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>
33 #include <linux/kmemleak.h>
36 #include <asm/fixmap.h>
37 #include <asm/kasan.h>
38 #include <asm/kernel-pgtable.h>
39 #include <asm/kvm_host.h>
40 #include <asm/memory.h>
42 #include <asm/sections.h>
43 #include <asm/setup.h>
44 #include <linux/sizes.h>
46 #include <asm/alternative.h>
47 #include <asm/xen/swiotlb-xen.h>
50 * We need to be able to catch inadvertent references to memstart_addr
51 * that occur (potentially in generic code) before arm64_memblock_init()
52 * executes, which assigns it its actual value. So use a default value
53 * that cannot be mistaken for a real physical address.
55 s64 memstart_addr __ro_after_init = -1;
56 EXPORT_SYMBOL(memstart_addr);
59 * If the corresponding config options are enabled, we create both ZONE_DMA
60 * and ZONE_DMA32. By default ZONE_DMA covers the 32-bit addressable memory
61 * unless restricted on specific platforms (e.g. 30-bit on Raspberry Pi 4).
62 * In such case, ZONE_DMA32 covers the rest of the 32-bit addressable memory,
63 * otherwise it is empty.
65 phys_addr_t arm64_dma_phys_limit __ro_after_init;
67 #ifdef CONFIG_KEXEC_CORE
69 * reserve_crashkernel() - reserves memory for crash kernel
71 * This function reserves memory area given in "crashkernel=" kernel command
72 * line parameter. The memory reserved is used by dump capture kernel when
73 * primary kernel is crashing.
75 static void __init reserve_crashkernel(void)
77 unsigned long long crash_base, crash_size;
78 unsigned long long crash_max = arm64_dma_phys_limit;
81 ret = parse_crashkernel(boot_command_line, memblock_phys_mem_size(),
82 &crash_size, &crash_base);
83 /* no crashkernel= or invalid value specified */
84 if (ret || !crash_size)
87 crash_size = PAGE_ALIGN(crash_size);
89 /* User specifies base address explicitly. */
91 crash_max = crash_base + crash_size;
93 /* Current arm64 boot protocol requires 2MB alignment */
94 crash_base = memblock_phys_alloc_range(crash_size, SZ_2M,
95 crash_base, crash_max);
97 pr_warn("cannot allocate crashkernel (size:0x%llx)\n",
102 pr_info("crashkernel reserved: 0x%016llx - 0x%016llx (%lld MB)\n",
103 crash_base, crash_base + crash_size, crash_size >> 20);
106 * The crashkernel memory will be removed from the kernel linear
107 * map. Inform kmemleak so that it won't try to access it.
109 kmemleak_ignore_phys(crash_base);
110 crashk_res.start = crash_base;
111 crashk_res.end = crash_base + crash_size - 1;
114 static void __init reserve_crashkernel(void)
117 #endif /* CONFIG_KEXEC_CORE */
120 * Return the maximum physical address for a zone accessible by the given bits
121 * limit. If DRAM starts above 32-bit, expand the zone to the maximum
122 * available memory, otherwise cap it at 32-bit.
124 static phys_addr_t __init max_zone_phys(unsigned int zone_bits)
126 phys_addr_t zone_mask = DMA_BIT_MASK(zone_bits);
127 phys_addr_t phys_start = memblock_start_of_DRAM();
129 if (phys_start > U32_MAX)
130 zone_mask = PHYS_ADDR_MAX;
131 else if (phys_start > zone_mask)
134 return min(zone_mask, memblock_end_of_DRAM() - 1) + 1;
137 static void __init zone_sizes_init(unsigned long min, unsigned long max)
139 unsigned long max_zone_pfns[MAX_NR_ZONES] = {0};
140 unsigned int __maybe_unused acpi_zone_dma_bits;
141 unsigned int __maybe_unused dt_zone_dma_bits;
142 phys_addr_t __maybe_unused dma32_phys_limit = max_zone_phys(32);
144 #ifdef CONFIG_ZONE_DMA
145 acpi_zone_dma_bits = fls64(acpi_iort_dma_get_max_cpu_address());
146 dt_zone_dma_bits = fls64(of_dma_get_max_cpu_address(NULL));
147 zone_dma_bits = min3(32U, dt_zone_dma_bits, acpi_zone_dma_bits);
148 arm64_dma_phys_limit = max_zone_phys(zone_dma_bits);
149 max_zone_pfns[ZONE_DMA] = PFN_DOWN(arm64_dma_phys_limit);
151 #ifdef CONFIG_ZONE_DMA32
152 max_zone_pfns[ZONE_DMA32] = PFN_DOWN(dma32_phys_limit);
153 if (!arm64_dma_phys_limit)
154 arm64_dma_phys_limit = dma32_phys_limit;
156 if (!arm64_dma_phys_limit)
157 arm64_dma_phys_limit = PHYS_MASK + 1;
158 max_zone_pfns[ZONE_NORMAL] = max;
160 free_area_init(max_zone_pfns);
163 int pfn_valid(unsigned long pfn)
165 phys_addr_t addr = PFN_PHYS(pfn);
166 struct mem_section *ms;
169 * Ensure the upper PAGE_SHIFT bits are clear in the
170 * pfn. Else it might lead to false positives when
171 * some of the upper bits are set, but the lower bits
174 if (PHYS_PFN(addr) != pfn)
177 if (pfn_to_section_nr(pfn) >= NR_MEM_SECTIONS)
180 ms = __pfn_to_section(pfn);
181 if (!valid_section(ms))
185 * ZONE_DEVICE memory does not have the memblock entries.
186 * memblock_is_map_memory() check for ZONE_DEVICE based
187 * addresses will always fail. Even the normal hotplugged
188 * memory will never have MEMBLOCK_NOMAP flag set in their
189 * memblock entries. Skip memblock search for all non early
190 * memory sections covering all of hotplug memory including
191 * both normal and ZONE_DEVICE based.
193 if (!early_section(ms))
194 return pfn_section_valid(ms, pfn);
196 return memblock_is_memory(addr);
198 EXPORT_SYMBOL(pfn_valid);
200 int pfn_is_map_memory(unsigned long pfn)
202 phys_addr_t addr = PFN_PHYS(pfn);
204 /* avoid false positives for bogus PFNs, see comment in pfn_valid() */
205 if (PHYS_PFN(addr) != pfn)
208 return memblock_is_map_memory(addr);
210 EXPORT_SYMBOL(pfn_is_map_memory);
212 static phys_addr_t memory_limit = PHYS_ADDR_MAX;
215 * Limit the memory size that was specified via FDT.
217 static int __init early_mem(char *p)
222 memory_limit = memparse(p, &p) & PAGE_MASK;
223 pr_notice("Memory limited to %lldMB\n", memory_limit >> 20);
227 early_param("mem", early_mem);
229 void __init arm64_memblock_init(void)
231 s64 linear_region_size = PAGE_END - _PAGE_OFFSET(vabits_actual);
234 * Corner case: 52-bit VA capable systems running KVM in nVHE mode may
235 * be limited in their ability to support a linear map that exceeds 51
236 * bits of VA space, depending on the placement of the ID map. Given
237 * that the placement of the ID map may be randomized, let's simply
238 * limit the kernel's linear map to 51 bits as well if we detect this
241 if (IS_ENABLED(CONFIG_KVM) && vabits_actual == 52 &&
242 is_hyp_mode_available() && !is_kernel_in_hyp_mode()) {
243 pr_info("Capping linear region to 51 bits for KVM in nVHE mode on LVA capable hardware.\n");
244 linear_region_size = min_t(u64, linear_region_size, BIT(51));
247 /* Remove memory above our supported physical address size */
248 memblock_remove(1ULL << PHYS_MASK_SHIFT, ULLONG_MAX);
251 * Select a suitable value for the base of physical memory.
253 memstart_addr = round_down(memblock_start_of_DRAM(),
254 ARM64_MEMSTART_ALIGN);
256 if ((memblock_end_of_DRAM() - memstart_addr) > linear_region_size)
257 pr_warn("Memory doesn't fit in the linear mapping, VA_BITS too small\n");
260 * Remove the memory that we will not be able to cover with the
261 * linear mapping. Take care not to clip the kernel which may be
264 memblock_remove(max_t(u64, memstart_addr + linear_region_size,
265 __pa_symbol(_end)), ULLONG_MAX);
266 if (memstart_addr + linear_region_size < memblock_end_of_DRAM()) {
267 /* ensure that memstart_addr remains sufficiently aligned */
268 memstart_addr = round_up(memblock_end_of_DRAM() - linear_region_size,
269 ARM64_MEMSTART_ALIGN);
270 memblock_remove(0, memstart_addr);
274 * If we are running with a 52-bit kernel VA config on a system that
275 * does not support it, we have to place the available physical
276 * memory in the 48-bit addressable part of the linear region, i.e.,
277 * we have to move it upward. Since memstart_addr represents the
278 * physical address of PAGE_OFFSET, we have to *subtract* from it.
280 if (IS_ENABLED(CONFIG_ARM64_VA_BITS_52) && (vabits_actual != 52))
281 memstart_addr -= _PAGE_OFFSET(48) - _PAGE_OFFSET(52);
284 * Apply the memory limit if it was set. Since the kernel may be loaded
285 * high up in memory, add back the kernel region that must be accessible
286 * via the linear mapping.
288 if (memory_limit != PHYS_ADDR_MAX) {
289 memblock_mem_limit_remove_map(memory_limit);
290 memblock_add(__pa_symbol(_text), (u64)(_end - _text));
293 if (IS_ENABLED(CONFIG_BLK_DEV_INITRD) && phys_initrd_size) {
295 * Add back the memory we just removed if it results in the
296 * initrd to become inaccessible via the linear mapping.
297 * Otherwise, this is a no-op
299 u64 base = phys_initrd_start & PAGE_MASK;
300 u64 size = PAGE_ALIGN(phys_initrd_start + phys_initrd_size) - base;
303 * We can only add back the initrd memory if we don't end up
304 * with more memory than we can address via the linear mapping.
305 * It is up to the bootloader to position the kernel and the
306 * initrd reasonably close to each other (i.e., within 32 GB of
307 * each other) so that all granule/#levels combinations can
308 * always access both.
310 if (WARN(base < memblock_start_of_DRAM() ||
311 base + size > memblock_start_of_DRAM() +
313 "initrd not fully accessible via the linear mapping -- please check your bootloader ...\n")) {
314 phys_initrd_size = 0;
316 memblock_remove(base, size); /* clear MEMBLOCK_ flags */
317 memblock_add(base, size);
318 memblock_reserve(base, size);
322 if (IS_ENABLED(CONFIG_RANDOMIZE_BASE)) {
323 extern u16 memstart_offset_seed;
324 u64 mmfr0 = read_cpuid(ID_AA64MMFR0_EL1);
325 int parange = cpuid_feature_extract_unsigned_field(
326 mmfr0, ID_AA64MMFR0_PARANGE_SHIFT);
327 s64 range = linear_region_size -
328 BIT(id_aa64mmfr0_parange_to_phys_shift(parange));
331 * If the size of the linear region exceeds, by a sufficient
332 * margin, the size of the region that the physical memory can
333 * span, randomize the linear region as well.
335 if (memstart_offset_seed > 0 && range >= (s64)ARM64_MEMSTART_ALIGN) {
336 range /= ARM64_MEMSTART_ALIGN;
337 memstart_addr -= ARM64_MEMSTART_ALIGN *
338 ((range * memstart_offset_seed) >> 16);
343 * Register the kernel text, kernel data, initrd, and initial
344 * pagetables with memblock.
346 memblock_reserve(__pa_symbol(_stext), _end - _stext);
347 if (IS_ENABLED(CONFIG_BLK_DEV_INITRD) && phys_initrd_size) {
348 /* the generic initrd code expects virtual addresses */
349 initrd_start = __phys_to_virt(phys_initrd_start);
350 initrd_end = initrd_start + phys_initrd_size;
353 early_init_fdt_scan_reserved_mem();
355 high_memory = __va(memblock_end_of_DRAM() - 1) + 1;
358 void __init bootmem_init(void)
360 unsigned long min, max;
362 min = PFN_UP(memblock_start_of_DRAM());
363 max = PFN_DOWN(memblock_end_of_DRAM());
365 early_memtest(min << PAGE_SHIFT, max << PAGE_SHIFT);
367 max_pfn = max_low_pfn = max;
373 * must be done after arch_numa_init() which calls numa_init() to
374 * initialize node_online_map that gets used in hugetlb_cma_reserve()
375 * while allocating required CMA size across online nodes.
377 #if defined(CONFIG_HUGETLB_PAGE) && defined(CONFIG_CMA)
378 arm64_hugetlb_cma_reserve();
381 dma_pernuma_cma_reserve();
386 * sparse_init() tries to allocate memory from memblock, so must be
387 * done after the fixed reservations
390 zone_sizes_init(min, max);
393 * Reserve the CMA area after arm64_dma_phys_limit was initialised.
395 dma_contiguous_reserve(arm64_dma_phys_limit);
398 * request_standard_resources() depends on crashkernel's memory being
399 * reserved, so do it here.
401 reserve_crashkernel();
407 * mem_init() marks the free areas in the mem_map and tells us how much memory
408 * is free. This is done after various parts of the system have claimed their
409 * memory after the kernel image.
411 void __init mem_init(void)
413 if (swiotlb_force == SWIOTLB_FORCE ||
414 max_pfn > PFN_DOWN(arm64_dma_phys_limit))
416 else if (!xen_swiotlb_detect())
417 swiotlb_force = SWIOTLB_NO_FORCE;
419 set_max_mapnr(max_pfn - PHYS_PFN_OFFSET);
421 /* this will put all unused low memory onto the freelists */
425 * Check boundaries twice: Some fundamental inconsistencies can be
426 * detected at build time already.
429 BUILD_BUG_ON(TASK_SIZE_32 > DEFAULT_MAP_WINDOW_64);
433 * Selected page table levels should match when derived from
434 * scratch using the virtual address range and page size.
436 BUILD_BUG_ON(ARM64_HW_PGTABLE_LEVELS(CONFIG_ARM64_VA_BITS) !=
437 CONFIG_PGTABLE_LEVELS);
439 if (PAGE_SIZE >= 16384 && get_num_physpages() <= 128) {
440 extern int sysctl_overcommit_memory;
442 * On a machine this small we won't get anywhere without
443 * overcommit, so turn it on by default.
445 sysctl_overcommit_memory = OVERCOMMIT_ALWAYS;
449 void free_initmem(void)
451 free_reserved_area(lm_alias(__init_begin),
452 lm_alias(__init_end),
453 POISON_FREE_INITMEM, "unused kernel");
455 * Unmap the __init region but leave the VM area in place. This
456 * prevents the region from being reused for kernel modules, which
457 * is not supported by kallsyms.
459 vunmap_range((u64)__init_begin, (u64)__init_end);
462 void dump_mem_limit(void)
464 if (memory_limit != PHYS_ADDR_MAX) {
465 pr_emerg("Memory Limit: %llu MB\n", memory_limit >> 20);
467 pr_emerg("Memory Limit: none\n");