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-map-ops.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>
31 #include <linux/hugetlb.h>
32 #include <linux/acpi_iort.h>
35 #include <asm/fixmap.h>
36 #include <asm/kasan.h>
37 #include <asm/kernel-pgtable.h>
38 #include <asm/kvm_host.h>
39 #include <asm/memory.h>
41 #include <asm/sections.h>
42 #include <asm/setup.h>
43 #include <linux/sizes.h>
45 #include <asm/alternative.h>
48 * We need to be able to catch inadvertent references to memstart_addr
49 * that occur (potentially in generic code) before arm64_memblock_init()
50 * executes, which assigns it its actual value. So use a default value
51 * that cannot be mistaken for a real physical address.
53 s64 memstart_addr __ro_after_init = -1;
54 EXPORT_SYMBOL(memstart_addr);
57 * If the corresponding config options are enabled, we create both ZONE_DMA
58 * and ZONE_DMA32. By default ZONE_DMA covers the 32-bit addressable memory
59 * unless restricted on specific platforms (e.g. 30-bit on Raspberry Pi 4).
60 * In such case, ZONE_DMA32 covers the rest of the 32-bit addressable memory,
61 * otherwise it is empty.
63 phys_addr_t arm64_dma_phys_limit __ro_after_init;
65 #ifdef CONFIG_KEXEC_CORE
67 * reserve_crashkernel() - reserves memory for crash kernel
69 * This function reserves memory area given in "crashkernel=" kernel command
70 * line parameter. The memory reserved is used by dump capture kernel when
71 * primary kernel is crashing.
73 static void __init reserve_crashkernel(void)
75 unsigned long long crash_base, crash_size;
78 ret = parse_crashkernel(boot_command_line, memblock_phys_mem_size(),
79 &crash_size, &crash_base);
80 /* no crashkernel= or invalid value specified */
81 if (ret || !crash_size)
84 crash_size = PAGE_ALIGN(crash_size);
86 if (crash_base == 0) {
87 /* Current arm64 boot protocol requires 2MB alignment */
88 crash_base = memblock_find_in_range(0, arm64_dma_phys_limit,
90 if (crash_base == 0) {
91 pr_warn("cannot allocate crashkernel (size:0x%llx)\n",
96 /* User specifies base address explicitly. */
97 if (!memblock_is_region_memory(crash_base, crash_size)) {
98 pr_warn("cannot reserve crashkernel: region is not memory\n");
102 if (memblock_is_region_reserved(crash_base, crash_size)) {
103 pr_warn("cannot reserve crashkernel: region overlaps reserved memory\n");
107 if (!IS_ALIGNED(crash_base, SZ_2M)) {
108 pr_warn("cannot reserve crashkernel: base address is not 2MB aligned\n");
112 memblock_reserve(crash_base, crash_size);
114 pr_info("crashkernel reserved: 0x%016llx - 0x%016llx (%lld MB)\n",
115 crash_base, crash_base + crash_size, crash_size >> 20);
117 crashk_res.start = crash_base;
118 crashk_res.end = crash_base + crash_size - 1;
121 static void __init reserve_crashkernel(void)
124 #endif /* CONFIG_KEXEC_CORE */
126 #ifdef CONFIG_CRASH_DUMP
127 static int __init early_init_dt_scan_elfcorehdr(unsigned long node,
128 const char *uname, int depth, void *data)
133 if (depth != 1 || strcmp(uname, "chosen") != 0)
136 reg = of_get_flat_dt_prop(node, "linux,elfcorehdr", &len);
137 if (!reg || (len < (dt_root_addr_cells + dt_root_size_cells)))
140 elfcorehdr_addr = dt_mem_next_cell(dt_root_addr_cells, ®);
141 elfcorehdr_size = dt_mem_next_cell(dt_root_size_cells, ®);
147 * reserve_elfcorehdr() - reserves memory for elf core header
149 * This function reserves the memory occupied by an elf core header
150 * described in the device tree. This region contains all the
151 * information about primary kernel's core image and is used by a dump
152 * capture kernel to access the system memory on primary kernel.
154 static void __init reserve_elfcorehdr(void)
156 of_scan_flat_dt(early_init_dt_scan_elfcorehdr, NULL);
158 if (!elfcorehdr_size)
161 if (memblock_is_region_reserved(elfcorehdr_addr, elfcorehdr_size)) {
162 pr_warn("elfcorehdr is overlapped\n");
166 memblock_reserve(elfcorehdr_addr, elfcorehdr_size);
168 pr_info("Reserving %lldKB of memory at 0x%llx for elfcorehdr\n",
169 elfcorehdr_size >> 10, elfcorehdr_addr);
172 static void __init reserve_elfcorehdr(void)
175 #endif /* CONFIG_CRASH_DUMP */
178 * Return the maximum physical address for a zone accessible by the given bits
179 * limit. If DRAM starts above 32-bit, expand the zone to the maximum
180 * available memory, otherwise cap it at 32-bit.
182 static phys_addr_t __init max_zone_phys(unsigned int zone_bits)
184 phys_addr_t zone_mask = DMA_BIT_MASK(zone_bits);
185 phys_addr_t phys_start = memblock_start_of_DRAM();
187 if (phys_start > U32_MAX)
188 zone_mask = PHYS_ADDR_MAX;
189 else if (phys_start > zone_mask)
192 return min(zone_mask, memblock_end_of_DRAM() - 1) + 1;
195 static void __init zone_sizes_init(unsigned long min, unsigned long max)
197 unsigned long max_zone_pfns[MAX_NR_ZONES] = {0};
198 unsigned int __maybe_unused acpi_zone_dma_bits;
199 unsigned int __maybe_unused dt_zone_dma_bits;
200 phys_addr_t __maybe_unused dma32_phys_limit = max_zone_phys(32);
202 #ifdef CONFIG_ZONE_DMA
203 acpi_zone_dma_bits = fls64(acpi_iort_dma_get_max_cpu_address());
204 dt_zone_dma_bits = fls64(of_dma_get_max_cpu_address(NULL));
205 zone_dma_bits = min3(32U, dt_zone_dma_bits, acpi_zone_dma_bits);
206 arm64_dma_phys_limit = max_zone_phys(zone_dma_bits);
207 max_zone_pfns[ZONE_DMA] = PFN_DOWN(arm64_dma_phys_limit);
209 #ifdef CONFIG_ZONE_DMA32
210 max_zone_pfns[ZONE_DMA32] = PFN_DOWN(dma32_phys_limit);
211 if (!arm64_dma_phys_limit)
212 arm64_dma_phys_limit = dma32_phys_limit;
214 if (!arm64_dma_phys_limit)
215 arm64_dma_phys_limit = PHYS_MASK + 1;
216 max_zone_pfns[ZONE_NORMAL] = max;
218 free_area_init(max_zone_pfns);
221 int pfn_valid(unsigned long pfn)
223 phys_addr_t addr = PFN_PHYS(pfn);
224 struct mem_section *ms;
227 * Ensure the upper PAGE_SHIFT bits are clear in the
228 * pfn. Else it might lead to false positives when
229 * some of the upper bits are set, but the lower bits
232 if (PHYS_PFN(addr) != pfn)
235 if (pfn_to_section_nr(pfn) >= NR_MEM_SECTIONS)
238 ms = __pfn_to_section(pfn);
239 if (!valid_section(ms))
243 * ZONE_DEVICE memory does not have the memblock entries.
244 * memblock_is_map_memory() check for ZONE_DEVICE based
245 * addresses will always fail. Even the normal hotplugged
246 * memory will never have MEMBLOCK_NOMAP flag set in their
247 * memblock entries. Skip memblock search for all non early
248 * memory sections covering all of hotplug memory including
249 * both normal and ZONE_DEVICE based.
251 if (!early_section(ms))
252 return pfn_section_valid(ms, pfn);
254 return memblock_is_map_memory(addr);
256 EXPORT_SYMBOL(pfn_valid);
258 static phys_addr_t memory_limit = PHYS_ADDR_MAX;
261 * Limit the memory size that was specified via FDT.
263 static int __init early_mem(char *p)
268 memory_limit = memparse(p, &p) & PAGE_MASK;
269 pr_notice("Memory limited to %lldMB\n", memory_limit >> 20);
273 early_param("mem", early_mem);
275 static int __init early_init_dt_scan_usablemem(unsigned long node,
276 const char *uname, int depth, void *data)
278 struct memblock_region *usablemem = data;
282 if (depth != 1 || strcmp(uname, "chosen") != 0)
285 reg = of_get_flat_dt_prop(node, "linux,usable-memory-range", &len);
286 if (!reg || (len < (dt_root_addr_cells + dt_root_size_cells)))
289 usablemem->base = dt_mem_next_cell(dt_root_addr_cells, ®);
290 usablemem->size = dt_mem_next_cell(dt_root_size_cells, ®);
295 static void __init fdt_enforce_memory_region(void)
297 struct memblock_region reg = {
301 of_scan_flat_dt(early_init_dt_scan_usablemem, ®);
304 memblock_cap_memory_range(reg.base, reg.size);
307 void __init arm64_memblock_init(void)
309 const s64 linear_region_size = PAGE_END - _PAGE_OFFSET(vabits_actual);
311 /* Handle linux,usable-memory-range property */
312 fdt_enforce_memory_region();
314 /* Remove memory above our supported physical address size */
315 memblock_remove(1ULL << PHYS_MASK_SHIFT, ULLONG_MAX);
318 * Select a suitable value for the base of physical memory.
320 memstart_addr = round_down(memblock_start_of_DRAM(),
321 ARM64_MEMSTART_ALIGN);
323 if ((memblock_end_of_DRAM() - memstart_addr) > linear_region_size)
324 pr_warn("Memory doesn't fit in the linear mapping, VA_BITS too small\n");
327 * Remove the memory that we will not be able to cover with the
328 * linear mapping. Take care not to clip the kernel which may be
331 memblock_remove(max_t(u64, memstart_addr + linear_region_size,
332 __pa_symbol(_end)), ULLONG_MAX);
333 if (memstart_addr + linear_region_size < memblock_end_of_DRAM()) {
334 /* ensure that memstart_addr remains sufficiently aligned */
335 memstart_addr = round_up(memblock_end_of_DRAM() - linear_region_size,
336 ARM64_MEMSTART_ALIGN);
337 memblock_remove(0, memstart_addr);
341 * If we are running with a 52-bit kernel VA config on a system that
342 * does not support it, we have to place the available physical
343 * memory in the 48-bit addressable part of the linear region, i.e.,
344 * we have to move it upward. Since memstart_addr represents the
345 * physical address of PAGE_OFFSET, we have to *subtract* from it.
347 if (IS_ENABLED(CONFIG_ARM64_VA_BITS_52) && (vabits_actual != 52))
348 memstart_addr -= _PAGE_OFFSET(48) - _PAGE_OFFSET(52);
351 * Apply the memory limit if it was set. Since the kernel may be loaded
352 * high up in memory, add back the kernel region that must be accessible
353 * via the linear mapping.
355 if (memory_limit != PHYS_ADDR_MAX) {
356 memblock_mem_limit_remove_map(memory_limit);
357 memblock_add(__pa_symbol(_text), (u64)(_end - _text));
360 if (IS_ENABLED(CONFIG_BLK_DEV_INITRD) && phys_initrd_size) {
362 * Add back the memory we just removed if it results in the
363 * initrd to become inaccessible via the linear mapping.
364 * Otherwise, this is a no-op
366 u64 base = phys_initrd_start & PAGE_MASK;
367 u64 size = PAGE_ALIGN(phys_initrd_start + phys_initrd_size) - base;
370 * We can only add back the initrd memory if we don't end up
371 * with more memory than we can address via the linear mapping.
372 * It is up to the bootloader to position the kernel and the
373 * initrd reasonably close to each other (i.e., within 32 GB of
374 * each other) so that all granule/#levels combinations can
375 * always access both.
377 if (WARN(base < memblock_start_of_DRAM() ||
378 base + size > memblock_start_of_DRAM() +
380 "initrd not fully accessible via the linear mapping -- please check your bootloader ...\n")) {
381 phys_initrd_size = 0;
383 memblock_remove(base, size); /* clear MEMBLOCK_ flags */
384 memblock_add(base, size);
385 memblock_reserve(base, size);
389 if (IS_ENABLED(CONFIG_RANDOMIZE_BASE)) {
390 extern u16 memstart_offset_seed;
391 u64 mmfr0 = read_cpuid(ID_AA64MMFR0_EL1);
392 int parange = cpuid_feature_extract_unsigned_field(
393 mmfr0, ID_AA64MMFR0_PARANGE_SHIFT);
394 s64 range = linear_region_size -
395 BIT(id_aa64mmfr0_parange_to_phys_shift(parange));
398 * If the size of the linear region exceeds, by a sufficient
399 * margin, the size of the region that the physical memory can
400 * span, randomize the linear region as well.
402 if (memstart_offset_seed > 0 && range >= (s64)ARM64_MEMSTART_ALIGN) {
403 range /= ARM64_MEMSTART_ALIGN;
404 memstart_addr -= ARM64_MEMSTART_ALIGN *
405 ((range * memstart_offset_seed) >> 16);
410 * Register the kernel text, kernel data, initrd, and initial
411 * pagetables with memblock.
413 memblock_reserve(__pa_symbol(_stext), _end - _stext);
414 if (IS_ENABLED(CONFIG_BLK_DEV_INITRD) && phys_initrd_size) {
415 /* the generic initrd code expects virtual addresses */
416 initrd_start = __phys_to_virt(phys_initrd_start);
417 initrd_end = initrd_start + phys_initrd_size;
420 early_init_fdt_scan_reserved_mem();
422 reserve_elfcorehdr();
424 high_memory = __va(memblock_end_of_DRAM() - 1) + 1;
427 void __init bootmem_init(void)
429 unsigned long min, max;
431 min = PFN_UP(memblock_start_of_DRAM());
432 max = PFN_DOWN(memblock_end_of_DRAM());
434 early_memtest(min << PAGE_SHIFT, max << PAGE_SHIFT);
436 max_pfn = max_low_pfn = max;
442 * must be done after arch_numa_init() which calls numa_init() to
443 * initialize node_online_map that gets used in hugetlb_cma_reserve()
444 * while allocating required CMA size across online nodes.
446 #if defined(CONFIG_HUGETLB_PAGE) && defined(CONFIG_CMA)
447 arm64_hugetlb_cma_reserve();
450 dma_pernuma_cma_reserve();
455 * sparse_init() tries to allocate memory from memblock, so must be
456 * done after the fixed reservations
459 zone_sizes_init(min, max);
462 * Reserve the CMA area after arm64_dma_phys_limit was initialised.
464 dma_contiguous_reserve(arm64_dma_phys_limit);
467 * request_standard_resources() depends on crashkernel's memory being
468 * reserved, so do it here.
470 reserve_crashkernel();
476 * mem_init() marks the free areas in the mem_map and tells us how much memory
477 * is free. This is done after various parts of the system have claimed their
478 * memory after the kernel image.
480 void __init mem_init(void)
482 if (swiotlb_force == SWIOTLB_FORCE ||
483 max_pfn > PFN_DOWN(arm64_dma_phys_limit))
486 swiotlb_force = SWIOTLB_NO_FORCE;
488 set_max_mapnr(max_pfn - PHYS_PFN_OFFSET);
490 /* this will put all unused low memory onto the freelists */
494 * Check boundaries twice: Some fundamental inconsistencies can be
495 * detected at build time already.
498 BUILD_BUG_ON(TASK_SIZE_32 > DEFAULT_MAP_WINDOW_64);
501 if (PAGE_SIZE >= 16384 && get_num_physpages() <= 128) {
502 extern int sysctl_overcommit_memory;
504 * On a machine this small we won't get anywhere without
505 * overcommit, so turn it on by default.
507 sysctl_overcommit_memory = OVERCOMMIT_ALWAYS;
511 void free_initmem(void)
513 free_reserved_area(lm_alias(__init_begin),
514 lm_alias(__init_end),
515 POISON_FREE_INITMEM, "unused kernel");
517 * Unmap the __init region but leave the VM area in place. This
518 * prevents the region from being reused for kernel modules, which
519 * is not supported by kallsyms.
521 vunmap_range((u64)__init_begin, (u64)__init_end);
524 void dump_mem_limit(void)
526 if (memory_limit != PHYS_ADDR_MAX) {
527 pr_emerg("Memory Limit: %llu MB\n", memory_limit >> 20);
529 pr_emerg("Memory Limit: none\n");