1 | // SPDX-License-Identifier: GPL-2.0-only |
2 | /* |
3 | * Based on arch/arm/mm/init.c |
4 | * |
5 | * Copyright (C) 1995-2005 Russell King |
6 | * Copyright (C) 2012 ARM Ltd. |
7 | */ |
8 | |
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/math.h> |
20 | #include <linux/memblock.h> |
21 | #include <linux/sort.h> |
22 | #include <linux/of.h> |
23 | #include <linux/of_fdt.h> |
24 | #include <linux/dma-direct.h> |
25 | #include <linux/dma-map-ops.h> |
26 | #include <linux/efi.h> |
27 | #include <linux/swiotlb.h> |
28 | #include <linux/vmalloc.h> |
29 | #include <linux/mm.h> |
30 | #include <linux/kexec.h> |
31 | #include <linux/crash_dump.h> |
32 | #include <linux/hugetlb.h> |
33 | #include <linux/acpi_iort.h> |
34 | #include <linux/kmemleak.h> |
35 | |
36 | #include <asm/boot.h> |
37 | #include <asm/fixmap.h> |
38 | #include <asm/kasan.h> |
39 | #include <asm/kernel-pgtable.h> |
40 | #include <asm/kvm_host.h> |
41 | #include <asm/memory.h> |
42 | #include <asm/numa.h> |
43 | #include <asm/sections.h> |
44 | #include <asm/setup.h> |
45 | #include <linux/sizes.h> |
46 | #include <asm/tlb.h> |
47 | #include <asm/alternative.h> |
48 | #include <asm/xen/swiotlb-xen.h> |
49 | |
50 | /* |
51 | * We need to be able to catch inadvertent references to memstart_addr |
52 | * that occur (potentially in generic code) before arm64_memblock_init() |
53 | * executes, which assigns it its actual value. So use a default value |
54 | * that cannot be mistaken for a real physical address. |
55 | */ |
56 | s64 memstart_addr __ro_after_init = -1; |
57 | EXPORT_SYMBOL(memstart_addr); |
58 | |
59 | /* |
60 | * If the corresponding config options are enabled, we create both ZONE_DMA |
61 | * and ZONE_DMA32. By default ZONE_DMA covers the 32-bit addressable memory |
62 | * unless restricted on specific platforms (e.g. 30-bit on Raspberry Pi 4). |
63 | * In such case, ZONE_DMA32 covers the rest of the 32-bit addressable memory, |
64 | * otherwise it is empty. |
65 | */ |
66 | phys_addr_t __ro_after_init arm64_dma_phys_limit; |
67 | |
68 | /* |
69 | * To make optimal use of block mappings when laying out the linear |
70 | * mapping, round down the base of physical memory to a size that can |
71 | * be mapped efficiently, i.e., either PUD_SIZE (4k granule) or PMD_SIZE |
72 | * (64k granule), or a multiple that can be mapped using contiguous bits |
73 | * in the page tables: 32 * PMD_SIZE (16k granule) |
74 | */ |
75 | #if defined(CONFIG_ARM64_4K_PAGES) |
76 | #define ARM64_MEMSTART_SHIFT PUD_SHIFT |
77 | #elif defined(CONFIG_ARM64_16K_PAGES) |
78 | #define ARM64_MEMSTART_SHIFT CONT_PMD_SHIFT |
79 | #else |
80 | #define ARM64_MEMSTART_SHIFT PMD_SHIFT |
81 | #endif |
82 | |
83 | /* |
84 | * sparsemem vmemmap imposes an additional requirement on the alignment of |
85 | * memstart_addr, due to the fact that the base of the vmemmap region |
86 | * has a direct correspondence, and needs to appear sufficiently aligned |
87 | * in the virtual address space. |
88 | */ |
89 | #if ARM64_MEMSTART_SHIFT < SECTION_SIZE_BITS |
90 | #define ARM64_MEMSTART_ALIGN (1UL << SECTION_SIZE_BITS) |
91 | #else |
92 | #define ARM64_MEMSTART_ALIGN (1UL << ARM64_MEMSTART_SHIFT) |
93 | #endif |
94 | |
95 | static void __init arch_reserve_crashkernel(void) |
96 | { |
97 | unsigned long long low_size = 0; |
98 | unsigned long long crash_base, crash_size; |
99 | char *cmdline = boot_command_line; |
100 | bool high = false; |
101 | int ret; |
102 | |
103 | if (!IS_ENABLED(CONFIG_KEXEC_CORE)) |
104 | return; |
105 | |
106 | ret = parse_crashkernel(cmdline, system_ram: memblock_phys_mem_size(), |
107 | crash_size: &crash_size, crash_base: &crash_base, |
108 | low_size: &low_size, high: &high); |
109 | if (ret) |
110 | return; |
111 | |
112 | reserve_crashkernel_generic(cmdline, crash_size, crash_base, |
113 | crash_low_size: low_size, high); |
114 | } |
115 | |
116 | /* |
117 | * Return the maximum physical address for a zone accessible by the given bits |
118 | * limit. If DRAM starts above 32-bit, expand the zone to the maximum |
119 | * available memory, otherwise cap it at 32-bit. |
120 | */ |
121 | static phys_addr_t __init max_zone_phys(unsigned int zone_bits) |
122 | { |
123 | phys_addr_t zone_mask = DMA_BIT_MASK(zone_bits); |
124 | phys_addr_t phys_start = memblock_start_of_DRAM(); |
125 | |
126 | if (phys_start > U32_MAX) |
127 | zone_mask = PHYS_ADDR_MAX; |
128 | else if (phys_start > zone_mask) |
129 | zone_mask = U32_MAX; |
130 | |
131 | return min(zone_mask, memblock_end_of_DRAM() - 1) + 1; |
132 | } |
133 | |
134 | static void __init zone_sizes_init(void) |
135 | { |
136 | unsigned long max_zone_pfns[MAX_NR_ZONES] = {0}; |
137 | unsigned int __maybe_unused acpi_zone_dma_bits; |
138 | unsigned int __maybe_unused dt_zone_dma_bits; |
139 | phys_addr_t __maybe_unused dma32_phys_limit = max_zone_phys(zone_bits: 32); |
140 | |
141 | #ifdef CONFIG_ZONE_DMA |
142 | acpi_zone_dma_bits = fls64(x: acpi_iort_dma_get_max_cpu_address()); |
143 | dt_zone_dma_bits = fls64(x: of_dma_get_max_cpu_address(NULL)); |
144 | zone_dma_bits = min3(32U, dt_zone_dma_bits, acpi_zone_dma_bits); |
145 | arm64_dma_phys_limit = max_zone_phys(zone_bits: zone_dma_bits); |
146 | max_zone_pfns[ZONE_DMA] = PFN_DOWN(arm64_dma_phys_limit); |
147 | #endif |
148 | #ifdef CONFIG_ZONE_DMA32 |
149 | max_zone_pfns[ZONE_DMA32] = PFN_DOWN(dma32_phys_limit); |
150 | if (!arm64_dma_phys_limit) |
151 | arm64_dma_phys_limit = dma32_phys_limit; |
152 | #endif |
153 | if (!arm64_dma_phys_limit) |
154 | arm64_dma_phys_limit = PHYS_MASK + 1; |
155 | max_zone_pfns[ZONE_NORMAL] = max_pfn; |
156 | |
157 | free_area_init(max_zone_pfn: max_zone_pfns); |
158 | } |
159 | |
160 | int pfn_is_map_memory(unsigned long pfn) |
161 | { |
162 | phys_addr_t addr = PFN_PHYS(pfn); |
163 | |
164 | /* avoid false positives for bogus PFNs, see comment in pfn_valid() */ |
165 | if (PHYS_PFN(addr) != pfn) |
166 | return 0; |
167 | |
168 | return memblock_is_map_memory(addr); |
169 | } |
170 | EXPORT_SYMBOL(pfn_is_map_memory); |
171 | |
172 | static phys_addr_t memory_limit __ro_after_init = PHYS_ADDR_MAX; |
173 | |
174 | /* |
175 | * Limit the memory size that was specified via FDT. |
176 | */ |
177 | static int __init early_mem(char *p) |
178 | { |
179 | if (!p) |
180 | return 1; |
181 | |
182 | memory_limit = memparse(ptr: p, retptr: &p) & PAGE_MASK; |
183 | pr_notice("Memory limited to %lldMB\n" , memory_limit >> 20); |
184 | |
185 | return 0; |
186 | } |
187 | early_param("mem" , early_mem); |
188 | |
189 | void __init arm64_memblock_init(void) |
190 | { |
191 | s64 linear_region_size = PAGE_END - _PAGE_OFFSET(vabits_actual); |
192 | |
193 | /* |
194 | * Corner case: 52-bit VA capable systems running KVM in nVHE mode may |
195 | * be limited in their ability to support a linear map that exceeds 51 |
196 | * bits of VA space, depending on the placement of the ID map. Given |
197 | * that the placement of the ID map may be randomized, let's simply |
198 | * limit the kernel's linear map to 51 bits as well if we detect this |
199 | * configuration. |
200 | */ |
201 | if (IS_ENABLED(CONFIG_KVM) && vabits_actual == 52 && |
202 | is_hyp_mode_available() && !is_kernel_in_hyp_mode()) { |
203 | pr_info("Capping linear region to 51 bits for KVM in nVHE mode on LVA capable hardware.\n" ); |
204 | linear_region_size = min_t(u64, linear_region_size, BIT(51)); |
205 | } |
206 | |
207 | /* Remove memory above our supported physical address size */ |
208 | memblock_remove(base: 1ULL << PHYS_MASK_SHIFT, ULLONG_MAX); |
209 | |
210 | /* |
211 | * Select a suitable value for the base of physical memory. |
212 | */ |
213 | memstart_addr = round_down(memblock_start_of_DRAM(), |
214 | ARM64_MEMSTART_ALIGN); |
215 | |
216 | if ((memblock_end_of_DRAM() - memstart_addr) > linear_region_size) |
217 | pr_warn("Memory doesn't fit in the linear mapping, VA_BITS too small\n" ); |
218 | |
219 | /* |
220 | * Remove the memory that we will not be able to cover with the |
221 | * linear mapping. Take care not to clip the kernel which may be |
222 | * high in memory. |
223 | */ |
224 | memblock_remove(max_t(u64, memstart_addr + linear_region_size, |
225 | __pa_symbol(_end)), ULLONG_MAX); |
226 | if (memstart_addr + linear_region_size < memblock_end_of_DRAM()) { |
227 | /* ensure that memstart_addr remains sufficiently aligned */ |
228 | memstart_addr = round_up(memblock_end_of_DRAM() - linear_region_size, |
229 | ARM64_MEMSTART_ALIGN); |
230 | memblock_remove(base: 0, size: memstart_addr); |
231 | } |
232 | |
233 | /* |
234 | * If we are running with a 52-bit kernel VA config on a system that |
235 | * does not support it, we have to place the available physical |
236 | * memory in the 48-bit addressable part of the linear region, i.e., |
237 | * we have to move it upward. Since memstart_addr represents the |
238 | * physical address of PAGE_OFFSET, we have to *subtract* from it. |
239 | */ |
240 | if (IS_ENABLED(CONFIG_ARM64_VA_BITS_52) && (vabits_actual != 52)) |
241 | memstart_addr -= _PAGE_OFFSET(48) - _PAGE_OFFSET(52); |
242 | |
243 | /* |
244 | * Apply the memory limit if it was set. Since the kernel may be loaded |
245 | * high up in memory, add back the kernel region that must be accessible |
246 | * via the linear mapping. |
247 | */ |
248 | if (memory_limit != PHYS_ADDR_MAX) { |
249 | memblock_mem_limit_remove_map(limit: memory_limit); |
250 | memblock_add(__pa_symbol(_text), size: (u64)(_end - _text)); |
251 | } |
252 | |
253 | if (IS_ENABLED(CONFIG_BLK_DEV_INITRD) && phys_initrd_size) { |
254 | /* |
255 | * Add back the memory we just removed if it results in the |
256 | * initrd to become inaccessible via the linear mapping. |
257 | * Otherwise, this is a no-op |
258 | */ |
259 | u64 base = phys_initrd_start & PAGE_MASK; |
260 | u64 size = PAGE_ALIGN(phys_initrd_start + phys_initrd_size) - base; |
261 | |
262 | /* |
263 | * We can only add back the initrd memory if we don't end up |
264 | * with more memory than we can address via the linear mapping. |
265 | * It is up to the bootloader to position the kernel and the |
266 | * initrd reasonably close to each other (i.e., within 32 GB of |
267 | * each other) so that all granule/#levels combinations can |
268 | * always access both. |
269 | */ |
270 | if (WARN(base < memblock_start_of_DRAM() || |
271 | base + size > memblock_start_of_DRAM() + |
272 | linear_region_size, |
273 | "initrd not fully accessible via the linear mapping -- please check your bootloader ...\n" )) { |
274 | phys_initrd_size = 0; |
275 | } else { |
276 | memblock_add(base, size); |
277 | memblock_clear_nomap(base, size); |
278 | memblock_reserve(base, size); |
279 | } |
280 | } |
281 | |
282 | if (IS_ENABLED(CONFIG_RANDOMIZE_BASE)) { |
283 | extern u16 memstart_offset_seed; |
284 | u64 mmfr0 = read_cpuid(ID_AA64MMFR0_EL1); |
285 | int parange = cpuid_feature_extract_unsigned_field( |
286 | mmfr0, ID_AA64MMFR0_EL1_PARANGE_SHIFT); |
287 | s64 range = linear_region_size - |
288 | BIT(id_aa64mmfr0_parange_to_phys_shift(parange)); |
289 | |
290 | /* |
291 | * If the size of the linear region exceeds, by a sufficient |
292 | * margin, the size of the region that the physical memory can |
293 | * span, randomize the linear region as well. |
294 | */ |
295 | if (memstart_offset_seed > 0 && range >= (s64)ARM64_MEMSTART_ALIGN) { |
296 | range /= ARM64_MEMSTART_ALIGN; |
297 | memstart_addr -= ARM64_MEMSTART_ALIGN * |
298 | ((range * memstart_offset_seed) >> 16); |
299 | } |
300 | } |
301 | |
302 | /* |
303 | * Register the kernel text, kernel data, initrd, and initial |
304 | * pagetables with memblock. |
305 | */ |
306 | memblock_reserve(__pa_symbol(_stext), size: _end - _stext); |
307 | if (IS_ENABLED(CONFIG_BLK_DEV_INITRD) && phys_initrd_size) { |
308 | /* the generic initrd code expects virtual addresses */ |
309 | initrd_start = __phys_to_virt(phys_initrd_start); |
310 | initrd_end = initrd_start + phys_initrd_size; |
311 | } |
312 | |
313 | early_init_fdt_scan_reserved_mem(); |
314 | |
315 | high_memory = __va(memblock_end_of_DRAM() - 1) + 1; |
316 | } |
317 | |
318 | void __init bootmem_init(void) |
319 | { |
320 | unsigned long min, max; |
321 | |
322 | min = PFN_UP(memblock_start_of_DRAM()); |
323 | max = PFN_DOWN(memblock_end_of_DRAM()); |
324 | |
325 | early_memtest(start: min << PAGE_SHIFT, end: max << PAGE_SHIFT); |
326 | |
327 | max_pfn = max_low_pfn = max; |
328 | min_low_pfn = min; |
329 | |
330 | arch_numa_init(); |
331 | |
332 | /* |
333 | * must be done after arch_numa_init() which calls numa_init() to |
334 | * initialize node_online_map that gets used in hugetlb_cma_reserve() |
335 | * while allocating required CMA size across online nodes. |
336 | */ |
337 | #if defined(CONFIG_HUGETLB_PAGE) && defined(CONFIG_CMA) |
338 | arm64_hugetlb_cma_reserve(); |
339 | #endif |
340 | |
341 | kvm_hyp_reserve(); |
342 | |
343 | /* |
344 | * sparse_init() tries to allocate memory from memblock, so must be |
345 | * done after the fixed reservations |
346 | */ |
347 | sparse_init(); |
348 | zone_sizes_init(); |
349 | |
350 | /* |
351 | * Reserve the CMA area after arm64_dma_phys_limit was initialised. |
352 | */ |
353 | dma_contiguous_reserve(addr_limit: arm64_dma_phys_limit); |
354 | |
355 | /* |
356 | * request_standard_resources() depends on crashkernel's memory being |
357 | * reserved, so do it here. |
358 | */ |
359 | arch_reserve_crashkernel(); |
360 | |
361 | memblock_dump_all(); |
362 | } |
363 | |
364 | /* |
365 | * mem_init() marks the free areas in the mem_map and tells us how much memory |
366 | * is free. This is done after various parts of the system have claimed their |
367 | * memory after the kernel image. |
368 | */ |
369 | void __init mem_init(void) |
370 | { |
371 | bool swiotlb = max_pfn > PFN_DOWN(arm64_dma_phys_limit); |
372 | |
373 | if (IS_ENABLED(CONFIG_DMA_BOUNCE_UNALIGNED_KMALLOC) && !swiotlb) { |
374 | /* |
375 | * If no bouncing needed for ZONE_DMA, reduce the swiotlb |
376 | * buffer for kmalloc() bouncing to 1MB per 1GB of RAM. |
377 | */ |
378 | unsigned long size = |
379 | DIV_ROUND_UP(memblock_phys_mem_size(), 1024); |
380 | swiotlb_adjust_size(min(swiotlb_size_or_default(), size)); |
381 | swiotlb = true; |
382 | } |
383 | |
384 | swiotlb_init(addressing_limited: swiotlb, SWIOTLB_VERBOSE); |
385 | |
386 | /* this will put all unused low memory onto the freelists */ |
387 | memblock_free_all(); |
388 | |
389 | /* |
390 | * Check boundaries twice: Some fundamental inconsistencies can be |
391 | * detected at build time already. |
392 | */ |
393 | #ifdef CONFIG_COMPAT |
394 | BUILD_BUG_ON(TASK_SIZE_32 > DEFAULT_MAP_WINDOW_64); |
395 | #endif |
396 | |
397 | /* |
398 | * Selected page table levels should match when derived from |
399 | * scratch using the virtual address range and page size. |
400 | */ |
401 | BUILD_BUG_ON(ARM64_HW_PGTABLE_LEVELS(CONFIG_ARM64_VA_BITS) != |
402 | CONFIG_PGTABLE_LEVELS); |
403 | |
404 | if (PAGE_SIZE >= 16384 && get_num_physpages() <= 128) { |
405 | extern int sysctl_overcommit_memory; |
406 | /* |
407 | * On a machine this small we won't get anywhere without |
408 | * overcommit, so turn it on by default. |
409 | */ |
410 | sysctl_overcommit_memory = OVERCOMMIT_ALWAYS; |
411 | } |
412 | } |
413 | |
414 | void free_initmem(void) |
415 | { |
416 | free_reserved_area(lm_alias(__init_begin), |
417 | lm_alias(__init_end), |
418 | POISON_FREE_INITMEM, s: "unused kernel" ); |
419 | /* |
420 | * Unmap the __init region but leave the VM area in place. This |
421 | * prevents the region from being reused for kernel modules, which |
422 | * is not supported by kallsyms. |
423 | */ |
424 | vunmap_range(addr: (u64)__init_begin, end: (u64)__init_end); |
425 | } |
426 | |
427 | void dump_mem_limit(void) |
428 | { |
429 | if (memory_limit != PHYS_ADDR_MAX) { |
430 | pr_emerg("Memory Limit: %llu MB\n" , memory_limit >> 20); |
431 | } else { |
432 | pr_emerg("Memory Limit: none\n" ); |
433 | } |
434 | } |
435 | |