init_64.c

  1// SPDX-License-Identifier: GPL-2.0-or-later
  2/*
  3 *  PowerPC version
  4 *    Copyright (C) 1995-1996 Gary Thomas (gdt@linuxppc.org)
  5 *
  6 *  Modifications by Paul Mackerras (PowerMac) (paulus@cs.anu.edu.au)
  7 *  and Cort Dougan (PReP) (cort@cs.nmt.edu)
  8 *    Copyright (C) 1996 Paul Mackerras
  9 *
 10 *  Derived from "arch/i386/mm/init.c"
 11 *    Copyright (C) 1991, 1992, 1993, 1994  Linus Torvalds
 12 *
 13 *  Dave Engebretsen <engebret@us.ibm.com>
 14 *      Rework for PPC64 port.
 15 */
 16
 17#undef DEBUG
 18
 19#include <linux/signal.h>
 20#include <linux/sched.h>
 21#include <linux/kernel.h>
 22#include <linux/errno.h>
 23#include <linux/string.h>
 24#include <linux/types.h>
 25#include <linux/mman.h>
 26#include <linux/mm.h>
 27#include <linux/swap.h>
 28#include <linux/stddef.h>
 29#include <linux/vmalloc.h>
 30#include <linux/init.h>
 31#include <linux/delay.h>
 32#include <linux/highmem.h>
 33#include <linux/idr.h>
 34#include <linux/nodemask.h>
 35#include <linux/module.h>
 36#include <linux/poison.h>
 37#include <linux/memblock.h>
 38#include <linux/hugetlb.h>
 39#include <linux/slab.h>
 40#include <linux/of_fdt.h>
 41#include <linux/libfdt.h>
 42#include <linux/memremap.h>
 43#include <linux/memory.h>
 44#include <linux/bootmem_info.h>
 45
 46#include <asm/pgalloc.h>
 47#include <asm/page.h>
 48#include <asm/prom.h>
 49#include <asm/rtas.h>
 50#include <asm/io.h>
 51#include <asm/mmu_context.h>
 52#include <asm/mmu.h>
 53#include <linux/uaccess.h>
 54#include <asm/smp.h>
 55#include <asm/machdep.h>
 56#include <asm/tlb.h>
 57#include <asm/eeh.h>
 58#include <asm/processor.h>
 59#include <asm/mmzone.h>
 60#include <asm/cputable.h>
 61#include <asm/sections.h>
 62#include <asm/iommu.h>
 63#include <asm/vdso.h>
 64#include <asm/hugetlb.h>
 65
 66#include <mm/mmu_decl.h>
 67
 68#ifdef CONFIG_SPARSEMEM_VMEMMAP
 69/*
 70 * Given an address within the vmemmap, determine the page that
 71 * represents the start of the subsection it is within.  Note that we have to
 72 * do this by hand as the proffered address may not be correctly aligned.
 73 * Subtraction of non-aligned pointers produces undefined results.
 74 */
 75static struct page * __meminit vmemmap_subsection_start(unsigned long vmemmap_addr)
 76{
 77	unsigned long start_pfn;
 78	unsigned long offset = vmemmap_addr - ((unsigned long)(vmemmap));
 79
 80	/* Return the pfn of the start of the section. */
 81	start_pfn = (offset / sizeof(struct page)) & PAGE_SUBSECTION_MASK;
 82	return pfn_to_page(start_pfn);
 83}
 84
 85/*
 86 * Since memory is added in sub-section chunks, before creating a new vmemmap
 87 * mapping, the kernel should check whether there is an existing memmap mapping
 88 * covering the new subsection added. This is needed because kernel can map
 89 * vmemmap area using 16MB pages which will cover a memory range of 16G. Such
 90 * a range covers multiple subsections (2M)
 91 *
 92 * If any subsection in the 16G range mapped by vmemmap is valid we consider the
 93 * vmemmap populated (There is a page table entry already present). We can't do
 94 * a page table lookup here because with the hash translation we don't keep
 95 * vmemmap details in linux page table.
 96 */
 97int __meminit vmemmap_populated(unsigned long vmemmap_addr, int vmemmap_map_size)
 98{
 99	struct page *start;
100	unsigned long vmemmap_end = vmemmap_addr + vmemmap_map_size;
101	start = vmemmap_subsection_start(vmemmap_addr);
102
103	for (; (unsigned long)start < vmemmap_end; start += PAGES_PER_SUBSECTION)
104		/*
105		 * pfn valid check here is intended to really check
106		 * whether we have any subsection already initialized
107		 * in this range.
108		 */
109		if (pfn_valid(page_to_pfn(start)))
110			return 1;
111
112	return 0;
113}
114
115/*
116 * vmemmap virtual address space management does not have a traditional page
117 * table to track which virtual struct pages are backed by physical mapping.
118 * The virtual to physical mappings are tracked in a simple linked list
119 * format. 'vmemmap_list' maintains the entire vmemmap physical mapping at
120 * all times where as the 'next' list maintains the available
121 * vmemmap_backing structures which have been deleted from the
122 * 'vmemmap_global' list during system runtime (memory hotplug remove
123 * operation). The freed 'vmemmap_backing' structures are reused later when
124 * new requests come in without allocating fresh memory. This pointer also
125 * tracks the allocated 'vmemmap_backing' structures as we allocate one
126 * full page memory at a time when we dont have any.
127 */
128struct vmemmap_backing *vmemmap_list;
129static struct vmemmap_backing *next;
130
131/*
132 * The same pointer 'next' tracks individual chunks inside the allocated
133 * full page during the boot time and again tracks the freed nodes during
134 * runtime. It is racy but it does not happen as they are separated by the
135 * boot process. Will create problem if some how we have memory hotplug
136 * operation during boot !!
137 */
138static int num_left;
139static int num_freed;
140
141static __meminit struct vmemmap_backing * vmemmap_list_alloc(int node)
142{
143	struct vmemmap_backing *vmem_back;
144	/* get from freed entries first */
145	if (num_freed) {
146		num_freed--;
147		vmem_back = next;
148		next = next->list;
149
150		return vmem_back;
151	}
152
153	/* allocate a page when required and hand out chunks */
154	if (!num_left) {
155		next = vmemmap_alloc_block(PAGE_SIZE, node);
156		if (unlikely(!next)) {
157			WARN_ON(1);
158			return NULL;
159		}
160		num_left = PAGE_SIZE / sizeof(struct vmemmap_backing);
161	}
162
163	num_left--;
164
165	return next++;
166}
167
168static __meminit int vmemmap_list_populate(unsigned long phys,
169					   unsigned long start,
170					   int node)
171{
172	struct vmemmap_backing *vmem_back;
173
174	vmem_back = vmemmap_list_alloc(node);
175	if (unlikely(!vmem_back)) {
176		pr_debug("vmemap list allocation failed\n");
177		return -ENOMEM;
178	}
179
180	vmem_back->phys = phys;
181	vmem_back->virt_addr = start;
182	vmem_back->list = vmemmap_list;
183
184	vmemmap_list = vmem_back;
185	return 0;
186}
187
188bool altmap_cross_boundary(struct vmem_altmap *altmap, unsigned long start,
189			   unsigned long page_size)
190{
191	unsigned long nr_pfn = page_size / sizeof(struct page);
192	unsigned long start_pfn = page_to_pfn((struct page *)start);
193
194	if ((start_pfn + nr_pfn - 1) > altmap->end_pfn)
195		return true;
196
197	if (start_pfn < altmap->base_pfn)
198		return true;
199
200	return false;
201}
202
203static int __meminit __vmemmap_populate(unsigned long start, unsigned long end, int node,
204					struct vmem_altmap *altmap)
205{
206	bool altmap_alloc;
207	unsigned long page_size = 1 << mmu_psize_defs[mmu_vmemmap_psize].shift;
208
209	/* Align to the page size of the linear mapping. */
210	start = ALIGN_DOWN(start, page_size);
211
212	pr_debug("vmemmap_populate %lx..%lx, node %d\n", start, end, node);
213
214	for (; start < end; start += page_size) {
215		void *p = NULL;
216		int rc;
217
218		/*
219		 * This vmemmap range is backing different subsections. If any
220		 * of that subsection is marked valid, that means we already
221		 * have initialized a page table covering this range and hence
222		 * the vmemmap range is populated.
223		 */
224		if (vmemmap_populated(start, page_size))
225			continue;
226
227		/*
228		 * Allocate from the altmap first if we have one. This may
229		 * fail due to alignment issues when using 16MB hugepages, so
230		 * fall back to system memory if the altmap allocation fail.
231		 */
232		if (altmap && !altmap_cross_boundary(altmap, start, page_size)) {
233			p = vmemmap_alloc_block_buf(page_size, node, altmap);
234			if (!p)
235				pr_debug("altmap block allocation failed, falling back to system memory");
236			else
237				altmap_alloc = true;
238		}
239		if (!p) {
240			p = vmemmap_alloc_block_buf(page_size, node, NULL);
241			altmap_alloc = false;
242		}
243		if (!p)
244			return -ENOMEM;
245
246		if (vmemmap_list_populate(__pa(p), start, node)) {
247			/*
248			 * If we don't populate vmemap list, we don't have
249			 * the ability to free the allocated vmemmap
250			 * pages in section_deactivate. Hence free them
251			 * here.
252			 */
253			int nr_pfns = page_size >> PAGE_SHIFT;
254			unsigned long page_order = get_order(page_size);
255
256			if (altmap_alloc)
257				vmem_altmap_free(altmap, nr_pfns);
258			else
259				free_pages((unsigned long)p, page_order);
260			return -ENOMEM;
261		}
262
263		pr_debug("      * %016lx..%016lx allocated at %p\n",
264			 start, start + page_size, p);
265
266		rc = vmemmap_create_mapping(start, page_size, __pa(p));
267		if (rc < 0) {
268			pr_warn("%s: Unable to create vmemmap mapping: %d\n",
269				__func__, rc);
270			return -EFAULT;
271		}
272	}
273
274	return 0;
275}
276
277int __meminit vmemmap_populate(unsigned long start, unsigned long end, int node,
278			       struct vmem_altmap *altmap)
279{
280
281#ifdef CONFIG_PPC_BOOK3S_64
282	if (radix_enabled())
283		return radix__vmemmap_populate(start, end, node, altmap);
284#endif
285
286	return __vmemmap_populate(start, end, node, altmap);
287}
288
289#ifdef CONFIG_MEMORY_HOTPLUG
290static unsigned long vmemmap_list_free(unsigned long start)
291{
292	struct vmemmap_backing *vmem_back, *vmem_back_prev;
293
294	vmem_back_prev = vmem_back = vmemmap_list;
295
296	/* look for it with prev pointer recorded */
297	for (; vmem_back; vmem_back = vmem_back->list) {
298		if (vmem_back->virt_addr == start)
299			break;
300		vmem_back_prev = vmem_back;
301	}
302
303	if (unlikely(!vmem_back))
304		return 0;
305
306	/* remove it from vmemmap_list */
307	if (vmem_back == vmemmap_list) /* remove head */
308		vmemmap_list = vmem_back->list;
309	else
310		vmem_back_prev->list = vmem_back->list;
311
312	/* next point to this freed entry */
313	vmem_back->list = next;
314	next = vmem_back;
315	num_freed++;
316
317	return vmem_back->phys;
318}
319
320static void __ref __vmemmap_free(unsigned long start, unsigned long end,
321				 struct vmem_altmap *altmap)
322{
323	unsigned long page_size = 1 << mmu_psize_defs[mmu_vmemmap_psize].shift;
324	unsigned long page_order = get_order(page_size);
325	unsigned long alt_start = ~0, alt_end = ~0;
326	unsigned long base_pfn;
327
328	start = ALIGN_DOWN(start, page_size);
329	if (altmap) {
330		alt_start = altmap->base_pfn;
331		alt_end = altmap->base_pfn + altmap->reserve + altmap->free;
332	}
333
334	pr_debug("vmemmap_free %lx...%lx\n", start, end);
335
336	for (; start < end; start += page_size) {
337		unsigned long nr_pages, addr;
338		struct page *page;
339
340		/*
341		 * We have already marked the subsection we are trying to remove
342		 * invalid. So if we want to remove the vmemmap range, we
343		 * need to make sure there is no subsection marked valid
344		 * in this range.
345		 */
346		if (vmemmap_populated(start, page_size))
347			continue;
348
349		addr = vmemmap_list_free(start);
350		if (!addr)
351			continue;
352
353		page = pfn_to_page(addr >> PAGE_SHIFT);
354		nr_pages = 1 << page_order;
355		base_pfn = PHYS_PFN(addr);
356
357		if (base_pfn >= alt_start && base_pfn < alt_end) {
358			vmem_altmap_free(altmap, nr_pages);
359		} else if (PageReserved(page)) {
360			/* allocated from bootmem */
361			if (page_size < PAGE_SIZE) {
362				/*
363				 * this shouldn't happen, but if it is
364				 * the case, leave the memory there
365				 */
366				WARN_ON_ONCE(1);
367			} else {
368				while (nr_pages--)
369					free_reserved_page(page++);
370			}
371		} else {
372			free_pages((unsigned long)(__va(addr)), page_order);
373		}
374
375		vmemmap_remove_mapping(start, page_size);
376	}
377}
378
379void __ref vmemmap_free(unsigned long start, unsigned long end,
380			struct vmem_altmap *altmap)
381{
382#ifdef CONFIG_PPC_BOOK3S_64
383	if (radix_enabled())
384		return radix__vmemmap_free(start, end, altmap);
385#endif
386	return __vmemmap_free(start, end, altmap);
387}
388
389#endif
390
391#ifdef CONFIG_HAVE_BOOTMEM_INFO_NODE
392void register_page_bootmem_memmap(unsigned long section_nr,
393				  struct page *start_page, unsigned long size)
394{
395}
396#endif /* CONFIG_HAVE_BOOTMEM_INFO_NODE */
397
398#endif /* CONFIG_SPARSEMEM_VMEMMAP */
399
400#ifdef CONFIG_PPC_BOOK3S_64
401unsigned int mmu_lpid_bits;
402#ifdef CONFIG_KVM_BOOK3S_HV_POSSIBLE
403EXPORT_SYMBOL_GPL(mmu_lpid_bits);
404#endif
405unsigned int mmu_pid_bits;
406
407static bool disable_radix = !IS_ENABLED(CONFIG_PPC_RADIX_MMU_DEFAULT);
408
409static int __init parse_disable_radix(char *p)
410{
411	bool val;
412
413	if (!p)
414		val = true;
415	else if (kstrtobool(p, &val))
416		return -EINVAL;
417
418	disable_radix = val;
419
420	return 0;
421}
422early_param("disable_radix", parse_disable_radix);
423
424/*
425 * If we're running under a hypervisor, we need to check the contents of
426 * /chosen/ibm,architecture-vec-5 to see if the hypervisor is willing to do
427 * radix.  If not, we clear the radix feature bit so we fall back to hash.
428 */
429static void __init early_check_vec5(void)
430{
431	unsigned long root, chosen;
432	int size;
433	const u8 *vec5;
434	u8 mmu_supported;
435
436	root = of_get_flat_dt_root();
437	chosen = of_get_flat_dt_subnode_by_name(root, "chosen");
438	if (chosen == -FDT_ERR_NOTFOUND) {
439		cur_cpu_spec->mmu_features &= ~MMU_FTR_TYPE_RADIX;
440		return;
441	}
442	vec5 = of_get_flat_dt_prop(chosen, "ibm,architecture-vec-5", &size);
443	if (!vec5) {
444		cur_cpu_spec->mmu_features &= ~MMU_FTR_TYPE_RADIX;
445		return;
446	}
447	if (size <= OV5_INDX(OV5_MMU_SUPPORT)) {
448		cur_cpu_spec->mmu_features &= ~MMU_FTR_TYPE_RADIX;
449		return;
450	}
451
452	/* Check for supported configuration */
453	mmu_supported = vec5[OV5_INDX(OV5_MMU_SUPPORT)] &
454			OV5_FEAT(OV5_MMU_SUPPORT);
455	if (mmu_supported == OV5_FEAT(OV5_MMU_RADIX)) {
456		/* Hypervisor only supports radix - check enabled && GTSE */
457		if (!early_radix_enabled()) {
458			pr_warn("WARNING: Ignoring cmdline option disable_radix\n");
459		}
460		if (!(vec5[OV5_INDX(OV5_RADIX_GTSE)] &
461						OV5_FEAT(OV5_RADIX_GTSE))) {
462			cur_cpu_spec->mmu_features &= ~MMU_FTR_GTSE;
463		} else
464			cur_cpu_spec->mmu_features |= MMU_FTR_GTSE;
465		/* Do radix anyway - the hypervisor said we had to */
466		cur_cpu_spec->mmu_features |= MMU_FTR_TYPE_RADIX;
467	} else if (mmu_supported == OV5_FEAT(OV5_MMU_HASH)) {
468		/* Hypervisor only supports hash - disable radix */
469		cur_cpu_spec->mmu_features &= ~MMU_FTR_TYPE_RADIX;
470		cur_cpu_spec->mmu_features &= ~MMU_FTR_GTSE;
471	}
472}
473
474static int __init dt_scan_mmu_pid_width(unsigned long node,
475					   const char *uname, int depth,
476					   void *data)
477{
478	int size = 0;
479	const __be32 *prop;
480	const char *type = of_get_flat_dt_prop(node, "device_type", NULL);
481
482	/* We are scanning "cpu" nodes only */
483	if (type == NULL || strcmp(type, "cpu") != 0)
484		return 0;
485
486	/* Find MMU LPID, PID register size */
487	prop = of_get_flat_dt_prop(node, "ibm,mmu-lpid-bits", &size);
488	if (prop && size == 4)
489		mmu_lpid_bits = be32_to_cpup(prop);
490
491	prop = of_get_flat_dt_prop(node, "ibm,mmu-pid-bits", &size);
492	if (prop && size == 4)
493		mmu_pid_bits = be32_to_cpup(prop);
494
495	if (!mmu_pid_bits && !mmu_lpid_bits)
496		return 0;
497
498	return 1;
499}
500
501/*
502 * Outside hotplug the kernel uses this value to map the kernel direct map
503 * with radix. To be compatible with older kernels, let's keep this value
504 * as 16M which is also SECTION_SIZE with SPARSEMEM. We can ideally map
505 * things with 1GB size in the case where we don't support hotplug.
506 */
507#ifndef CONFIG_MEMORY_HOTPLUG
508#define DEFAULT_MEMORY_BLOCK_SIZE	SZ_16M
509#else
510#define DEFAULT_MEMORY_BLOCK_SIZE	MIN_MEMORY_BLOCK_SIZE
511#endif
512
513static void update_memory_block_size(unsigned long *block_size, unsigned long mem_size)
514{
515	unsigned long min_memory_block_size = DEFAULT_MEMORY_BLOCK_SIZE;
516
517	for (; *block_size > min_memory_block_size; *block_size >>= 2) {
518		if ((mem_size & *block_size) == 0)
519			break;
520	}
521}
522
523static int __init probe_memory_block_size(unsigned long node, const char *uname, int
524					  depth, void *data)
525{
526	const char *type;
527	unsigned long *block_size = (unsigned long *)data;
528	const __be32 *reg, *endp;
529	int l;
530
531	if (depth != 1)
532		return 0;
533	/*
534	 * If we have dynamic-reconfiguration-memory node, use the
535	 * lmb value.
536	 */
537	if (strcmp(uname, "ibm,dynamic-reconfiguration-memory") == 0) {
538
539		const __be32 *prop;
540
541		prop = of_get_flat_dt_prop(node, "ibm,lmb-size", &l);
542
543		if (!prop || l < dt_root_size_cells * sizeof(__be32))
544			/*
545			 * Nothing in the device tree
546			 */
547			*block_size = DEFAULT_MEMORY_BLOCK_SIZE;
548		else
549			*block_size = of_read_number(prop, dt_root_size_cells);
550		/*
551		 * We have found the final value. Don't probe further.
552		 */
553		return 1;
554	}
555	/*
556	 * Find all the device tree nodes of memory type and make sure
557	 * the area can be mapped using the memory block size value
558	 * we end up using. We start with 1G value and keep reducing
559	 * it such that we can map the entire area using memory_block_size.
560	 * This will be used on powernv and older pseries that don't
561	 * have ibm,lmb-size node.
562	 * For ex: with P5 we can end up with
563	 * memory@0 -> 128MB
564	 * memory@128M -> 64M
565	 * This will end up using 64MB  memory block size value.
566	 */
567	type = of_get_flat_dt_prop(node, "device_type", NULL);
568	if (type == NULL || strcmp(type, "memory") != 0)
569		return 0;
570
571	reg = of_get_flat_dt_prop(node, "linux,usable-memory", &l);
572	if (!reg)
573		reg = of_get_flat_dt_prop(node, "reg", &l);
574	if (!reg)
575		return 0;
576
577	endp = reg + (l / sizeof(__be32));
578	while ((endp - reg) >= (dt_root_addr_cells + dt_root_size_cells)) {
579		const char *compatible;
580		u64 size;
581
582		dt_mem_next_cell(dt_root_addr_cells, &reg);
583		size = dt_mem_next_cell(dt_root_size_cells, &reg);
584
585		if (size) {
586			update_memory_block_size(block_size, size);
587			continue;
588		}
589		/*
590		 * ibm,coherent-device-memory with linux,usable-memory = 0
591		 * Force 256MiB block size. Work around for GPUs on P9 PowerNV
592		 * linux,usable-memory == 0 implies driver managed memory and
593		 * we can't use large memory block size due to hotplug/unplug
594		 * limitations.
595		 */
596		compatible = of_get_flat_dt_prop(node, "compatible", NULL);
597		if (compatible && !strcmp(compatible, "ibm,coherent-device-memory")) {
598			if (*block_size > SZ_256M)
599				*block_size = SZ_256M;
600			/*
601			 * We keep 256M as the upper limit with GPU present.
602			 */
603			return 0;
604		}
605	}
606	/* continue looking for other memory device types */
607	return 0;
608}
609
610/*
611 * start with 1G memory block size. Early init will
612 * fix this with correct value.
613 */
614unsigned long memory_block_size __ro_after_init = 1UL << 30;
615static void __init early_init_memory_block_size(void)
616{
617	/*
618	 * We need to do memory_block_size probe early so that
619	 * radix__early_init_mmu() can use this as limit for
620	 * mapping page size.
621	 */
622	of_scan_flat_dt(probe_memory_block_size, &memory_block_size);
623}
624
625void __init mmu_early_init_devtree(void)
626{
627	bool hvmode = !!(mfmsr() & MSR_HV);
628
629	/* Disable radix mode based on kernel command line. */
630	if (disable_radix) {
631		if (IS_ENABLED(CONFIG_PPC_64S_HASH_MMU))
632			cur_cpu_spec->mmu_features &= ~MMU_FTR_TYPE_RADIX;
633		else
634			pr_warn("WARNING: Ignoring cmdline option disable_radix\n");
635	}
636
637	of_scan_flat_dt(dt_scan_mmu_pid_width, NULL);
638	if (hvmode && !mmu_lpid_bits) {
639		if (early_cpu_has_feature(CPU_FTR_ARCH_207S))
640			mmu_lpid_bits = 12; /* POWER8-10 */
641		else
642			mmu_lpid_bits = 10; /* POWER7 */
643	}
644	if (!mmu_pid_bits) {
645		if (early_cpu_has_feature(CPU_FTR_ARCH_300))
646			mmu_pid_bits = 20; /* POWER9-10 */
647	}
648
649	/*
650	 * Check /chosen/ibm,architecture-vec-5 if running as a guest.
651	 * When running bare-metal, we can use radix if we like
652	 * even though the ibm,architecture-vec-5 property created by
653	 * skiboot doesn't have the necessary bits set.
654	 */
655	if (!hvmode)
656		early_check_vec5();
657
658	early_init_memory_block_size();
659
660	if (early_radix_enabled()) {
661		radix__early_init_devtree();
662
663		/*
664		 * We have finalized the translation we are going to use by now.
665		 * Radix mode is not limited by RMA / VRMA addressing.
666		 * Hence don't limit memblock allocations.
667		 */
668		ppc64_rma_size = ULONG_MAX;
669		memblock_set_current_limit(MEMBLOCK_ALLOC_ANYWHERE);
670	} else
671		hash__early_init_devtree();
672
673	if (IS_ENABLED(CONFIG_HUGETLB_PAGE_SIZE_VARIABLE))
674		hugetlbpage_init_defaultsize();
675
676	if (!(cur_cpu_spec->mmu_features & MMU_FTR_HPTE_TABLE) &&
677	    !(cur_cpu_spec->mmu_features & MMU_FTR_TYPE_RADIX))
678		panic("kernel does not support any MMU type offered by platform");
679}
680#endif /* CONFIG_PPC_BOOK3S_64 */