaries);
492 }
493 #endif /* CONFIG_SMP || CONFIG_KEXEC_CORE */
494 
495 /*
496  * Initialize some remaining members of the ppc64_caches and systemcfg
497  * structures
498  * (at least until we get rid of them completely). This is mostly some
499  * cache informations about the CPU that will be used by cache flush
500  * routines and/or provided to userland
501  */
502 
503 static void init_cache_info(struct ppc_cache_info *info, u32 size, u32 lsize,
504 			    u32 bsize, u32 sets)
505 {
506 	info->size = size;
507 	info->sets = sets;
508 	info->line_size = lsize;
509 	info->block_size = bsize;
510 	info->log_block_size = __ilog2(bsize);
511 	if (bsize)
512 		info->blocks_per_page = PAGE_SIZE / bsize;
513 	else
514 		info->blocks_per_page = 0;
515 
516 	if (sets == 0)
517 		info->assoc = 0xffff;
518 	else
519 		info->assoc = size / (sets * lsize);
520 }
521 
522 static bool __init parse_cache_info(struct device_node *np,
523 				    bool icache,
524 				    struct ppc_cache_info *info)
525 {
526 	static const char *ipropnames[] __initdata = {
527 		"i-cache-size",
528 		"i-cache-sets",
529 		"i-cache-block-size",
530 		"i-cache-line-size",
531 	};
532 	static const char *dpropnames[] __initdata = {
533 		"d-cache-size",
534 		"d-cache-sets",
535 		"d-cache-block-size",
536 		"d-cache-line-size",
537 	};
538 	const char **propnames = icache ? ipropnames : dpropnames;
539 	const __be32 *sizep, *lsizep, *bsizep, *setsp;
540 	u32 size, lsize, bsize, sets;
541 	bool success = true;
542 
543 	size = 0;
544 	sets = -1u;
545 	lsize = bsize = cur_cpu_spec->dcache_bsize;
546 	sizep = of_get_property(np, propnames[0], NULL);
547 	if (sizep != NULL)
548 		size = be32_to_cpu(*sizep);
549 	setsp = of_get_property(np, propnames[1], NULL);
550 	if (setsp != NULL)
551 		sets = be32_to_cpu(*setsp);
552 	bsizep = of_get_property(np, propnames[2], NULL);
553 	lsizep = of_get_property(np, propnames[3], NULL);
554 	if (bsizep == NULL)
555 		bsizep = lsizep;
556 	if (lsizep == NULL)
557 		lsizep = bsizep;
558 	if (lsizep != NULL)
559 		lsize = be32_to_cpu(*lsizep);
560 	if (bsizep != NULL)
561 		bsize = be32_to_cpu(*bsizep);
562 	if (sizep == NULL || bsizep == NULL || lsizep == NULL)
563 		success = false;
564 
565 	/*
566 	 * OF is weird .. it represents fully associative caches
567 	 * as "1 way" which doesn't make much sense and doesn't
568 	 * leave room for direct mapped. We'll assume that 0
569 	 * in OF means direct mapped for that reason.
570 	 */
571 	if (sets == 1)
572 		sets = 0;
573 	else if (sets == 0)
574 		sets = 1;
575 
576 	init_cache_info(info, size, lsize, bsize, sets);
577 
578 	return success;
579 }
580 
581 void __init initialize_cache_info(void)
582 {
583 	struct device_node *cpu = NULL, *l2, *l3 = NULL;
584 	u32 pvr;
585 
586 	/*
587 	 * All shipping POWER8 machines have a firmware bug that
588 	 * puts incorrect information in the device-tree. This will
589 	 * be (hopefully) fixed for future chips but for now hard
590 	 * code the values if we are running on one of these
591 	 */
592 	pvr = PVR_VER(mfspr(SPRN_PVR));
593 	if (pvr == PVR_POWER8 || pvr == PVR_POWER8E ||
594 	    pvr == PVR_POWER8NVL) {
595 						/* size    lsize   blk  sets */
596 		init_cache_info(&ppc64_caches.l1i, 0x8000,   128,  128, 32);
597 		init_cache_info(&ppc64_caches.l1d, 0x10000,  128,  128, 64);
598 		init_cache_info(&ppc64_caches.l2,  0x80000,  128,  0,   512);
599 		init_cache_info(&ppc64_caches.l3,  0x800000, 128,  0,   8192);
600 	} else
601 		cpu = of_find_node_by_type(NULL, "cpu");
602 
603 	/*
604 	 * We're assuming *all* of the CPUs have the same
605 	 * d-cache and i-cache sizes... -Peter
606 	 */
607 	if (cpu) {
608 		if (!parse_cache_info(cpu, false, &ppc64_caches.l1d))
609 			pr_warn("Argh, can't find dcache properties !\n");
610 
611 		if (!parse_cache_info(cpu, true, &ppc64_caches.l1i))
612 			pr_warn("Argh, can't find icache properties !\n");
613 
614 		/*
615 		 * Try to find the L2 and L3 if any. Assume they are
616 		 * unified and use the D-side properties.
617 		 */
618 		l2 = of_find_next_cache_node(cpu);
619 		of_node_put(cpu);
620 		if (l2) {
621 			parse_cache_info(l2, false, &ppc64_caches.l2);
622 			l3 = of_find_next_cache_node(l2);
623 			of_node_put(l2);
624 		}
625 		if (l3) {
626 			parse_cache_info(l3, false, &ppc64_caches.l3);
627 			of_node_put(l3);
628 		}
629 	}
630 
631 	/* For use by binfmt_elf */
632 	dcache_bsize = ppc64_caches.l1d.block_size;
633 	icache_bsize = ppc64_caches.l1i.block_size;
634 
635 	cur_cpu_spec->dcache_bsize = dcache_bsize;
636 	cur_cpu_spec->icache_bsize = icache_bsize;
637 }
638 
639 /*
640  * This returns the limit below which memory accesses to the linear
641  * mapping are guarnateed not to cause an architectural exception (e.g.,
642  * TLB or SLB miss fault).
643  *
644  * This is used to allocate PACAs and various interrupt stacks that
645  * that are accessed early in interrupt handlers that must not cause
646  * re-entrant interrupts.
647  */
648 __init u64 ppc64_bolted_size(void)
649 {
650 #ifdef CONFIG_PPC_BOOK3E
651 	/* Freescale BookE bolts the entire linear mapping */
652 	/* XXX: BookE ppc64_rma_limit setup seems to disagree? */
653 	if (early_mmu_has_feature(MMU_FTR_TYPE_FSL_E))
654 		return linear_map_top;
655 	/* Other BookE, we assume the first GB is bolted */
656 	return 1ul << 30;
657 #else
658 	/* BookS radix, does not take faults on linear mapping */
659 	if (early_radix_enabled())
660 		return ULONG_MAX;
661 
662 	/* BookS hash, the first segment is bolted */
663 	if (early_mmu_has_feature(MMU_FTR_1T_SEGMENT))
664 		return 1UL << SID_SHIFT_1T;
665 	return 1UL << SID_SHIFT;
666 #endif
667 }
668 
669 static void *__init alloc_stack(unsigned long limit, int cpu)
670 {
671 	void *ptr;
672 
673 	BUILD_BUG_ON(STACK_INT_FRAME_SIZE % 16);
674 
675 	ptr = memblock_alloc_try_nid(THREAD_SIZE, THREAD_ALIGN,
676 				     MEMBLOCK_LOW_LIMIT, limit,
677 				     early_cpu_to_node(cpu));
678 	if (!ptr)
679 		panic("cannot allocate stacks");
680 
681 	return ptr;
682 }
683 
684 void __init irqstack_early_init(void)
685 {
686 	u64 limit = ppc64_bolted_size();
687 	unsigned int i;
688 
689 	/*
690 	 * Interrupt stacks must be in the first segment since we
691 	 * cannot afford to take SLB misses on them. They are not
692 	 * accessed in realmode.
693 	 */
694 	for_each_possible_cpu(i) {
695 		softirq_ctx[i] = alloc_stack(limit, i);
696 		hardirq_ctx[i] = alloc_stack(limit, i);
697 	}
698 }
699 
700 #ifdef CONFIG_PPC_BOOK3E
701 void __init exc_lvl_early_init(void)
702 {
703 	unsigned int i;
704 
705 	for_each_possible_cpu(i) {
706 		void *sp;
707 
708 		sp = alloc_stack(ULONG_MAX, i);
709 		critirq_ctx[i] = sp;
710 		paca_ptrs[i]->crit_kstack = sp + THREAD_SIZE;
711 
712 		sp = alloc_stack(ULONG_MAX, i);
713 		dbgirq_ctx[i] = sp;
714 		paca_ptrs[i]->dbg_kstack = sp + THREAD_SIZE;
715 
716 		sp = alloc_stack(ULONG_MAX, i);
717 		mcheckirq_ctx[i] = sp;
718 		paca_ptrs[i]->mc_kstack = sp + THREAD_SIZE;
719 	}
720 
721 	if (cpu_has_feature(CPU_FTR_DEBUG_LVL_EXC))
722 		patch_exception(0x040, exc_debug_debug_book3e);
723 }
724 #endif
725 
726 /*
727  * Stack space used when we detect a bad kernel stack pointer, and
728  * early in SMP boots before relocation is enabled. Exclusive emergency
729  * stack for machine checks.
730  */
731 void __init emergency_stack_init(void)
732 {
733 	u64 limit, mce_limit;
734 	unsigned int i;
735 
736 	/*
737 	 * Emergency stacks must be under 256MB, we cannot afford to take
738 	 * SLB misses on them. The ABI also requires them to be 128-byte
739 	 * aligned.
740 	 *
741 	 * Since we use these as temporary stacks during secondary CPU
742 	 * bringup, machine check, system reset, and HMI, we need to get
743 	 * at them in real mode. This means they must also be within the RMO
744 	 * region.
745 	 *
746 	 * The IRQ stacks allocated elsewhere in this file are zeroed and
747 	 * initialized in kernel/irq.c. These are initialized here in order
748 	 * to have emergency stacks available as early as possible.
749 	 */
750 	limit = mce_limit = min(ppc64_bolted_size(), ppc64_rma_size);
751 
752 	/*
753 	 * Machine check on pseries calls rtas, but can't use the static
754 	 * rtas_args due to a machine check hitting while the lock is held.
755 	 * rtas args have to be under 4GB, so the machine check stack is
756 	 * limited to 4GB so args can be put on stack.
757 	 */
758 	if (firmware_has_feature(FW_FEATURE_LPAR) && mce_limit > SZ_4G)
759 		mce_limit = SZ_4G;
760 
761 	for_each_possible_cpu(i) {
762 		paca_ptrs[i]->emergency_sp = alloc_stack(limit, i) + THREAD_SIZE;
763 
764 #ifdef CONFIG_PPC_BOOK3S_64
765 		/* emergency stack for NMI exception handling. */
766 		paca_ptrs[i]->nmi_emergency_sp = alloc_stack(limit, i) + THREAD_SIZE;
767 
768 		/* emergency stack for machine check exception handling. */
769 		paca_ptrs[i]->mc_emergency_sp = alloc_stack(mce_limit, i) + THREAD_SIZE;
770 #endif
771 	}
772 }
773 
774 #ifdef CONFIG_SMP
775 /**
776  * pcpu_alloc_bootmem - NUMA friendly alloc_bootmem wrapper for percpu
777  * @cpu: cpu to allocate for
778  * @size: size allocation in bytes
779  * @align: alignment
780  *
781  * Allocate @size bytes aligned at @align for cpu @cpu.  This wrapper
782  * does the right thing for NUMA regardless of the current
783  * configuration.
784  *
785  * RETURNS:
786  * Pointer to the allocated area on success, NULL on failure.
787  */
788 static void * __init pcpu_alloc_bootmem(unsigned int cpu, size_t size,
789 					size_t align)
790 {
791 	const unsigned long goal = __pa(MAX_DMA_ADDRESS);
792 #ifdef CONFIG_NUMA
793 	int node = early_cpu_to_node(cpu);
794 	void *ptr;
795 
796 	if (!node_online(node) || !NODE_DATA(node)) {
797 		ptr = memblock_alloc_from(size, align, goal);
798 		pr_info("cpu %d has no node %d or node-local memory\n",
799 			cpu, node);
800 		pr_debug("per cpu data for cpu%d %lu bytes at %016lx\n",
801 			 cpu, size, __pa(ptr));
802 	} else {
803 		ptr = memblock_alloc_try_nid(size, align, goal,
804 					     MEMBLOCK_ALLOC_ACCESSIBLE, node);
805 		pr_debug("per cpu data for cpu%d %lu bytes on node%d at "
806 			 "%016lx\n", cpu, size, node, __pa(ptr));
807 	}
808 	return ptr;
809 #else
810 	return memblock_alloc_from(size, align, goal);
811 #endif
812 }
813 
814 static void __init pcpu_free_bootmem(void *ptr, size_t size)
815 {
816 	memblock_free(__pa(ptr), size);
817 }
818 
819 static int pcpu_cpu_distance(unsigned int from, unsigned int to)
820 {
821 	if (early_cpu_to_node(from) == early_cpu_to_node(to))
822 		return LOCAL_DISTANCE;
823 	else
824 		return REMOTE_DISTANCE;
825 }
826 
827 unsigned long __per_cpu_offset[NR_CPUS] __read_mostly;
828 EXPORT_SYMBOL(__per_cpu_offset);
829 
830 static void __init pcpu_populate_pte(unsigned long addr)
831 {
832 	pgd_t *pgd = pgd_offset_k(addr);
833 	p4d_t *p4d;
834 	pud_t *pud;
835 	pmd_t *pmd;
836 
837 	p4d = p4d_offset(pgd, addr);
838 	if (p4d_none(*p4d)) {
839 		pud_t *new;
840 
841 		new = memblock_alloc(PUD_TABLE_SIZE, PUD_TABLE_SIZE);
842 		if (!new)
843 			goto err_alloc;
844 		p4d_populate(&init_mm, p4d, new);
845 	}
846 
847 	pud = pud_offset(p4d, addr);
848 	if (pud_none(*pud)) {
849 		pmd_t *new;
850 
851 		new = memblock_alloc(PMD_TABLE_SIZE, PMD_TABLE_SIZE);
852 		if (!new)
853 			goto err_alloc;
854 		pud_populate(&init_mm, pud, new);
855 	}
856 
857 	pmd = pmd_offset(pud, addr);
858 	if (!pmd_present(*pmd)) {
859 		pte_t *new;
860 
861 		new = memblock_alloc(PTE_TABLE_SIZE, PTE_TABLE_SIZE);
862 		if (!new)
863 			goto err_alloc;
864 		pmd_populate_kernel(&init_mm, pmd, new);
865 	}
866 
867 	return;
868 
869 err_alloc:
870 	panic("%s: Failed to allocate %lu bytes align=%lx from=%lx\n",
871 	      __func__, PAGE_SIZE, PAGE_SIZE, PAGE_SIZE);
872 }
873 
874 
875 void __init setup_per_cpu_areas(void)
876 {
877 	const size_t dyn_size = PERCPU_MODULE_RESERVE + PERCPU_DYNAMIC_RESERVE;
878 	size_t atom_size;
879 	unsigned long delta;
880 	unsigned int cpu;
881 	int rc = -EINVAL;
882 
883 	/*
884 	 * Linear mapping is one of 4K, 1M and 16M.  For 4K, no need
885 	 * to group units.  For larger mappings, use 1M atom which
886 	 * should be large enough to contain a number of units.
887 	 */
888 	if (mmu_linear_psize == MMU_PAGE_4K)
889 		atom_size = PAGE_SIZE;
890 	else
891 		atom_size = 1 << 20;
892 
893 	if (pcpu_chosen_fc != PCPU_FC_PAGE) {
894 		rc = pcpu_embed_first_chunk(0, dyn_size, atom_size, pcpu_cpu_distance,
895 					    pcpu_alloc_bootmem, pcpu_free_bootmem);
896 		if (rc)
897 			pr_warn("PERCPU: %s allocator failed (%d), "
898 				"falling back to page size\n",
899 				pcpu_fc_names[pcpu_chosen_fc], rc);
900 	}
901 
902 	if (rc < 0)
903 		rc = pcpu_page_first_chunk(0, pcpu_alloc_bootmem, pcpu_free_bootmem,
904 					   pcpu_populate_pte);
905 	if (rc < 0)
906 		panic("cannot initialize percpu area (err=%d)", rc);
907 
908 	delta = (unsigned long)pcpu_base_addr - (unsigned long)__per_cpu_start;
909 	for_each_possible_cpu(cpu) {
910                 __per_cpu_offset[cpu] = delta + pcpu_unit_offsets[cpu];
911 		paca_ptrs[cpu]->data_offset = __per_cpu_offset[cpu];
912 	}
913 }
914 #endif
915 
916 #ifdef CONFIG_MEMORY_HOTPLUG_SPARSE
917 unsigned long memory_block_size_bytes(void)
918 {
919 	if (ppc_md.memory_block_size)
920 		return ppc_md.memory_block_size();
921 
922 	return MIN_MEMORY_BLOCK_SIZE;
923 }
924 #endif
925 
926 #if defined(CONFIG_PPC_INDIRECT_PIO) || defined(CONFIG_PPC_INDIRECT_MMIO)
927 struct ppc_pci_io ppc_pci_io;
928 EXPORT_SYMBOL(ppc_pci_io);
929 #endif
930 
931 #ifdef CONFIG_HARDLOCKUP_DETECTOR_PERF
932 u64 hw_nmi_get_sample_period(int watchdog_thresh)
933 {
934 	return ppc_proc_freq * watchdog_thresh;
935 }
936 #endif
937 
938 /*
939  * The perf based hardlockup detector breaks PMU event based branches, so
940  * disable it by default. Book3S has a soft-nmi hardlockup detector based
941  * on the decrementer interrupt, so it does not suffer from this problem.
942  *
943  * It is likely to get false positives in KVM guests, so disable it there
944  * by default too. PowerVM will not stop or arbitrarily oversubscribe
945  * CPUs, but give a minimum regular allotment even with SPLPAR, so enable
946  * the detector for non-KVM guests, assume PowerVM.
947  */
948 static int __init disable_hardlockup_detector(void)
949 {
950 #ifdef CONFIG_HARDLOCKUP_DETECTOR_PERF
951 	hardlockup_detector_disable();
952 #else
953 	if (firmware_has_feature(FW_FEATURE_LPAR)) {
954 		if (is_kvm_guest())
955 			hardlockup_detector_disable();
956 	}
957 #endif
958 
959 	return 0;
960 }
961 early_initcall(disable_hardlockup_detector);
962