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
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
117 /*
118 * Return the maximum physical address for a zone accessible by the given bits
119 * limit. If DRAM starts above 32-bit, expand the zone to the maximum
120 * available memory, otherwise cap it at 32-bit.
121 */
122 static phys_addr_t __init max_zone_phys(unsigned int zone_bits)
123 {
124 phys_addr_t zone_mask = DMA_BIT_MASK(zone_bits);
125 phys_addr_t phys_start = memblock_start_of_DRAM();
126
127 if (phys_start > U32_MAX)
128 zone_mask = PHYS_ADDR_MAX;
129 else if (phys_start > zone_mask)
130 zone_mask = U32_MAX;
131
132 return min(zone_mask, memblock_end_of_DRAM() - 1) + 1;
133 }
134
135 static void __init zone_sizes_init(void)
136 {
137 unsigned long max_zone_pfns[MAX_NR_ZONES] = {0};
138 unsigned int __maybe_unused acpi_zone_dma_bits;
139 unsigned int __maybe_unused dt_zone_dma_bits;
140 phys_addr_t __maybe_unused dma32_phys_limit = max_zone_phys(32);
141
142 #ifdef CONFIG_ZONE_DMA
143 acpi_zone_dma_bits = fls64(acpi_iort_dma_get_max_cpu_address());
144 dt_zone_dma_bits = fls64(of_dma_get_max_cpu_address(NULL));
145 zone_dma_bits = min3(32U, dt_zone_dma_bits, acpi_zone_dma_bits);
146 arm64_dma_phys_limit = max_zone_phys(zone_dma_bits);
147 max_zone_pfns[ZONE_DMA] = PFN_DOWN(arm64_dma_phys_limit);
148 #endif
149 #ifdef CONFIG_ZONE_DMA32
150 max_zone_pfns[ZONE_DMA32] = PFN_DOWN(dma32_phys_limit);
151 if (!arm64_dma_phys_limit)
152 arm64_dma_phys_limit = dma32_phys_limit;
153 #endif
154 if (!arm64_dma_phys_limit)
155 arm64_dma_phys_limit = PHYS_MASK + 1;
156 max_zone_pfns[ZONE_NORMAL] = max_pfn;
157
158 free_area_init(max_zone_pfns);
159 }
160
161 int pfn_is_map_memory(unsigned long pfn)
162 {
163 phys_addr_t addr = PFN_PHYS(pfn);
164
165 /* avoid false positives for bogus PFNs, see comment in pfn_valid() */
166 if (PHYS_PFN(addr) != pfn)
167 return 0;
168
169 return memblock_is_map_memory(addr);
170 }
171 EXPORT_SYMBOL(pfn_is_map_memory);
172
173 static phys_addr_t memory_limit __ro_after_init = PHYS_ADDR_MAX;
174
175 /*
176 * Limit the memory size that was specified via FDT.
177 */
178 static int __init early_mem(char *p)
179 {
180 if (!p)
181 return 1;
182
183 memory_limit = memparse(p, &p) & PAGE_MASK;
184 pr_notice("Memory limited to %lldMB\n", memory_limit >> 20);
185
186 return 0;
187 }
188 early_param("mem", early_mem);
189
190 void __init arm64_memblock_init(void)
191 {
192 s64 linear_region_size = PAGE_END - _PAGE_OFFSET(vabits_actual);
193
194 /*
195 * Corner case: 52-bit VA capable systems running KVM in nVHE mode may
196 * be limited in their ability to support a linear map that exceeds 51
197 * bits of VA space, depending on the placement of the ID map. Given
198 * that the placement of the ID map may be randomized, let's simply
199 * limit the kernel's linear map to 51 bits as well if we detect this
200 * configuration.
201 */
202 if (IS_ENABLED(CONFIG_KVM) && vabits_actual == 52 &&
203 is_hyp_mode_available() && !is_kernel_in_hyp_mode()) {
204 pr_info("Capping linear region to 51 bits for KVM in nVHE mode on LVA capable hardware.\n");
205 linear_region_size = min_t(u64, linear_region_size, BIT(51));
206 }
207
208 /* Remove memory above our supported physical address size */
209 memblock_remove(1ULL << PHYS_MASK_SHIFT, ULLONG_MAX);
210
211 /*
212 * Select a suitable value for the base of physical memory.
213 */
214 memstart_addr = round_down(memblock_start_of_DRAM(),
215 ARM64_MEMSTART_ALIGN);
216
217 if ((memblock_end_of_DRAM() - memstart_addr) > linear_region_size)
218 pr_warn("Memory doesn't fit in the linear mapping, VA_BITS too small\n");
219
220 /*
221 * Remove the memory that we will not be able to cover with the
222 * linear mapping. Take care not to clip the kernel which may be
223 * high in memory.
224 */
225 memblock_remove(max_t(u64, memstart_addr + linear_region_size,
226 __pa_symbol(_end)), ULLONG_MAX);
227 if (memstart_addr + linear_region_size < memblock_end_of_DRAM()) {
228 /* ensure that memstart_addr remains sufficiently aligned */
229 memstart_addr = round_up(memblock_end_of_DRAM() - linear_region_size,
230 ARM64_MEMSTART_ALIGN);
231 memblock_remove(0, memstart_addr);
232 }
233
234 /*
235 * If we are running with a 52-bit kernel VA config on a system that
236 * does not support it, we have to place the available physical
237 * memory in the 48-bit addressable part of the linear region, i.e.,
238 * we have to move it upward. Since memstart_addr represents the
239 * physical address of PAGE_OFFSET, we have to *subtract* from it.
240 */
241 if (IS_ENABLED(CONFIG_ARM64_VA_BITS_52) && (vabits_actual != 52))
242 memstart_addr -= _PAGE_OFFSET(vabits_actual) - _PAGE_OFFSET(52);
243
244 /*
245 * Apply the memory limit if it was set. Since the kernel may be loaded
246 * high up in memory, add back the kernel region that must be accessible
247 * via the linear mapping.
248 */
249 if (memory_limit != PHYS_ADDR_MAX) {
250 memblock_mem_limit_remove_map(memory_limit);
251 memblock_add(__pa_symbol(_text), (u64)(_end - _text));
252 }
253
254 if (IS_ENABLED(CONFIG_BLK_DEV_INITRD) && phys_initrd_size) {
255 /*
256 * Add back the memory we just removed if it results in the
257 * initrd to become inaccessible via the linear mapping.
258 * Otherwise, this is a no-op
259 */
260 u64 base = phys_initrd_start & PAGE_MASK;
261 u64 size = PAGE_ALIGN(phys_initrd_start + phys_initrd_size) - base;
262
263 /*
264 * We can only add back the initrd memory if we don't end up
265 * with more memory than we can address via the linear mapping.
266 * It is up to the bootloader to position the kernel and the
267 * initrd reasonably close to each other (i.e., within 32 GB of
268 * each other) so that all granule/#levels combinations can
269 * always access both.
270 */
271 if (WARN(base < memblock_start_of_DRAM() ||
272 base + size > memblock_start_of_DRAM() +
273 linear_region_size,
274 "initrd not fully accessible via the linear mapping -- please check your bootloader ...\n")) {
275 phys_initrd_size = 0;
276 } else {
277 memblock_add(base, size);
278 memblock_clear_nomap(base, size);
279 memblock_reserve(base, size);
280 }
281 }
282
283 if (IS_ENABLED(CONFIG_RANDOMIZE_BASE)) {
284 extern u16 memstart_offset_seed;
285 u64 mmfr0 = read_cpuid(ID_AA64MMFR0_EL1);
286 int parange = cpuid_feature_extract_unsigned_field(
287 mmfr0, ID_AA64MMFR0_EL1_PARANGE_SHIFT);
288 s64 range = linear_region_size -
289 BIT(id_aa64mmfr0_parange_to_phys_shift(parange));
290
291 /*
292 * If the size of the linear region exceeds, by a sufficient
293 * margin, the size of the region that the physical memory can
294 * span, randomize the linear region as well.
295 */
296 if (memstart_offset_seed > 0 && range >= (s64)ARM64_MEMSTART_ALIGN) {
297 range /= ARM64_MEMSTART_ALIGN;
298 memstart_addr -= ARM64_MEMSTART_ALIGN *
299 ((range * memstart_offset_seed) >> 16);
300 }
301 }
302
303 /*
304 * Register the kernel text, kernel data, initrd, and initial
305 * pagetables with memblock.
306 */
307 memblock_reserve(__pa_symbol(_stext), _end - _stext);
308 if (IS_ENABLED(CONFIG_BLK_DEV_INITRD) && phys_initrd_size) {
309 /* the generic initrd code expects virtual addresses */
310 initrd_start = __phys_to_virt(phys_initrd_start);
311 initrd_end = initrd_start + phys_initrd_size;
312 }
313
314 early_init_fdt_scan_reserved_mem();
315
316 high_memory = __va(memblock_end_of_DRAM() - 1) + 1;
317 }
318
319 void __init bootmem_init(void)
320 {
321 unsigned long min, max;
322
323 min = PFN_UP(memblock_start_of_DRAM());
324 max = PFN_DOWN(memblock_end_of_DRAM());
325
326 early_memtest(min << PAGE_SHIFT, max << PAGE_SHIFT);
327
328 max_pfn = max_low_pfn = max;
329 min_low_pfn = min;
330
331 arch_numa_init();
332
333 /*
334 * must be done after arch_numa_init() which calls numa_init() to
335 * initialize node_online_map that gets used in hugetlb_cma_reserve()
336 * while allocating required CMA size across online nodes.
337 */
338 #if defined(CONFIG_HUGETLB_PAGE) && defined(CONFIG_CMA)
339 arm64_hugetlb_cma_reserve();
340 #endif
341
342 kvm_hyp_reserve();
343
344 /*
345 * sparse_init() tries to allocate memory from memblock, so must be
346 * done after the fixed reservations
347 */
348 sparse_init();
349 zone_sizes_init();
350
351 /*
352 * Reserve the CMA area after arm64_dma_phys_limit was initialised.
353 */
354 dma_contiguous_reserve(arm64_dma_phys_limit);
355
356 /*
357 * request_standard_resources() depends on crashkernel's memory being
358 * reserved, so do it here.
359 */
360 arch_reserve_crashkernel();
361
362 memblock_dump_all();
363 }
364
365 /*
366 * mem_init() marks the free areas in the mem_map and tells us how much memory
367 * is free. This is done after various parts of the system have claimed their
368 * memory after the kernel image.
369 */
370 void __init mem_init(void)
371 {
372 bool swiotlb = max_pfn > PFN_DOWN(arm64_dma_phys_limit);
373
374 if (IS_ENABLED(CONFIG_DMA_BOUNCE_UNALIGNED_KMALLOC) && !swiotlb) {
375 /*
376 * If no bouncing needed for ZONE_DMA, reduce the swiotlb
377 * buffer for kmalloc() bouncing to 1MB per 1GB of RAM.
378 */
379 unsigned long size =
380 DIV_ROUND_UP(memblock_phys_mem_size(), 1024);
381 swiotlb_adjust_size(min(swiotlb_size_or_default(), size));
382 swiotlb = true;
383 }
384
385 swiotlb_init(swiotlb, SWIOTLB_VERBOSE);
386
387 /* this will put all unused low memory onto the freelists */
388 memblock_free_all();
389
390 /*
391 * Check boundaries twice: Some fundamental inconsistencies can be
392 * detected at build time already.
393 */
394 #ifdef CONFIG_COMPAT
395 BUILD_BUG_ON(TASK_SIZE_32 > DEFAULT_MAP_WINDOW_64);
396 #endif
397
398 /*
399 * Selected page table levels should match when derived from
400 * scratch using the virtual address range and page size.
401 */
402 BUILD_BUG_ON(ARM64_HW_PGTABLE_LEVELS(CONFIG_ARM64_VA_BITS) !=
403 CONFIG_PGTABLE_LEVELS);
404
405 if (PAGE_SIZE >= 16384 && get_num_physpages() <= 128) {
406 extern int sysctl_overcommit_memory;
407 /*
408 * On a machine this small we won't get anywhere without
409 * overcommit, so turn it on by default.
410 */
411 sysctl_overcommit_memory = OVERCOMMIT_ALWAYS;
412 }
413 }
414
415 void free_initmem(void)
416 {
417 free_reserved_area(lm_alias(__init_begin),
418 lm_alias(__init_end),
419 POISON_FREE_INITMEM, "unused kernel");
420 /*
421 * Unmap the __init region but leave the VM area in place. This
422 * prevents the region from being reused for kernel modules, which
423 * is not supported by kallsyms.
424 */
425 vunmap_range((u64)__init_begin, (u64)__init_end);
426 }
427
428 void dump_mem_limit(void)
429 {
430 if (memory_limit != PHYS_ADDR_MAX) {
431 pr_emerg("Memory Limit: %llu MB\n", memory_limit >> 20);
432 } else {
433 pr_emerg("Memory Limit: none\n");
434 }
435 }
436
437 #ifdef CONFIG_EXECMEM
438 static u64 module_direct_base __ro_after_init = 0;
439 static u64 module_plt_base __ro_after_init = 0;
440
441 /*
442 * Choose a random page-aligned base address for a window of 'size' bytes which
443 * entirely contains the interval [start, end - 1].
444 */
445 static u64 __init random_bounding_box(u64 size, u64 start, u64 end)
446 {
447 u64 max_pgoff, pgoff;
448
449 if ((end - start) >= size)
450 return 0;
451
452 max_pgoff = (size - (end - start)) / PAGE_SIZE;
453 pgoff = get_random_u32_inclusive(0, max_pgoff);
454
455 return start - pgoff * PAGE_SIZE;
456 }
457
458 /*
459 * Modules may directly reference data and text anywhere within the kernel
460 * image and other modules. References using PREL32 relocations have a +/-2G
461 * range, and so we need to ensure that the entire kernel image and all modules
462 * fall within a 2G window such that these are always within range.
463 *
464 * Modules may directly branch to functions and code within the kernel text,
465 * and to functions and code within other modules. These branches will use
466 * CALL26/JUMP26 relocations with a +/-128M range. Without PLTs, we must ensure
467 * that the entire kernel text and all module text falls within a 128M window
468 * such that these are always within range. With PLTs, we can expand this to a
469 * 2G window.
470 *
471 * We chose the 128M region to surround the entire kernel image (rather than
472 * just the text) as using the same bounds for the 128M and 2G regions ensures
473 * by construction that we never select a 128M region that is not a subset of
474 * the 2G region. For very large and unusual kernel configurations this means
475 * we may fall back to PLTs where they could have been avoided, but this keeps
476 * the logic significantly simpler.
477 */
478 static int __init module_init_limits(void)
479 {
480 u64 kernel_end = (u64)_end;
481 u64 kernel_start = (u64)_text;
482 u64 kernel_size = kernel_end - kernel_start;
483
484 /*
485 * The default modules region is placed immediately below the kernel
486 * image, and is large enough to use the full 2G relocation range.
487 */
488 BUILD_BUG_ON(KIMAGE_VADDR != MODULES_END);
489 BUILD_BUG_ON(MODULES_VSIZE < SZ_2G);
490
491 if (!kaslr_enabled()) {
492 if (kernel_size < SZ_128M)
493 module_direct_base = kernel_end - SZ_128M;
494 if (kernel_size < SZ_2G)
495 module_plt_base = kernel_end - SZ_2G;
496 } else {
497 u64 min = kernel_start;
498 u64 max = kernel_end;
499
500 if (IS_ENABLED(CONFIG_RANDOMIZE_MODULE_REGION_FULL)) {
501 pr_info("2G module region forced by RANDOMIZE_MODULE_REGION_FULL\n");
502 } else {
503 module_direct_base = random_bounding_box(SZ_128M, min, max);
504 if (module_direct_base) {
505 min = module_direct_base;
506 max = module_direct_base + SZ_128M;
507 }
508 }
509
510 module_plt_base = random_bounding_box(SZ_2G, min, max);
511 }
512
513 pr_info("%llu pages in range for non-PLT usage",
514 module_direct_base ? (SZ_128M - kernel_size) / PAGE_SIZE : 0);
515 pr_info("%llu pages in range for PLT usage",
516 module_plt_base ? (SZ_2G - kernel_size) / PAGE_SIZE : 0);
517
518 return 0;
519 }
520
521 static struct execmem_info execmem_info __ro_after_init;
522
523 struct execmem_info __init *execmem_arch_setup(void)
524 {
525 unsigned long fallback_start = 0, fallback_end = 0;
526 unsigned long start = 0, end = 0;
527
528 module_init_limits();
529
530 /*
531 * Where possible, prefer to allocate within direct branch range of the
532 * kernel such that no PLTs are necessary.
533 */
534 if (module_direct_base) {
535 start = module_direct_base;
536 end = module_direct_base + SZ_128M;
537
538 if (module_plt_base) {
539 fallback_start = module_plt_base;
540 fallback_end = module_plt_base + SZ_2G;
541 }
542 } else if (module_plt_base) {
543 start = module_plt_base;
544 end = module_plt_base + SZ_2G;
545 }
546
547 execmem_info = (struct execmem_info){
548 .ranges = {
549 [EXECMEM_DEFAULT] = {
550 .start = start,
551 .end = end,
552 .pgprot = PAGE_KERNEL,
553 .alignment = 1,
554 .fallback_start = fallback_start,
555 .fallback_end = fallback_end,
556 },
557 [EXECMEM_KPROBES] = {
558 .start = VMALLOC_START,
559 .end = VMALLOC_END,
560 .pgprot = PAGE_KERNEL_ROX,
561 .alignment = 1,
562 },
563 [EXECMEM_BPF] = {
564 .start = VMALLOC_START,
565 .end = VMALLOC_END,
566 .pgprot = PAGE_KERNEL,
567 .alignment = 1,
568 },
569 },
570 };
571
572 return &execmem_info;
573 }
574 #endif /* CONFIG_EXECMEM */
575
Linux® is a registered trademark of Linus Torvalds in the United States and other countries.
TOMOYO® is a registered trademark of NTT DATA CORPORATION.