1 // SPDX-License-Identifier: GPL-2.0-only
3 * Based on arch/arm/mm/init.c
5 * Copyright (C) 1995-2005 Russell King
6 * Copyright (C) 2012 ARM Ltd.
9 #include <linux/kernel.h>
10 #include <linux/export.h>
11 #include <linux/errno.h>
12 #include <linux/swap.h>
13 #include <linux/init.h>
14 #include <linux/cache.h>
15 #include <linux/mman.h>
16 #include <linux/nodemask.h>
17 #include <linux/initrd.h>
18 #include <linux/gfp.h>
19 #include <linux/memblock.h>
20 #include <linux/sort.h>
22 #include <linux/of_fdt.h>
23 #include <linux/dma-direct.h>
24 #include <linux/dma-mapping.h>
25 #include <linux/dma-contiguous.h>
26 #include <linux/efi.h>
27 #include <linux/swiotlb.h>
28 #include <linux/vmalloc.h>
30 #include <linux/kexec.h>
31 #include <linux/crash_dump.h>
32 #include <linux/hugetlb.h>
35 #include <asm/fixmap.h>
36 #include <asm/kasan.h>
37 #include <asm/kernel-pgtable.h>
38 #include <asm/memory.h>
40 #include <asm/sections.h>
41 #include <asm/setup.h>
42 #include <linux/sizes.h>
44 #include <asm/alternative.h>
46 #define ARM64_ZONE_DMA_BITS 30
49 * We need to be able to catch inadvertent references to memstart_addr
50 * that occur (potentially in generic code) before arm64_memblock_init()
51 * executes, which assigns it its actual value. So use a default value
52 * that cannot be mistaken for a real physical address.
54 s64 memstart_addr __ro_after_init = -1;
55 EXPORT_SYMBOL(memstart_addr);
57 s64 physvirt_offset __ro_after_init;
58 EXPORT_SYMBOL(physvirt_offset);
60 struct page *vmemmap __ro_after_init;
61 EXPORT_SYMBOL(vmemmap);
64 * We create both ZONE_DMA and ZONE_DMA32. ZONE_DMA covers the first 1G of
65 * memory as some devices, namely the Raspberry Pi 4, have peripherals with
66 * this limited view of the memory. ZONE_DMA32 will cover the rest of the 32
67 * bit addressable memory area.
69 phys_addr_t arm64_dma_phys_limit __ro_after_init;
70 static phys_addr_t arm64_dma32_phys_limit __ro_after_init;
72 #ifdef CONFIG_KEXEC_CORE
74 * reserve_crashkernel() - reserves memory for crash kernel
76 * This function reserves memory area given in "crashkernel=" kernel command
77 * line parameter. The memory reserved is used by dump capture kernel when
78 * primary kernel is crashing.
80 static void __init reserve_crashkernel(void)
82 unsigned long long crash_base, crash_size;
85 ret = parse_crashkernel(boot_command_line, memblock_phys_mem_size(),
86 &crash_size, &crash_base);
87 /* no crashkernel= or invalid value specified */
88 if (ret || !crash_size)
91 crash_size = PAGE_ALIGN(crash_size);
93 if (crash_base == 0) {
94 /* Current arm64 boot protocol requires 2MB alignment */
95 crash_base = memblock_find_in_range(0, arm64_dma32_phys_limit,
97 if (crash_base == 0) {
98 pr_warn("cannot allocate crashkernel (size:0x%llx)\n",
103 /* User specifies base address explicitly. */
104 if (!memblock_is_region_memory(crash_base, crash_size)) {
105 pr_warn("cannot reserve crashkernel: region is not memory\n");
109 if (memblock_is_region_reserved(crash_base, crash_size)) {
110 pr_warn("cannot reserve crashkernel: region overlaps reserved memory\n");
114 if (!IS_ALIGNED(crash_base, SZ_2M)) {
115 pr_warn("cannot reserve crashkernel: base address is not 2MB aligned\n");
119 memblock_reserve(crash_base, crash_size);
121 pr_info("crashkernel reserved: 0x%016llx - 0x%016llx (%lld MB)\n",
122 crash_base, crash_base + crash_size, crash_size >> 20);
124 crashk_res.start = crash_base;
125 crashk_res.end = crash_base + crash_size - 1;
128 static void __init reserve_crashkernel(void)
131 #endif /* CONFIG_KEXEC_CORE */
133 #ifdef CONFIG_CRASH_DUMP
134 static int __init early_init_dt_scan_elfcorehdr(unsigned long node,
135 const char *uname, int depth, void *data)
140 if (depth != 1 || strcmp(uname, "chosen") != 0)
143 reg = of_get_flat_dt_prop(node, "linux,elfcorehdr", &len);
144 if (!reg || (len < (dt_root_addr_cells + dt_root_size_cells)))
147 elfcorehdr_addr = dt_mem_next_cell(dt_root_addr_cells, ®);
148 elfcorehdr_size = dt_mem_next_cell(dt_root_size_cells, ®);
154 * reserve_elfcorehdr() - reserves memory for elf core header
156 * This function reserves the memory occupied by an elf core header
157 * described in the device tree. This region contains all the
158 * information about primary kernel's core image and is used by a dump
159 * capture kernel to access the system memory on primary kernel.
161 static void __init reserve_elfcorehdr(void)
163 of_scan_flat_dt(early_init_dt_scan_elfcorehdr, NULL);
165 if (!elfcorehdr_size)
168 if (memblock_is_region_reserved(elfcorehdr_addr, elfcorehdr_size)) {
169 pr_warn("elfcorehdr is overlapped\n");
173 memblock_reserve(elfcorehdr_addr, elfcorehdr_size);
175 pr_info("Reserving %lldKB of memory at 0x%llx for elfcorehdr\n",
176 elfcorehdr_size >> 10, elfcorehdr_addr);
179 static void __init reserve_elfcorehdr(void)
182 #endif /* CONFIG_CRASH_DUMP */
185 * Return the maximum physical address for a zone with a given address size
186 * limit. It currently assumes that for memory starting above 4G, 32-bit
187 * devices will use a DMA offset.
189 static phys_addr_t __init max_zone_phys(unsigned int zone_bits)
191 phys_addr_t offset = memblock_start_of_DRAM() & GENMASK_ULL(63, zone_bits);
192 return min(offset + (1ULL << zone_bits), memblock_end_of_DRAM());
197 static void __init zone_sizes_init(unsigned long min, unsigned long max)
199 unsigned long max_zone_pfns[MAX_NR_ZONES] = {0};
201 #ifdef CONFIG_ZONE_DMA
202 max_zone_pfns[ZONE_DMA] = PFN_DOWN(arm64_dma_phys_limit);
204 #ifdef CONFIG_ZONE_DMA32
205 max_zone_pfns[ZONE_DMA32] = PFN_DOWN(arm64_dma32_phys_limit);
207 max_zone_pfns[ZONE_NORMAL] = max;
209 free_area_init_nodes(max_zone_pfns);
214 static void __init zone_sizes_init(unsigned long min, unsigned long max)
216 struct memblock_region *reg;
217 unsigned long zone_size[MAX_NR_ZONES], zhole_size[MAX_NR_ZONES];
218 unsigned long __maybe_unused max_dma, max_dma32;
220 memset(zone_size, 0, sizeof(zone_size));
222 max_dma = max_dma32 = min;
223 #ifdef CONFIG_ZONE_DMA
224 max_dma = max_dma32 = PFN_DOWN(arm64_dma_phys_limit);
225 zone_size[ZONE_DMA] = max_dma - min;
227 #ifdef CONFIG_ZONE_DMA32
228 max_dma32 = PFN_DOWN(arm64_dma32_phys_limit);
229 zone_size[ZONE_DMA32] = max_dma32 - max_dma;
231 zone_size[ZONE_NORMAL] = max - max_dma32;
233 memcpy(zhole_size, zone_size, sizeof(zhole_size));
235 for_each_memblock(memory, reg) {
236 unsigned long start = memblock_region_memory_base_pfn(reg);
237 unsigned long end = memblock_region_memory_end_pfn(reg);
239 #ifdef CONFIG_ZONE_DMA
240 if (start >= min && start < max_dma) {
241 unsigned long dma_end = min(end, max_dma);
242 zhole_size[ZONE_DMA] -= dma_end - start;
246 #ifdef CONFIG_ZONE_DMA32
247 if (start >= max_dma && start < max_dma32) {
248 unsigned long dma32_end = min(end, max_dma32);
249 zhole_size[ZONE_DMA32] -= dma32_end - start;
253 if (start >= max_dma32 && start < max) {
254 unsigned long normal_end = min(end, max);
255 zhole_size[ZONE_NORMAL] -= normal_end - start;
259 free_area_init_node(0, zone_size, min, zhole_size);
262 #endif /* CONFIG_NUMA */
264 int pfn_valid(unsigned long pfn)
266 phys_addr_t addr = pfn << PAGE_SHIFT;
268 if ((addr >> PAGE_SHIFT) != pfn)
271 #ifdef CONFIG_SPARSEMEM
272 if (pfn_to_section_nr(pfn) >= NR_MEM_SECTIONS)
275 if (!valid_section(__nr_to_section(pfn_to_section_nr(pfn))))
278 return memblock_is_map_memory(addr);
280 EXPORT_SYMBOL(pfn_valid);
282 static phys_addr_t memory_limit = PHYS_ADDR_MAX;
285 * Limit the memory size that was specified via FDT.
287 static int __init early_mem(char *p)
292 memory_limit = memparse(p, &p) & PAGE_MASK;
293 pr_notice("Memory limited to %lldMB\n", memory_limit >> 20);
297 early_param("mem", early_mem);
299 static int __init early_init_dt_scan_usablemem(unsigned long node,
300 const char *uname, int depth, void *data)
302 struct memblock_region *usablemem = data;
306 if (depth != 1 || strcmp(uname, "chosen") != 0)
309 reg = of_get_flat_dt_prop(node, "linux,usable-memory-range", &len);
310 if (!reg || (len < (dt_root_addr_cells + dt_root_size_cells)))
313 usablemem->base = dt_mem_next_cell(dt_root_addr_cells, ®);
314 usablemem->size = dt_mem_next_cell(dt_root_size_cells, ®);
319 static void __init fdt_enforce_memory_region(void)
321 struct memblock_region reg = {
325 of_scan_flat_dt(early_init_dt_scan_usablemem, ®);
328 memblock_cap_memory_range(reg.base, reg.size);
331 void __init arm64_memblock_init(void)
333 const s64 linear_region_size = BIT(vabits_actual - 1);
335 /* Handle linux,usable-memory-range property */
336 fdt_enforce_memory_region();
338 /* Remove memory above our supported physical address size */
339 memblock_remove(1ULL << PHYS_MASK_SHIFT, ULLONG_MAX);
342 * Select a suitable value for the base of physical memory.
344 memstart_addr = round_down(memblock_start_of_DRAM(),
345 ARM64_MEMSTART_ALIGN);
347 physvirt_offset = PHYS_OFFSET - PAGE_OFFSET;
349 vmemmap = ((struct page *)VMEMMAP_START - (memstart_addr >> PAGE_SHIFT));
352 * If we are running with a 52-bit kernel VA config on a system that
353 * does not support it, we have to offset our vmemmap and physvirt_offset
354 * s.t. we avoid the 52-bit portion of the direct linear map
356 if (IS_ENABLED(CONFIG_ARM64_VA_BITS_52) && (vabits_actual != 52)) {
357 vmemmap += (_PAGE_OFFSET(48) - _PAGE_OFFSET(52)) >> PAGE_SHIFT;
358 physvirt_offset = PHYS_OFFSET - _PAGE_OFFSET(48);
362 * Remove the memory that we will not be able to cover with the
363 * linear mapping. Take care not to clip the kernel which may be
366 memblock_remove(max_t(u64, memstart_addr + linear_region_size,
367 __pa_symbol(_end)), ULLONG_MAX);
368 if (memstart_addr + linear_region_size < memblock_end_of_DRAM()) {
369 /* ensure that memstart_addr remains sufficiently aligned */
370 memstart_addr = round_up(memblock_end_of_DRAM() - linear_region_size,
371 ARM64_MEMSTART_ALIGN);
372 memblock_remove(0, memstart_addr);
376 * Apply the memory limit if it was set. Since the kernel may be loaded
377 * high up in memory, add back the kernel region that must be accessible
378 * via the linear mapping.
380 if (memory_limit != PHYS_ADDR_MAX) {
381 memblock_mem_limit_remove_map(memory_limit);
382 memblock_add(__pa_symbol(_text), (u64)(_end - _text));
385 if (IS_ENABLED(CONFIG_BLK_DEV_INITRD) && phys_initrd_size) {
387 * Add back the memory we just removed if it results in the
388 * initrd to become inaccessible via the linear mapping.
389 * Otherwise, this is a no-op
391 u64 base = phys_initrd_start & PAGE_MASK;
392 u64 size = PAGE_ALIGN(phys_initrd_start + phys_initrd_size) - base;
395 * We can only add back the initrd memory if we don't end up
396 * with more memory than we can address via the linear mapping.
397 * It is up to the bootloader to position the kernel and the
398 * initrd reasonably close to each other (i.e., within 32 GB of
399 * each other) so that all granule/#levels combinations can
400 * always access both.
402 if (WARN(base < memblock_start_of_DRAM() ||
403 base + size > memblock_start_of_DRAM() +
405 "initrd not fully accessible via the linear mapping -- please check your bootloader ...\n")) {
406 phys_initrd_size = 0;
408 memblock_remove(base, size); /* clear MEMBLOCK_ flags */
409 memblock_add(base, size);
410 memblock_reserve(base, size);
414 if (IS_ENABLED(CONFIG_RANDOMIZE_BASE)) {
415 extern u16 memstart_offset_seed;
416 u64 range = linear_region_size -
417 (memblock_end_of_DRAM() - memblock_start_of_DRAM());
420 * If the size of the linear region exceeds, by a sufficient
421 * margin, the size of the region that the available physical
422 * memory spans, randomize the linear region as well.
424 if (memstart_offset_seed > 0 && range >= ARM64_MEMSTART_ALIGN) {
425 range /= ARM64_MEMSTART_ALIGN;
426 memstart_addr -= ARM64_MEMSTART_ALIGN *
427 ((range * memstart_offset_seed) >> 16);
432 * Register the kernel text, kernel data, initrd, and initial
433 * pagetables with memblock.
435 memblock_reserve(__pa_symbol(_text), _end - _text);
436 if (IS_ENABLED(CONFIG_BLK_DEV_INITRD) && phys_initrd_size) {
437 /* the generic initrd code expects virtual addresses */
438 initrd_start = __phys_to_virt(phys_initrd_start);
439 initrd_end = initrd_start + phys_initrd_size;
442 early_init_fdt_scan_reserved_mem();
444 if (IS_ENABLED(CONFIG_ZONE_DMA)) {
445 zone_dma_bits = ARM64_ZONE_DMA_BITS;
446 arm64_dma_phys_limit = max_zone_phys(ARM64_ZONE_DMA_BITS);
449 if (IS_ENABLED(CONFIG_ZONE_DMA32))
450 arm64_dma32_phys_limit = max_zone_phys(32);
452 arm64_dma32_phys_limit = PHYS_MASK + 1;
454 reserve_crashkernel();
456 reserve_elfcorehdr();
458 high_memory = __va(memblock_end_of_DRAM() - 1) + 1;
460 dma_contiguous_reserve(arm64_dma32_phys_limit);
462 #ifdef CONFIG_ARM64_4K_PAGES
463 hugetlb_cma_reserve(PUD_SHIFT - PAGE_SHIFT);
468 void __init bootmem_init(void)
470 unsigned long min, max;
472 min = PFN_UP(memblock_start_of_DRAM());
473 max = PFN_DOWN(memblock_end_of_DRAM());
475 early_memtest(min << PAGE_SHIFT, max << PAGE_SHIFT);
477 max_pfn = max_low_pfn = max;
482 * Sparsemem tries to allocate bootmem in memory_present(), so must be
483 * done after the fixed reservations.
488 zone_sizes_init(min, max);
493 #ifndef CONFIG_SPARSEMEM_VMEMMAP
494 static inline void free_memmap(unsigned long start_pfn, unsigned long end_pfn)
496 struct page *start_pg, *end_pg;
497 unsigned long pg, pgend;
500 * Convert start_pfn/end_pfn to a struct page pointer.
502 start_pg = pfn_to_page(start_pfn - 1) + 1;
503 end_pg = pfn_to_page(end_pfn - 1) + 1;
506 * Convert to physical addresses, and round start upwards and end
509 pg = (unsigned long)PAGE_ALIGN(__pa(start_pg));
510 pgend = (unsigned long)__pa(end_pg) & PAGE_MASK;
513 * If there are free pages between these, free the section of the
517 memblock_free(pg, pgend - pg);
521 * The mem_map array can get very big. Free the unused area of the memory map.
523 static void __init free_unused_memmap(void)
525 unsigned long start, prev_end = 0;
526 struct memblock_region *reg;
528 for_each_memblock(memory, reg) {
529 start = __phys_to_pfn(reg->base);
531 #ifdef CONFIG_SPARSEMEM
533 * Take care not to free memmap entries that don't exist due
534 * to SPARSEMEM sections which aren't present.
536 start = min(start, ALIGN(prev_end, PAGES_PER_SECTION));
539 * If we had a previous bank, and there is a space between the
540 * current bank and the previous, free it.
542 if (prev_end && prev_end < start)
543 free_memmap(prev_end, start);
546 * Align up here since the VM subsystem insists that the
547 * memmap entries are valid from the bank end aligned to
548 * MAX_ORDER_NR_PAGES.
550 prev_end = ALIGN(__phys_to_pfn(reg->base + reg->size),
554 #ifdef CONFIG_SPARSEMEM
555 if (!IS_ALIGNED(prev_end, PAGES_PER_SECTION))
556 free_memmap(prev_end, ALIGN(prev_end, PAGES_PER_SECTION));
559 #endif /* !CONFIG_SPARSEMEM_VMEMMAP */
562 * mem_init() marks the free areas in the mem_map and tells us how much memory
563 * is free. This is done after various parts of the system have claimed their
564 * memory after the kernel image.
566 void __init mem_init(void)
568 if (swiotlb_force == SWIOTLB_FORCE ||
569 max_pfn > PFN_DOWN(arm64_dma_phys_limit ? : arm64_dma32_phys_limit))
572 swiotlb_force = SWIOTLB_NO_FORCE;
574 set_max_mapnr(max_pfn - PHYS_PFN_OFFSET);
576 #ifndef CONFIG_SPARSEMEM_VMEMMAP
577 free_unused_memmap();
579 /* this will put all unused low memory onto the freelists */
582 mem_init_print_info(NULL);
585 * Check boundaries twice: Some fundamental inconsistencies can be
586 * detected at build time already.
589 BUILD_BUG_ON(TASK_SIZE_32 > DEFAULT_MAP_WINDOW_64);
592 if (PAGE_SIZE >= 16384 && get_num_physpages() <= 128) {
593 extern int sysctl_overcommit_memory;
595 * On a machine this small we won't get anywhere without
596 * overcommit, so turn it on by default.
598 sysctl_overcommit_memory = OVERCOMMIT_ALWAYS;
602 void free_initmem(void)
604 free_reserved_area(lm_alias(__init_begin),
605 lm_alias(__init_end),
606 POISON_FREE_INITMEM, "unused kernel");
608 * Unmap the __init region but leave the VM area in place. This
609 * prevents the region from being reused for kernel modules, which
610 * is not supported by kallsyms.
612 unmap_kernel_range((u64)__init_begin, (u64)(__init_end - __init_begin));
616 * Dump out memory limit information on panic.
618 static int dump_mem_limit(struct notifier_block *self, unsigned long v, void *p)
620 if (memory_limit != PHYS_ADDR_MAX) {
621 pr_emerg("Memory Limit: %llu MB\n", memory_limit >> 20);
623 pr_emerg("Memory Limit: none\n");
628 static struct notifier_block mem_limit_notifier = {
629 .notifier_call = dump_mem_limit,
632 static int __init register_mem_limit_dumper(void)
634 atomic_notifier_chain_register(&panic_notifier_list,
635 &mem_limit_notifier);
638 __initcall(register_mem_limit_dumper);