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-mapping.h>
24 #include <linux/dma-contiguous.h>
25 #include <linux/efi.h>
26 #include <linux/swiotlb.h>
27 #include <linux/vmalloc.h>
29 #include <linux/kexec.h>
30 #include <linux/crash_dump.h>
33 #include <asm/fixmap.h>
34 #include <asm/kasan.h>
35 #include <asm/kernel-pgtable.h>
36 #include <asm/memory.h>
38 #include <asm/sections.h>
39 #include <asm/setup.h>
40 #include <linux/sizes.h>
42 #include <asm/alternative.h>
45 * We need to be able to catch inadvertent references to memstart_addr
46 * that occur (potentially in generic code) before arm64_memblock_init()
47 * executes, which assigns it its actual value. So use a default value
48 * that cannot be mistaken for a real physical address.
50 s64 memstart_addr __ro_after_init = -1;
51 EXPORT_SYMBOL(memstart_addr);
53 s64 physvirt_offset __ro_after_init;
54 EXPORT_SYMBOL(physvirt_offset);
56 struct page *vmemmap __ro_after_init;
57 EXPORT_SYMBOL(vmemmap);
59 phys_addr_t arm64_dma_phys_limit __ro_after_init;
61 #ifdef CONFIG_KEXEC_CORE
63 * reserve_crashkernel() - reserves memory for crash kernel
65 * This function reserves memory area given in "crashkernel=" kernel command
66 * line parameter. The memory reserved is used by dump capture kernel when
67 * primary kernel is crashing.
69 static void __init reserve_crashkernel(void)
71 unsigned long long crash_base, crash_size;
74 ret = parse_crashkernel(boot_command_line, memblock_phys_mem_size(),
75 &crash_size, &crash_base);
76 /* no crashkernel= or invalid value specified */
77 if (ret || !crash_size)
80 crash_size = PAGE_ALIGN(crash_size);
82 if (crash_base == 0) {
83 /* Current arm64 boot protocol requires 2MB alignment */
84 crash_base = memblock_find_in_range(0, ARCH_LOW_ADDRESS_LIMIT,
86 if (crash_base == 0) {
87 pr_warn("cannot allocate crashkernel (size:0x%llx)\n",
92 /* User specifies base address explicitly. */
93 if (!memblock_is_region_memory(crash_base, crash_size)) {
94 pr_warn("cannot reserve crashkernel: region is not memory\n");
98 if (memblock_is_region_reserved(crash_base, crash_size)) {
99 pr_warn("cannot reserve crashkernel: region overlaps reserved memory\n");
103 if (!IS_ALIGNED(crash_base, SZ_2M)) {
104 pr_warn("cannot reserve crashkernel: base address is not 2MB aligned\n");
108 memblock_reserve(crash_base, crash_size);
110 pr_info("crashkernel reserved: 0x%016llx - 0x%016llx (%lld MB)\n",
111 crash_base, crash_base + crash_size, crash_size >> 20);
113 crashk_res.start = crash_base;
114 crashk_res.end = crash_base + crash_size - 1;
117 static void __init reserve_crashkernel(void)
120 #endif /* CONFIG_KEXEC_CORE */
122 #ifdef CONFIG_CRASH_DUMP
123 static int __init early_init_dt_scan_elfcorehdr(unsigned long node,
124 const char *uname, int depth, void *data)
129 if (depth != 1 || strcmp(uname, "chosen") != 0)
132 reg = of_get_flat_dt_prop(node, "linux,elfcorehdr", &len);
133 if (!reg || (len < (dt_root_addr_cells + dt_root_size_cells)))
136 elfcorehdr_addr = dt_mem_next_cell(dt_root_addr_cells, ®);
137 elfcorehdr_size = dt_mem_next_cell(dt_root_size_cells, ®);
143 * reserve_elfcorehdr() - reserves memory for elf core header
145 * This function reserves the memory occupied by an elf core header
146 * described in the device tree. This region contains all the
147 * information about primary kernel's core image and is used by a dump
148 * capture kernel to access the system memory on primary kernel.
150 static void __init reserve_elfcorehdr(void)
152 of_scan_flat_dt(early_init_dt_scan_elfcorehdr, NULL);
154 if (!elfcorehdr_size)
157 if (memblock_is_region_reserved(elfcorehdr_addr, elfcorehdr_size)) {
158 pr_warn("elfcorehdr is overlapped\n");
162 memblock_reserve(elfcorehdr_addr, elfcorehdr_size);
164 pr_info("Reserving %lldKB of memory at 0x%llx for elfcorehdr\n",
165 elfcorehdr_size >> 10, elfcorehdr_addr);
168 static void __init reserve_elfcorehdr(void)
171 #endif /* CONFIG_CRASH_DUMP */
173 * Return the maximum physical address for ZONE_DMA32 (DMA_BIT_MASK(32)). It
174 * currently assumes that for memory starting above 4G, 32-bit devices will
177 static phys_addr_t __init max_zone_dma_phys(void)
179 phys_addr_t offset = memblock_start_of_DRAM() & GENMASK_ULL(63, 32);
180 return min(offset + (1ULL << 32), memblock_end_of_DRAM());
185 static void __init zone_sizes_init(unsigned long min, unsigned long max)
187 unsigned long max_zone_pfns[MAX_NR_ZONES] = {0};
189 #ifdef CONFIG_ZONE_DMA32
190 max_zone_pfns[ZONE_DMA32] = PFN_DOWN(max_zone_dma_phys());
192 max_zone_pfns[ZONE_NORMAL] = max;
194 free_area_init_nodes(max_zone_pfns);
199 static void __init zone_sizes_init(unsigned long min, unsigned long max)
201 struct memblock_region *reg;
202 unsigned long zone_size[MAX_NR_ZONES], zhole_size[MAX_NR_ZONES];
203 unsigned long max_dma = min;
205 memset(zone_size, 0, sizeof(zone_size));
207 /* 4GB maximum for 32-bit only capable devices */
208 #ifdef CONFIG_ZONE_DMA32
209 max_dma = PFN_DOWN(arm64_dma_phys_limit);
210 zone_size[ZONE_DMA32] = max_dma - min;
212 zone_size[ZONE_NORMAL] = max - max_dma;
214 memcpy(zhole_size, zone_size, sizeof(zhole_size));
216 for_each_memblock(memory, reg) {
217 unsigned long start = memblock_region_memory_base_pfn(reg);
218 unsigned long end = memblock_region_memory_end_pfn(reg);
223 #ifdef CONFIG_ZONE_DMA32
224 if (start < max_dma) {
225 unsigned long dma_end = min(end, max_dma);
226 zhole_size[ZONE_DMA32] -= dma_end - start;
230 unsigned long normal_end = min(end, max);
231 unsigned long normal_start = max(start, max_dma);
232 zhole_size[ZONE_NORMAL] -= normal_end - normal_start;
236 free_area_init_node(0, zone_size, min, zhole_size);
239 #endif /* CONFIG_NUMA */
241 int pfn_valid(unsigned long pfn)
243 phys_addr_t addr = pfn << PAGE_SHIFT;
245 if ((addr >> PAGE_SHIFT) != pfn)
248 #ifdef CONFIG_SPARSEMEM
249 if (pfn_to_section_nr(pfn) >= NR_MEM_SECTIONS)
252 if (!valid_section(__nr_to_section(pfn_to_section_nr(pfn))))
255 return memblock_is_map_memory(addr);
257 EXPORT_SYMBOL(pfn_valid);
259 static phys_addr_t memory_limit = PHYS_ADDR_MAX;
262 * Limit the memory size that was specified via FDT.
264 static int __init early_mem(char *p)
269 memory_limit = memparse(p, &p) & PAGE_MASK;
270 pr_notice("Memory limited to %lldMB\n", memory_limit >> 20);
274 early_param("mem", early_mem);
276 static int __init early_init_dt_scan_usablemem(unsigned long node,
277 const char *uname, int depth, void *data)
279 struct memblock_region *usablemem = data;
283 if (depth != 1 || strcmp(uname, "chosen") != 0)
286 reg = of_get_flat_dt_prop(node, "linux,usable-memory-range", &len);
287 if (!reg || (len < (dt_root_addr_cells + dt_root_size_cells)))
290 usablemem->base = dt_mem_next_cell(dt_root_addr_cells, ®);
291 usablemem->size = dt_mem_next_cell(dt_root_size_cells, ®);
296 static void __init fdt_enforce_memory_region(void)
298 struct memblock_region reg = {
302 of_scan_flat_dt(early_init_dt_scan_usablemem, ®);
305 memblock_cap_memory_range(reg.base, reg.size);
308 void __init arm64_memblock_init(void)
310 const s64 linear_region_size = BIT(vabits_actual - 1);
312 /* Handle linux,usable-memory-range property */
313 fdt_enforce_memory_region();
315 /* Remove memory above our supported physical address size */
316 memblock_remove(1ULL << PHYS_MASK_SHIFT, ULLONG_MAX);
319 * Select a suitable value for the base of physical memory.
321 memstart_addr = round_down(memblock_start_of_DRAM(),
322 ARM64_MEMSTART_ALIGN);
324 physvirt_offset = PHYS_OFFSET - PAGE_OFFSET;
326 vmemmap = ((struct page *)VMEMMAP_START - (memstart_addr >> PAGE_SHIFT));
329 * If we are running with a 52-bit kernel VA config on a system that
330 * does not support it, we have to offset our vmemmap and physvirt_offset
331 * s.t. we avoid the 52-bit portion of the direct linear map
333 if (IS_ENABLED(CONFIG_ARM64_VA_BITS_52) && (vabits_actual != 52)) {
334 vmemmap += (_PAGE_OFFSET(48) - _PAGE_OFFSET(52)) >> PAGE_SHIFT;
335 physvirt_offset = PHYS_OFFSET - _PAGE_OFFSET(48);
339 * Remove the memory that we will not be able to cover with the
340 * linear mapping. Take care not to clip the kernel which may be
343 memblock_remove(max_t(u64, memstart_addr + linear_region_size,
344 __pa_symbol(_end)), ULLONG_MAX);
345 if (memstart_addr + linear_region_size < memblock_end_of_DRAM()) {
346 /* ensure that memstart_addr remains sufficiently aligned */
347 memstart_addr = round_up(memblock_end_of_DRAM() - linear_region_size,
348 ARM64_MEMSTART_ALIGN);
349 memblock_remove(0, memstart_addr);
353 * Apply the memory limit if it was set. Since the kernel may be loaded
354 * high up in memory, add back the kernel region that must be accessible
355 * via the linear mapping.
357 if (memory_limit != PHYS_ADDR_MAX) {
358 memblock_mem_limit_remove_map(memory_limit);
359 memblock_add(__pa_symbol(_text), (u64)(_end - _text));
362 if (IS_ENABLED(CONFIG_BLK_DEV_INITRD) && phys_initrd_size) {
364 * Add back the memory we just removed if it results in the
365 * initrd to become inaccessible via the linear mapping.
366 * Otherwise, this is a no-op
368 u64 base = phys_initrd_start & PAGE_MASK;
369 u64 size = PAGE_ALIGN(phys_initrd_start + phys_initrd_size) - base;
372 * We can only add back the initrd memory if we don't end up
373 * with more memory than we can address via the linear mapping.
374 * It is up to the bootloader to position the kernel and the
375 * initrd reasonably close to each other (i.e., within 32 GB of
376 * each other) so that all granule/#levels combinations can
377 * always access both.
379 if (WARN(base < memblock_start_of_DRAM() ||
380 base + size > memblock_start_of_DRAM() +
382 "initrd not fully accessible via the linear mapping -- please check your bootloader ...\n")) {
383 phys_initrd_size = 0;
385 memblock_remove(base, size); /* clear MEMBLOCK_ flags */
386 memblock_add(base, size);
387 memblock_reserve(base, size);
391 if (IS_ENABLED(CONFIG_RANDOMIZE_BASE)) {
392 extern u16 memstart_offset_seed;
393 u64 range = linear_region_size -
394 (memblock_end_of_DRAM() - memblock_start_of_DRAM());
397 * If the size of the linear region exceeds, by a sufficient
398 * margin, the size of the region that the available physical
399 * memory spans, randomize the linear region as well.
401 if (memstart_offset_seed > 0 && range >= ARM64_MEMSTART_ALIGN) {
402 range /= ARM64_MEMSTART_ALIGN;
403 memstart_addr -= ARM64_MEMSTART_ALIGN *
404 ((range * memstart_offset_seed) >> 16);
409 * Register the kernel text, kernel data, initrd, and initial
410 * pagetables with memblock.
412 memblock_reserve(__pa_symbol(_text), _end - _text);
413 if (IS_ENABLED(CONFIG_BLK_DEV_INITRD) && phys_initrd_size) {
414 /* the generic initrd code expects virtual addresses */
415 initrd_start = __phys_to_virt(phys_initrd_start);
416 initrd_end = initrd_start + phys_initrd_size;
419 early_init_fdt_scan_reserved_mem();
421 /* 4GB maximum for 32-bit only capable devices */
422 if (IS_ENABLED(CONFIG_ZONE_DMA32))
423 arm64_dma_phys_limit = max_zone_dma_phys();
425 arm64_dma_phys_limit = PHYS_MASK + 1;
427 reserve_crashkernel();
429 reserve_elfcorehdr();
431 high_memory = __va(memblock_end_of_DRAM() - 1) + 1;
433 dma_contiguous_reserve(arm64_dma_phys_limit);
436 void __init bootmem_init(void)
438 unsigned long min, max;
440 min = PFN_UP(memblock_start_of_DRAM());
441 max = PFN_DOWN(memblock_end_of_DRAM());
443 early_memtest(min << PAGE_SHIFT, max << PAGE_SHIFT);
445 max_pfn = max_low_pfn = max;
450 * Sparsemem tries to allocate bootmem in memory_present(), so must be
451 * done after the fixed reservations.
456 zone_sizes_init(min, max);
461 #ifndef CONFIG_SPARSEMEM_VMEMMAP
462 static inline void free_memmap(unsigned long start_pfn, unsigned long end_pfn)
464 struct page *start_pg, *end_pg;
465 unsigned long pg, pgend;
468 * Convert start_pfn/end_pfn to a struct page pointer.
470 start_pg = pfn_to_page(start_pfn - 1) + 1;
471 end_pg = pfn_to_page(end_pfn - 1) + 1;
474 * Convert to physical addresses, and round start upwards and end
477 pg = (unsigned long)PAGE_ALIGN(__pa(start_pg));
478 pgend = (unsigned long)__pa(end_pg) & PAGE_MASK;
481 * If there are free pages between these, free the section of the
485 memblock_free(pg, pgend - pg);
489 * The mem_map array can get very big. Free the unused area of the memory map.
491 static void __init free_unused_memmap(void)
493 unsigned long start, prev_end = 0;
494 struct memblock_region *reg;
496 for_each_memblock(memory, reg) {
497 start = __phys_to_pfn(reg->base);
499 #ifdef CONFIG_SPARSEMEM
501 * Take care not to free memmap entries that don't exist due
502 * to SPARSEMEM sections which aren't present.
504 start = min(start, ALIGN(prev_end, PAGES_PER_SECTION));
507 * If we had a previous bank, and there is a space between the
508 * current bank and the previous, free it.
510 if (prev_end && prev_end < start)
511 free_memmap(prev_end, start);
514 * Align up here since the VM subsystem insists that the
515 * memmap entries are valid from the bank end aligned to
516 * MAX_ORDER_NR_PAGES.
518 prev_end = ALIGN(__phys_to_pfn(reg->base + reg->size),
522 #ifdef CONFIG_SPARSEMEM
523 if (!IS_ALIGNED(prev_end, PAGES_PER_SECTION))
524 free_memmap(prev_end, ALIGN(prev_end, PAGES_PER_SECTION));
527 #endif /* !CONFIG_SPARSEMEM_VMEMMAP */
530 * mem_init() marks the free areas in the mem_map and tells us how much memory
531 * is free. This is done after various parts of the system have claimed their
532 * memory after the kernel image.
534 void __init mem_init(void)
536 if (swiotlb_force == SWIOTLB_FORCE ||
537 max_pfn > (arm64_dma_phys_limit >> PAGE_SHIFT))
540 swiotlb_force = SWIOTLB_NO_FORCE;
542 set_max_mapnr(max_pfn - PHYS_PFN_OFFSET);
544 #ifndef CONFIG_SPARSEMEM_VMEMMAP
545 free_unused_memmap();
547 /* this will put all unused low memory onto the freelists */
550 mem_init_print_info(NULL);
553 * Check boundaries twice: Some fundamental inconsistencies can be
554 * detected at build time already.
557 BUILD_BUG_ON(TASK_SIZE_32 > DEFAULT_MAP_WINDOW_64);
560 if (PAGE_SIZE >= 16384 && get_num_physpages() <= 128) {
561 extern int sysctl_overcommit_memory;
563 * On a machine this small we won't get anywhere without
564 * overcommit, so turn it on by default.
566 sysctl_overcommit_memory = OVERCOMMIT_ALWAYS;
570 void free_initmem(void)
572 free_reserved_area(lm_alias(__init_begin),
573 lm_alias(__init_end),
576 * Unmap the __init region but leave the VM area in place. This
577 * prevents the region from being reused for kernel modules, which
578 * is not supported by kallsyms.
580 unmap_kernel_range((u64)__init_begin, (u64)(__init_end - __init_begin));
583 #ifdef CONFIG_BLK_DEV_INITRD
584 void __init free_initrd_mem(unsigned long start, unsigned long end)
586 unsigned long aligned_start, aligned_end;
588 aligned_start = __virt_to_phys(start) & PAGE_MASK;
589 aligned_end = PAGE_ALIGN(__virt_to_phys(end));
590 memblock_free(aligned_start, aligned_end - aligned_start);
591 free_reserved_area((void *)start, (void *)end, 0, "initrd");
596 * Dump out memory limit information on panic.
598 static int dump_mem_limit(struct notifier_block *self, unsigned long v, void *p)
600 if (memory_limit != PHYS_ADDR_MAX) {
601 pr_emerg("Memory Limit: %llu MB\n", memory_limit >> 20);
603 pr_emerg("Memory Limit: none\n");
608 static struct notifier_block mem_limit_notifier = {
609 .notifier_call = dump_mem_limit,
612 static int __init register_mem_limit_dumper(void)
614 atomic_notifier_chain_register(&panic_notifier_list,
615 &mem_limit_notifier);
618 __initcall(register_mem_limit_dumper);