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 #include <linux/execmem.h>
36
37 #include <asm/boot.h>
38 #include <asm/fixmap.h>
39 #include <asm/kasan.h>
40 #include <asm/kernel-pgtable.h>
41 #include <asm/kvm_host.h>
42 #include <asm/memory.h>
43 #include <asm/numa.h>
44 #include <asm/sections.h>
45 #include <asm/setup.h>
46 #include <linux/sizes.h>
47 #include <asm/tlb.h>
48 #include <asm/alternative.h>
49 #include <asm/xen/swiotlb-xen.h>
50
51 /*
52 * We need to be able to catch inadvertent references to memstart_addr
53 * that occur (potentially in generic code) before arm64_memblock_init()
54 * executes, which assigns it its actual value. So use a default value
55 * that cannot be mistaken for a real physical address.
56 */
57 s64 memstart_addr __ro_after_init = -1;
58 EXPORT_SYMBOL(memstart_addr);
59
60 /*
61 * If the corresponding config options are enabled, we create both ZONE_DMA
62 * and ZONE_DMA32. By default ZONE_DMA covers the 32-bit addressable memory
63 * unless restricted on specific platforms (e.g. 30-bit on Raspberry Pi 4).
64 * In such case, ZONE_DMA32 covers the rest of the 32-bit addressable memory,
65 * otherwise it is empty.
66 */
67 phys_addr_t __ro_after_init arm64_dma_phys_limit;
68
69 /*
70 * To make optimal use of block mappings when laying out the linear
71 * mapping, round down the base of physical memory to a size that can
72 * be mapped efficiently, i.e., either PUD_SIZE (4k granule) or PMD_SIZE
73 * (64k granule), or a multiple that can be mapped using contiguous bits
74 * in the page tables: 32 * PMD_SIZE (16k granule)
75 */
76 #if defined(CONFIG_ARM64_4K_PAGES)
77 #define ARM64_MEMSTART_SHIFT PUD_SHIFT
78 #elif defined(CONFIG_ARM64_16K_PAGES)
79 #define ARM64_MEMSTART_SHIFT CONT_PMD_SHIFT
80 #else
81 #define ARM64_MEMSTART_SHIFT PMD_SHIFT
82 #endif
83
84 /*
85 * sparsemem vmemmap imposes an additional requirement on the alignment of
86 * memstart_addr, due to the fact that the base of the vmemmap region
87 * has a direct correspondence, and needs to appear sufficiently aligned
88 * in the virtual address space.
89 */
90 #if ARM64_MEMSTART_SHIFT < SECTION_SIZE_BITS
91 #define ARM64_MEMSTART_ALIGN (1UL << SECTION_SIZE_BITS)
92 #else
93 #define ARM64_MEMSTART_ALIGN (1UL << ARM64_MEMSTART_SHIFT)
94 #endif
95
arch_reserve_crashkernel(void)96 static void __init arch_reserve_crashkernel(void)
97 {
98 unsigned long long low_size = 0;
99 unsigned long long crash_base, crash_size;
100 char *cmdline = boot_command_line;
101 bool high = false;
102 int ret;
103
104 if (!IS_ENABLED(CONFIG_CRASH_RESERVE))
105 return;
106
107 ret = parse_crashkernel(cmdline, memblock_phys_mem_size(),
108 &crash_size, &crash_base,
109 &low_size, &high);
110 if (ret)
111 return;
112
113 reserve_crashkernel_generic(cmdline, crash_size, crash_base,
114 low_size, high);
115 }
116
max_zone_phys(phys_addr_t zone_limit)117 static phys_addr_t __init max_zone_phys(phys_addr_t zone_limit)
118 {
119 /**
120 * Information we get from firmware (e.g. DT dma-ranges) describe DMA
121 * bus constraints. Devices using DMA might have their own limitations.
122 * Some of them rely on DMA zone in low 32-bit memory. Keep low RAM
123 * DMA zone on platforms that have RAM there.
124 */
125 if (memblock_start_of_DRAM() < U32_MAX)
126 zone_limit = min(zone_limit, U32_MAX);
127
128 return min(zone_limit, memblock_end_of_DRAM() - 1) + 1;
129 }
130
zone_sizes_init(void)131 static void __init zone_sizes_init(void)
132 {
133 unsigned long max_zone_pfns[MAX_NR_ZONES] = {0};
134 phys_addr_t __maybe_unused acpi_zone_dma_limit;
135 phys_addr_t __maybe_unused dt_zone_dma_limit;
136 phys_addr_t __maybe_unused dma32_phys_limit =
137 max_zone_phys(DMA_BIT_MASK(32));
138
139 #ifdef CONFIG_ZONE_DMA
140 acpi_zone_dma_limit = acpi_iort_dma_get_max_cpu_address();
141 dt_zone_dma_limit = of_dma_get_max_cpu_address(NULL);
142 zone_dma_limit = min(dt_zone_dma_limit, acpi_zone_dma_limit);
143 arm64_dma_phys_limit = max_zone_phys(zone_dma_limit);
144 max_zone_pfns[ZONE_DMA] = PFN_DOWN(arm64_dma_phys_limit);
145 #endif
146 #ifdef CONFIG_ZONE_DMA32
147 max_zone_pfns[ZONE_DMA32] = PFN_DOWN(dma32_phys_limit);
148 if (!arm64_dma_phys_limit)
149 arm64_dma_phys_limit = dma32_phys_limit;
150 #endif
151 if (!arm64_dma_phys_limit)
152 arm64_dma_phys_limit = PHYS_MASK + 1;
153 max_zone_pfns[ZONE_NORMAL] = max_pfn;
154
155 free_area_init(max_zone_pfns);
156 }
157
pfn_is_map_memory(unsigned long pfn)158 int pfn_is_map_memory(unsigned long pfn)
159 {
160 phys_addr_t addr = PFN_PHYS(pfn);
161
162 /* avoid false positives for bogus PFNs, see comment in pfn_valid() */
163 if (PHYS_PFN(addr) != pfn)
164 return 0;
165
166 return memblock_is_map_memory(addr);
167 }
168 EXPORT_SYMBOL(pfn_is_map_memory);
169
170 static phys_addr_t memory_limit __ro_after_init = PHYS_ADDR_MAX;
171
172 /*
173 * Limit the memory size that was specified via FDT.
174 */
early_mem(char * p)175 static int __init early_mem(char *p)
176 {
177 if (!p)
178 return 1;
179
180 memory_limit = memparse(p, &p) & PAGE_MASK;
181 pr_notice("Memory limited to %lldMB\n", memory_limit >> 20);
182
183 return 0;
184 }
185 early_param("mem", early_mem);
186
arm64_memblock_init(void)187 void __init arm64_memblock_init(void)
188 {
189 s64 linear_region_size = PAGE_END - _PAGE_OFFSET(vabits_actual);
190
191 /*
192 * Corner case: 52-bit VA capable systems running KVM in nVHE mode may
193 * be limited in their ability to support a linear map that exceeds 51
194 * bits of VA space, depending on the placement of the ID map. Given
195 * that the placement of the ID map may be randomized, let's simply
196 * limit the kernel's linear map to 51 bits as well if we detect this
197 * configuration.
198 */
199 if (IS_ENABLED(CONFIG_KVM) && vabits_actual == 52 &&
200 is_hyp_mode_available() && !is_kernel_in_hyp_mode()) {
201 pr_info("Capping linear region to 51 bits for KVM in nVHE mode on LVA capable hardware.\n");
202 linear_region_size = min_t(u64, linear_region_size, BIT(51));
203 }
204
205 /* Remove memory above our supported physical address size */
206 memblock_remove(1ULL << PHYS_MASK_SHIFT, ULLONG_MAX);
207
208 /*
209 * Select a suitable value for the base of physical memory.
210 */
211 memstart_addr = round_down(memblock_start_of_DRAM(),
212 ARM64_MEMSTART_ALIGN);
213
214 if ((memblock_end_of_DRAM() - memstart_addr) > linear_region_size)
215 pr_warn("Memory doesn't fit in the linear mapping, VA_BITS too small\n");
216
217 /*
218 * Remove the memory that we will not be able to cover with the
219 * linear mapping. Take care not to clip the kernel which may be
220 * high in memory.
221 */
222 memblock_remove(max_t(u64, memstart_addr + linear_region_size,
223 __pa_symbol(_end)), ULLONG_MAX);
224 if (memstart_addr + linear_region_size < memblock_end_of_DRAM()) {
225 /* ensure that memstart_addr remains sufficiently aligned */
226 memstart_addr = round_up(memblock_end_of_DRAM() - linear_region_size,
227 ARM64_MEMSTART_ALIGN);
228 memblock_remove(0, memstart_addr);
229 }
230
231 /*
232 * If we are running with a 52-bit kernel VA config on a system that
233 * does not support it, we have to place the available physical
234 * memory in the 48-bit addressable part of the linear region, i.e.,
235 * we have to move it upward. Since memstart_addr represents the
236 * physical address of PAGE_OFFSET, we have to *subtract* from it.
237 */
238 if (IS_ENABLED(CONFIG_ARM64_VA_BITS_52) && (vabits_actual != 52))
239 memstart_addr -= _PAGE_OFFSET(vabits_actual) - _PAGE_OFFSET(52);
240
241 /*
242 * Apply the memory limit if it was set. Since the kernel may be loaded
243 * high up in memory, add back the kernel region that must be accessible
244 * via the linear mapping.
245 */
246 if (memory_limit != PHYS_ADDR_MAX) {
247 memblock_mem_limit_remove_map(memory_limit);
248 memblock_add(__pa_symbol(_text), (u64)(_end - _text));
249 }
250
251 if (IS_ENABLED(CONFIG_BLK_DEV_INITRD) && phys_initrd_size) {
252 /*
253 * Add back the memory we just removed if it results in the
254 * initrd to become inaccessible via the linear mapping.
255 * Otherwise, this is a no-op
256 */
257 u64 base = phys_initrd_start & PAGE_MASK;
258 u64 size = PAGE_ALIGN(phys_initrd_start + phys_initrd_size) - base;
259
260 /*
261 * We can only add back the initrd memory if we don't end up
262 * with more memory than we can address via the linear mapping.
263 * It is up to the bootloader to position the kernel and the
264 * initrd reasonably close to each other (i.e., within 32 GB of
265 * each other) so that all granule/#levels combinations can
266 * always access both.
267 */
268 if (WARN(base < memblock_start_of_DRAM() ||
269 base + size > memblock_start_of_DRAM() +
270 linear_region_size,
271 "initrd not fully accessible via the linear mapping -- please check your bootloader ...\n")) {
272 phys_initrd_size = 0;
273 } else {
274 memblock_add(base, size);
275 memblock_clear_nomap(base, size);
276 memblock_reserve(base, size);
277 }
278 }
279
280 if (IS_ENABLED(CONFIG_RANDOMIZE_BASE)) {
281 extern u16 memstart_offset_seed;
282 u64 mmfr0 = read_cpuid(ID_AA64MMFR0_EL1);
283 int parange = cpuid_feature_extract_unsigned_field(
284 mmfr0, ID_AA64MMFR0_EL1_PARANGE_SHIFT);
285 s64 range = linear_region_size -
286 BIT(id_aa64mmfr0_parange_to_phys_shift(parange));
287
288 /*
289 * If the size of the linear region exceeds, by a sufficient
290 * margin, the size of the region that the physical memory can
291 * span, randomize the linear region as well.
292 */
293 if (memstart_offset_seed > 0 && range >= (s64)ARM64_MEMSTART_ALIGN) {
294 range /= ARM64_MEMSTART_ALIGN;
295 memstart_addr -= ARM64_MEMSTART_ALIGN *
296 ((range * memstart_offset_seed) >> 16);
297 }
298 }
299
300 /*
301 * Register the kernel text, kernel data, initrd, and initial
302 * pagetables with memblock.
303 */
304 memblock_reserve(__pa_symbol(_stext), _end - _stext);
305 if (IS_ENABLED(CONFIG_BLK_DEV_INITRD) && phys_initrd_size) {
306 /* the generic initrd code expects virtual addresses */
307 initrd_start = __phys_to_virt(phys_initrd_start);
308 initrd_end = initrd_start + phys_initrd_size;
309 }
310
311 early_init_fdt_scan_reserved_mem();
312
313 high_memory = __va(memblock_end_of_DRAM() - 1) + 1;
314 }
315
bootmem_init(void)316 void __init bootmem_init(void)
317 {
318 unsigned long min, max;
319
320 min = PFN_UP(memblock_start_of_DRAM());
321 max = PFN_DOWN(memblock_end_of_DRAM());
322
323 early_memtest(min << PAGE_SHIFT, max << PAGE_SHIFT);
324
325 max_pfn = max_low_pfn = max;
326 min_low_pfn = min;
327
328 arch_numa_init();
329
330 /*
331 * must be done after arch_numa_init() which calls numa_init() to
332 * initialize node_online_map that gets used in hugetlb_cma_reserve()
333 * while allocating required CMA size across online nodes.
334 */
335 #if defined(CONFIG_HUGETLB_PAGE) && defined(CONFIG_CMA)
336 arm64_hugetlb_cma_reserve();
337 #endif
338
339 kvm_hyp_reserve();
340
341 /*
342 * sparse_init() tries to allocate memory from memblock, so must be
343 * done after the fixed reservations
344 */
345 sparse_init();
346 zone_sizes_init();
347
348 /*
349 * Reserve the CMA area after arm64_dma_phys_limit was initialised.
350 */
351 dma_contiguous_reserve(arm64_dma_phys_limit);
352
353 /*
354 * request_standard_resources() depends on crashkernel's memory being
355 * reserved, so do it here.
356 */
357 arch_reserve_crashkernel();
358
359 memblock_dump_all();
360 }
361
362 /*
363 * mem_init() marks the free areas in the mem_map and tells us how much memory
364 * is free. This is done after various parts of the system have claimed their
365 * memory after the kernel image.
366 */
mem_init(void)367 void __init mem_init(void)
368 {
369 bool swiotlb = max_pfn > PFN_DOWN(arm64_dma_phys_limit);
370
371 if (IS_ENABLED(CONFIG_DMA_BOUNCE_UNALIGNED_KMALLOC) && !swiotlb) {
372 /*
373 * If no bouncing needed for ZONE_DMA, reduce the swiotlb
374 * buffer for kmalloc() bouncing to 1MB per 1GB of RAM.
375 */
376 unsigned long size =
377 DIV_ROUND_UP(memblock_phys_mem_size(), 1024);
378 swiotlb_adjust_size(min(swiotlb_size_or_default(), size));
379 swiotlb = true;
380 }
381
382 swiotlb_init(swiotlb, SWIOTLB_VERBOSE);
383
384 /* this will put all unused low memory onto the freelists */
385 memblock_free_all();
386
387 /*
388 * Check boundaries twice: Some fundamental inconsistencies can be
389 * detected at build time already.
390 */
391 #ifdef CONFIG_COMPAT
392 BUILD_BUG_ON(TASK_SIZE_32 > DEFAULT_MAP_WINDOW_64);
393 #endif
394
395 /*
396 * Selected page table levels should match when derived from
397 * scratch using the virtual address range and page size.
398 */
399 BUILD_BUG_ON(ARM64_HW_PGTABLE_LEVELS(CONFIG_ARM64_VA_BITS) !=
400 CONFIG_PGTABLE_LEVELS);
401
402 if (PAGE_SIZE >= 16384 && get_num_physpages() <= 128) {
403 extern int sysctl_overcommit_memory;
404 /*
405 * On a machine this small we won't get anywhere without
406 * overcommit, so turn it on by default.
407 */
408 sysctl_overcommit_memory = OVERCOMMIT_ALWAYS;
409 }
410 }
411
free_initmem(void)412 void free_initmem(void)
413 {
414 void *lm_init_begin = lm_alias(__init_begin);
415 void *lm_init_end = lm_alias(__init_end);
416
417 WARN_ON(!IS_ALIGNED((unsigned long)lm_init_begin, PAGE_SIZE));
418 WARN_ON(!IS_ALIGNED((unsigned long)lm_init_end, PAGE_SIZE));
419
420 /* Delete __init region from memblock.reserved. */
421 memblock_free(lm_init_begin, lm_init_end - lm_init_begin);
422
423 free_reserved_area(lm_init_begin, lm_init_end,
424 POISON_FREE_INITMEM, "unused kernel");
425 /*
426 * Unmap the __init region but leave the VM area in place. This
427 * prevents the region from being reused for kernel modules, which
428 * is not supported by kallsyms.
429 */
430 vunmap_range((u64)__init_begin, (u64)__init_end);
431 }
432
dump_mem_limit(void)433 void dump_mem_limit(void)
434 {
435 if (memory_limit != PHYS_ADDR_MAX) {
436 pr_emerg("Memory Limit: %llu MB\n", memory_limit >> 20);
437 } else {
438 pr_emerg("Memory Limit: none\n");
439 }
440 }
441
442 #ifdef CONFIG_EXECMEM
443 static u64 module_direct_base __ro_after_init = 0;
444 static u64 module_plt_base __ro_after_init = 0;
445
446 /*
447 * Choose a random page-aligned base address for a window of 'size' bytes which
448 * entirely contains the interval [start, end - 1].
449 */
random_bounding_box(u64 size,u64 start,u64 end)450 static u64 __init random_bounding_box(u64 size, u64 start, u64 end)
451 {
452 u64 max_pgoff, pgoff;
453
454 if ((end - start) >= size)
455 return 0;
456
457 max_pgoff = (size - (end - start)) / PAGE_SIZE;
458 pgoff = get_random_u32_inclusive(0, max_pgoff);
459
460 return start - pgoff * PAGE_SIZE;
461 }
462
463 /*
464 * Modules may directly reference data and text anywhere within the kernel
465 * image and other modules. References using PREL32 relocations have a +/-2G
466 * range, and so we need to ensure that the entire kernel image and all modules
467 * fall within a 2G window such that these are always within range.
468 *
469 * Modules may directly branch to functions and code within the kernel text,
470 * and to functions and code within other modules. These branches will use
471 * CALL26/JUMP26 relocations with a +/-128M range. Without PLTs, we must ensure
472 * that the entire kernel text and all module text falls within a 128M window
473 * such that these are always within range. With PLTs, we can expand this to a
474 * 2G window.
475 *
476 * We chose the 128M region to surround the entire kernel image (rather than
477 * just the text) as using the same bounds for the 128M and 2G regions ensures
478 * by construction that we never select a 128M region that is not a subset of
479 * the 2G region. For very large and unusual kernel configurations this means
480 * we may fall back to PLTs where they could have been avoided, but this keeps
481 * the logic significantly simpler.
482 */
module_init_limits(void)483 static int __init module_init_limits(void)
484 {
485 u64 kernel_end = (u64)_end;
486 u64 kernel_start = (u64)_text;
487 u64 kernel_size = kernel_end - kernel_start;
488
489 /*
490 * The default modules region is placed immediately below the kernel
491 * image, and is large enough to use the full 2G relocation range.
492 */
493 BUILD_BUG_ON(KIMAGE_VADDR != MODULES_END);
494 BUILD_BUG_ON(MODULES_VSIZE < SZ_2G);
495
496 if (!kaslr_enabled()) {
497 if (kernel_size < SZ_128M)
498 module_direct_base = kernel_end - SZ_128M;
499 if (kernel_size < SZ_2G)
500 module_plt_base = kernel_end - SZ_2G;
501 } else {
502 u64 min = kernel_start;
503 u64 max = kernel_end;
504
505 if (IS_ENABLED(CONFIG_RANDOMIZE_MODULE_REGION_FULL)) {
506 pr_info("2G module region forced by RANDOMIZE_MODULE_REGION_FULL\n");
507 } else {
508 module_direct_base = random_bounding_box(SZ_128M, min, max);
509 if (module_direct_base) {
510 min = module_direct_base;
511 max = module_direct_base + SZ_128M;
512 }
513 }
514
515 module_plt_base = random_bounding_box(SZ_2G, min, max);
516 }
517
518 pr_info("%llu pages in range for non-PLT usage",
519 module_direct_base ? (SZ_128M - kernel_size) / PAGE_SIZE : 0);
520 pr_info("%llu pages in range for PLT usage",
521 module_plt_base ? (SZ_2G - kernel_size) / PAGE_SIZE : 0);
522
523 return 0;
524 }
525
526 static struct execmem_info execmem_info __ro_after_init;
527
execmem_arch_setup(void)528 struct execmem_info __init *execmem_arch_setup(void)
529 {
530 unsigned long fallback_start = 0, fallback_end = 0;
531 unsigned long start = 0, end = 0;
532
533 module_init_limits();
534
535 /*
536 * Where possible, prefer to allocate within direct branch range of the
537 * kernel such that no PLTs are necessary.
538 */
539 if (module_direct_base) {
540 start = module_direct_base;
541 end = module_direct_base + SZ_128M;
542
543 if (module_plt_base) {
544 fallback_start = module_plt_base;
545 fallback_end = module_plt_base + SZ_2G;
546 }
547 } else if (module_plt_base) {
548 start = module_plt_base;
549 end = module_plt_base + SZ_2G;
550 }
551
552 execmem_info = (struct execmem_info){
553 .ranges = {
554 [EXECMEM_DEFAULT] = {
555 .start = start,
556 .end = end,
557 .pgprot = PAGE_KERNEL,
558 .alignment = 1,
559 .fallback_start = fallback_start,
560 .fallback_end = fallback_end,
561 },
562 [EXECMEM_KPROBES] = {
563 .start = VMALLOC_START,
564 .end = VMALLOC_END,
565 .pgprot = PAGE_KERNEL_ROX,
566 .alignment = 1,
567 },
568 [EXECMEM_BPF] = {
569 .start = VMALLOC_START,
570 .end = VMALLOC_END,
571 .pgprot = PAGE_KERNEL,
572 .alignment = 1,
573 },
574 },
575 };
576
577 return &execmem_info;
578 }
579 #endif /* CONFIG_EXECMEM */
580