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