xref: /linux/arch/arm64/mm/init.c (revision 7f71507851fc7764b36a3221839607d3a45c2025)
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/rsi.h>
45 #include <asm/sections.h>
46 #include <asm/setup.h>
47 #include <linux/sizes.h>
48 #include <asm/tlb.h>
49 #include <asm/alternative.h>
50 #include <asm/xen/swiotlb-xen.h>
51 
52 /*
53  * We need to be able to catch inadvertent references to memstart_addr
54  * that occur (potentially in generic code) before arm64_memblock_init()
55  * executes, which assigns it its actual value. So use a default value
56  * that cannot be mistaken for a real physical address.
57  */
58 s64 memstart_addr __ro_after_init = -1;
59 EXPORT_SYMBOL(memstart_addr);
60 
61 /*
62  * If the corresponding config options are enabled, we create both ZONE_DMA
63  * and ZONE_DMA32. By default ZONE_DMA covers the 32-bit addressable memory
64  * unless restricted on specific platforms (e.g. 30-bit on Raspberry Pi 4).
65  * In such case, ZONE_DMA32 covers the rest of the 32-bit addressable memory,
66  * otherwise it is empty.
67  */
68 phys_addr_t __ro_after_init arm64_dma_phys_limit;
69 
70 /*
71  * To make optimal use of block mappings when laying out the linear
72  * mapping, round down the base of physical memory to a size that can
73  * be mapped efficiently, i.e., either PUD_SIZE (4k granule) or PMD_SIZE
74  * (64k granule), or a multiple that can be mapped using contiguous bits
75  * in the page tables: 32 * PMD_SIZE (16k granule)
76  */
77 #if defined(CONFIG_ARM64_4K_PAGES)
78 #define ARM64_MEMSTART_SHIFT		PUD_SHIFT
79 #elif defined(CONFIG_ARM64_16K_PAGES)
80 #define ARM64_MEMSTART_SHIFT		CONT_PMD_SHIFT
81 #else
82 #define ARM64_MEMSTART_SHIFT		PMD_SHIFT
83 #endif
84 
85 /*
86  * sparsemem vmemmap imposes an additional requirement on the alignment of
87  * memstart_addr, due to the fact that the base of the vmemmap region
88  * has a direct correspondence, and needs to appear sufficiently aligned
89  * in the virtual address space.
90  */
91 #if ARM64_MEMSTART_SHIFT < SECTION_SIZE_BITS
92 #define ARM64_MEMSTART_ALIGN	(1UL << SECTION_SIZE_BITS)
93 #else
94 #define ARM64_MEMSTART_ALIGN	(1UL << ARM64_MEMSTART_SHIFT)
95 #endif
96 
97 static void __init arch_reserve_crashkernel(void)
98 {
99 	unsigned long long low_size = 0;
100 	unsigned long long crash_base, crash_size;
101 	char *cmdline = boot_command_line;
102 	bool high = false;
103 	int ret;
104 
105 	if (!IS_ENABLED(CONFIG_CRASH_RESERVE))
106 		return;
107 
108 	ret = parse_crashkernel(cmdline, memblock_phys_mem_size(),
109 				&crash_size, &crash_base,
110 				&low_size, &high);
111 	if (ret)
112 		return;
113 
114 	reserve_crashkernel_generic(cmdline, crash_size, crash_base,
115 				    low_size, high);
116 }
117 
118 static phys_addr_t __init max_zone_phys(phys_addr_t zone_limit)
119 {
120 	/**
121 	 * Information we get from firmware (e.g. DT dma-ranges) describe DMA
122 	 * bus constraints. Devices using DMA might have their own limitations.
123 	 * Some of them rely on DMA zone in low 32-bit memory. Keep low RAM
124 	 * DMA zone on platforms that have RAM there.
125 	 */
126 	if (memblock_start_of_DRAM() < U32_MAX)
127 		zone_limit = min(zone_limit, U32_MAX);
128 
129 	return min(zone_limit, memblock_end_of_DRAM() - 1) + 1;
130 }
131 
132 static void __init zone_sizes_init(void)
133 {
134 	unsigned long max_zone_pfns[MAX_NR_ZONES]  = {0};
135 	phys_addr_t __maybe_unused acpi_zone_dma_limit;
136 	phys_addr_t __maybe_unused dt_zone_dma_limit;
137 	phys_addr_t __maybe_unused dma32_phys_limit =
138 		max_zone_phys(DMA_BIT_MASK(32));
139 
140 #ifdef CONFIG_ZONE_DMA
141 	acpi_zone_dma_limit = acpi_iort_dma_get_max_cpu_address();
142 	dt_zone_dma_limit = of_dma_get_max_cpu_address(NULL);
143 	zone_dma_limit = min(dt_zone_dma_limit, acpi_zone_dma_limit);
144 	arm64_dma_phys_limit = max_zone_phys(zone_dma_limit);
145 	max_zone_pfns[ZONE_DMA] = PFN_DOWN(arm64_dma_phys_limit);
146 #endif
147 #ifdef CONFIG_ZONE_DMA32
148 	max_zone_pfns[ZONE_DMA32] = PFN_DOWN(dma32_phys_limit);
149 	if (!arm64_dma_phys_limit)
150 		arm64_dma_phys_limit = dma32_phys_limit;
151 #endif
152 	if (!arm64_dma_phys_limit)
153 		arm64_dma_phys_limit = PHYS_MASK + 1;
154 	max_zone_pfns[ZONE_NORMAL] = max_pfn;
155 
156 	free_area_init(max_zone_pfns);
157 }
158 
159 int pfn_is_map_memory(unsigned long pfn)
160 {
161 	phys_addr_t addr = PFN_PHYS(pfn);
162 
163 	/* avoid false positives for bogus PFNs, see comment in pfn_valid() */
164 	if (PHYS_PFN(addr) != pfn)
165 		return 0;
166 
167 	return memblock_is_map_memory(addr);
168 }
169 EXPORT_SYMBOL(pfn_is_map_memory);
170 
171 static phys_addr_t memory_limit __ro_after_init = PHYS_ADDR_MAX;
172 
173 /*
174  * Limit the memory size that was specified via FDT.
175  */
176 static int __init early_mem(char *p)
177 {
178 	if (!p)
179 		return 1;
180 
181 	memory_limit = memparse(p, &p) & PAGE_MASK;
182 	pr_notice("Memory limited to %lldMB\n", memory_limit >> 20);
183 
184 	return 0;
185 }
186 early_param("mem", early_mem);
187 
188 void __init arm64_memblock_init(void)
189 {
190 	s64 linear_region_size = PAGE_END - _PAGE_OFFSET(vabits_actual);
191 
192 	/*
193 	 * Corner case: 52-bit VA capable systems running KVM in nVHE mode may
194 	 * be limited in their ability to support a linear map that exceeds 51
195 	 * bits of VA space, depending on the placement of the ID map. Given
196 	 * that the placement of the ID map may be randomized, let's simply
197 	 * limit the kernel's linear map to 51 bits as well if we detect this
198 	 * configuration.
199 	 */
200 	if (IS_ENABLED(CONFIG_KVM) && vabits_actual == 52 &&
201 	    is_hyp_mode_available() && !is_kernel_in_hyp_mode()) {
202 		pr_info("Capping linear region to 51 bits for KVM in nVHE mode on LVA capable hardware.\n");
203 		linear_region_size = min_t(u64, linear_region_size, BIT(51));
204 	}
205 
206 	/* Remove memory above our supported physical address size */
207 	memblock_remove(1ULL << PHYS_MASK_SHIFT, ULLONG_MAX);
208 
209 	/*
210 	 * Select a suitable value for the base of physical memory.
211 	 */
212 	memstart_addr = round_down(memblock_start_of_DRAM(),
213 				   ARM64_MEMSTART_ALIGN);
214 
215 	if ((memblock_end_of_DRAM() - memstart_addr) > linear_region_size)
216 		pr_warn("Memory doesn't fit in the linear mapping, VA_BITS too small\n");
217 
218 	/*
219 	 * Remove the memory that we will not be able to cover with the
220 	 * linear mapping. Take care not to clip the kernel which may be
221 	 * high in memory.
222 	 */
223 	memblock_remove(max_t(u64, memstart_addr + linear_region_size,
224 			__pa_symbol(_end)), ULLONG_MAX);
225 	if (memstart_addr + linear_region_size < memblock_end_of_DRAM()) {
226 		/* ensure that memstart_addr remains sufficiently aligned */
227 		memstart_addr = round_up(memblock_end_of_DRAM() - linear_region_size,
228 					 ARM64_MEMSTART_ALIGN);
229 		memblock_remove(0, memstart_addr);
230 	}
231 
232 	/*
233 	 * If we are running with a 52-bit kernel VA config on a system that
234 	 * does not support it, we have to place the available physical
235 	 * memory in the 48-bit addressable part of the linear region, i.e.,
236 	 * we have to move it upward. Since memstart_addr represents the
237 	 * physical address of PAGE_OFFSET, we have to *subtract* from it.
238 	 */
239 	if (IS_ENABLED(CONFIG_ARM64_VA_BITS_52) && (vabits_actual != 52))
240 		memstart_addr -= _PAGE_OFFSET(vabits_actual) - _PAGE_OFFSET(52);
241 
242 	/*
243 	 * Apply the memory limit if it was set. Since the kernel may be loaded
244 	 * high up in memory, add back the kernel region that must be accessible
245 	 * via the linear mapping.
246 	 */
247 	if (memory_limit != PHYS_ADDR_MAX) {
248 		memblock_mem_limit_remove_map(memory_limit);
249 		memblock_add(__pa_symbol(_text), (u64)(_end - _text));
250 	}
251 
252 	if (IS_ENABLED(CONFIG_BLK_DEV_INITRD) && phys_initrd_size) {
253 		/*
254 		 * Add back the memory we just removed if it results in the
255 		 * initrd to become inaccessible via the linear mapping.
256 		 * Otherwise, this is a no-op
257 		 */
258 		u64 base = phys_initrd_start & PAGE_MASK;
259 		u64 size = PAGE_ALIGN(phys_initrd_start + phys_initrd_size) - base;
260 
261 		/*
262 		 * We can only add back the initrd memory if we don't end up
263 		 * with more memory than we can address via the linear mapping.
264 		 * It is up to the bootloader to position the kernel and the
265 		 * initrd reasonably close to each other (i.e., within 32 GB of
266 		 * each other) so that all granule/#levels combinations can
267 		 * always access both.
268 		 */
269 		if (WARN(base < memblock_start_of_DRAM() ||
270 			 base + size > memblock_start_of_DRAM() +
271 				       linear_region_size,
272 			"initrd not fully accessible via the linear mapping -- please check your bootloader ...\n")) {
273 			phys_initrd_size = 0;
274 		} else {
275 			memblock_add(base, size);
276 			memblock_clear_nomap(base, size);
277 			memblock_reserve(base, size);
278 		}
279 	}
280 
281 	if (IS_ENABLED(CONFIG_RANDOMIZE_BASE)) {
282 		extern u16 memstart_offset_seed;
283 		u64 mmfr0 = read_cpuid(ID_AA64MMFR0_EL1);
284 		int parange = cpuid_feature_extract_unsigned_field(
285 					mmfr0, ID_AA64MMFR0_EL1_PARANGE_SHIFT);
286 		s64 range = linear_region_size -
287 			    BIT(id_aa64mmfr0_parange_to_phys_shift(parange));
288 
289 		/*
290 		 * If the size of the linear region exceeds, by a sufficient
291 		 * margin, the size of the region that the physical memory can
292 		 * span, randomize the linear region as well.
293 		 */
294 		if (memstart_offset_seed > 0 && range >= (s64)ARM64_MEMSTART_ALIGN) {
295 			range /= ARM64_MEMSTART_ALIGN;
296 			memstart_addr -= ARM64_MEMSTART_ALIGN *
297 					 ((range * memstart_offset_seed) >> 16);
298 		}
299 	}
300 
301 	/*
302 	 * Register the kernel text, kernel data, initrd, and initial
303 	 * pagetables with memblock.
304 	 */
305 	memblock_reserve(__pa_symbol(_stext), _end - _stext);
306 	if (IS_ENABLED(CONFIG_BLK_DEV_INITRD) && phys_initrd_size) {
307 		/* the generic initrd code expects virtual addresses */
308 		initrd_start = __phys_to_virt(phys_initrd_start);
309 		initrd_end = initrd_start + phys_initrd_size;
310 	}
311 
312 	early_init_fdt_scan_reserved_mem();
313 
314 	high_memory = __va(memblock_end_of_DRAM() - 1) + 1;
315 }
316 
317 void __init bootmem_init(void)
318 {
319 	unsigned long min, max;
320 
321 	min = PFN_UP(memblock_start_of_DRAM());
322 	max = PFN_DOWN(memblock_end_of_DRAM());
323 
324 	early_memtest(min << PAGE_SHIFT, max << PAGE_SHIFT);
325 
326 	max_pfn = max_low_pfn = max;
327 	min_low_pfn = min;
328 
329 	arch_numa_init();
330 
331 	/*
332 	 * must be done after arch_numa_init() which calls numa_init() to
333 	 * initialize node_online_map that gets used in hugetlb_cma_reserve()
334 	 * while allocating required CMA size across online nodes.
335 	 */
336 #if defined(CONFIG_HUGETLB_PAGE) && defined(CONFIG_CMA)
337 	arm64_hugetlb_cma_reserve();
338 #endif
339 
340 	kvm_hyp_reserve();
341 
342 	/*
343 	 * sparse_init() tries to allocate memory from memblock, so must be
344 	 * done after the fixed reservations
345 	 */
346 	sparse_init();
347 	zone_sizes_init();
348 
349 	/*
350 	 * Reserve the CMA area after arm64_dma_phys_limit was initialised.
351 	 */
352 	dma_contiguous_reserve(arm64_dma_phys_limit);
353 
354 	/*
355 	 * request_standard_resources() depends on crashkernel's memory being
356 	 * reserved, so do it here.
357 	 */
358 	arch_reserve_crashkernel();
359 
360 	memblock_dump_all();
361 }
362 
363 /*
364  * mem_init() marks the free areas in the mem_map and tells us how much memory
365  * is free.  This is done after various parts of the system have claimed their
366  * memory after the kernel image.
367  */
368 void __init mem_init(void)
369 {
370 	unsigned int flags = SWIOTLB_VERBOSE;
371 	bool swiotlb = max_pfn > PFN_DOWN(arm64_dma_phys_limit);
372 
373 	if (is_realm_world()) {
374 		swiotlb = true;
375 		flags |= SWIOTLB_FORCE;
376 	}
377 
378 	if (IS_ENABLED(CONFIG_DMA_BOUNCE_UNALIGNED_KMALLOC) && !swiotlb) {
379 		/*
380 		 * If no bouncing needed for ZONE_DMA, reduce the swiotlb
381 		 * buffer for kmalloc() bouncing to 1MB per 1GB of RAM.
382 		 */
383 		unsigned long size =
384 			DIV_ROUND_UP(memblock_phys_mem_size(), 1024);
385 		swiotlb_adjust_size(min(swiotlb_size_or_default(), size));
386 		swiotlb = true;
387 	}
388 
389 	swiotlb_init(swiotlb, flags);
390 	swiotlb_update_mem_attributes();
391 
392 	/* this will put all unused low memory onto the freelists */
393 	memblock_free_all();
394 
395 	/*
396 	 * Check boundaries twice: Some fundamental inconsistencies can be
397 	 * detected at build time already.
398 	 */
399 #ifdef CONFIG_COMPAT
400 	BUILD_BUG_ON(TASK_SIZE_32 > DEFAULT_MAP_WINDOW_64);
401 #endif
402 
403 	/*
404 	 * Selected page table levels should match when derived from
405 	 * scratch using the virtual address range and page size.
406 	 */
407 	BUILD_BUG_ON(ARM64_HW_PGTABLE_LEVELS(CONFIG_ARM64_VA_BITS) !=
408 		     CONFIG_PGTABLE_LEVELS);
409 
410 	if (PAGE_SIZE >= 16384 && get_num_physpages() <= 128) {
411 		extern int sysctl_overcommit_memory;
412 		/*
413 		 * On a machine this small we won't get anywhere without
414 		 * overcommit, so turn it on by default.
415 		 */
416 		sysctl_overcommit_memory = OVERCOMMIT_ALWAYS;
417 	}
418 }
419 
420 void free_initmem(void)
421 {
422 	void *lm_init_begin = lm_alias(__init_begin);
423 	void *lm_init_end = lm_alias(__init_end);
424 
425 	WARN_ON(!IS_ALIGNED((unsigned long)lm_init_begin, PAGE_SIZE));
426 	WARN_ON(!IS_ALIGNED((unsigned long)lm_init_end, PAGE_SIZE));
427 
428 	/* Delete __init region from memblock.reserved. */
429 	memblock_free(lm_init_begin, lm_init_end - lm_init_begin);
430 
431 	free_reserved_area(lm_init_begin, lm_init_end,
432 			   POISON_FREE_INITMEM, "unused kernel");
433 	/*
434 	 * Unmap the __init region but leave the VM area in place. This
435 	 * prevents the region from being reused for kernel modules, which
436 	 * is not supported by kallsyms.
437 	 */
438 	vunmap_range((u64)__init_begin, (u64)__init_end);
439 }
440 
441 void dump_mem_limit(void)
442 {
443 	if (memory_limit != PHYS_ADDR_MAX) {
444 		pr_emerg("Memory Limit: %llu MB\n", memory_limit >> 20);
445 	} else {
446 		pr_emerg("Memory Limit: none\n");
447 	}
448 }
449 
450 #ifdef CONFIG_EXECMEM
451 static u64 module_direct_base __ro_after_init = 0;
452 static u64 module_plt_base __ro_after_init = 0;
453 
454 /*
455  * Choose a random page-aligned base address for a window of 'size' bytes which
456  * entirely contains the interval [start, end - 1].
457  */
458 static u64 __init random_bounding_box(u64 size, u64 start, u64 end)
459 {
460 	u64 max_pgoff, pgoff;
461 
462 	if ((end - start) >= size)
463 		return 0;
464 
465 	max_pgoff = (size - (end - start)) / PAGE_SIZE;
466 	pgoff = get_random_u32_inclusive(0, max_pgoff);
467 
468 	return start - pgoff * PAGE_SIZE;
469 }
470 
471 /*
472  * Modules may directly reference data and text anywhere within the kernel
473  * image and other modules. References using PREL32 relocations have a +/-2G
474  * range, and so we need to ensure that the entire kernel image and all modules
475  * fall within a 2G window such that these are always within range.
476  *
477  * Modules may directly branch to functions and code within the kernel text,
478  * and to functions and code within other modules. These branches will use
479  * CALL26/JUMP26 relocations with a +/-128M range. Without PLTs, we must ensure
480  * that the entire kernel text and all module text falls within a 128M window
481  * such that these are always within range. With PLTs, we can expand this to a
482  * 2G window.
483  *
484  * We chose the 128M region to surround the entire kernel image (rather than
485  * just the text) as using the same bounds for the 128M and 2G regions ensures
486  * by construction that we never select a 128M region that is not a subset of
487  * the 2G region. For very large and unusual kernel configurations this means
488  * we may fall back to PLTs where they could have been avoided, but this keeps
489  * the logic significantly simpler.
490  */
491 static int __init module_init_limits(void)
492 {
493 	u64 kernel_end = (u64)_end;
494 	u64 kernel_start = (u64)_text;
495 	u64 kernel_size = kernel_end - kernel_start;
496 
497 	/*
498 	 * The default modules region is placed immediately below the kernel
499 	 * image, and is large enough to use the full 2G relocation range.
500 	 */
501 	BUILD_BUG_ON(KIMAGE_VADDR != MODULES_END);
502 	BUILD_BUG_ON(MODULES_VSIZE < SZ_2G);
503 
504 	if (!kaslr_enabled()) {
505 		if (kernel_size < SZ_128M)
506 			module_direct_base = kernel_end - SZ_128M;
507 		if (kernel_size < SZ_2G)
508 			module_plt_base = kernel_end - SZ_2G;
509 	} else {
510 		u64 min = kernel_start;
511 		u64 max = kernel_end;
512 
513 		if (IS_ENABLED(CONFIG_RANDOMIZE_MODULE_REGION_FULL)) {
514 			pr_info("2G module region forced by RANDOMIZE_MODULE_REGION_FULL\n");
515 		} else {
516 			module_direct_base = random_bounding_box(SZ_128M, min, max);
517 			if (module_direct_base) {
518 				min = module_direct_base;
519 				max = module_direct_base + SZ_128M;
520 			}
521 		}
522 
523 		module_plt_base = random_bounding_box(SZ_2G, min, max);
524 	}
525 
526 	pr_info("%llu pages in range for non-PLT usage",
527 		module_direct_base ? (SZ_128M - kernel_size) / PAGE_SIZE : 0);
528 	pr_info("%llu pages in range for PLT usage",
529 		module_plt_base ? (SZ_2G - kernel_size) / PAGE_SIZE : 0);
530 
531 	return 0;
532 }
533 
534 static struct execmem_info execmem_info __ro_after_init;
535 
536 struct execmem_info __init *execmem_arch_setup(void)
537 {
538 	unsigned long fallback_start = 0, fallback_end = 0;
539 	unsigned long start = 0, end = 0;
540 
541 	module_init_limits();
542 
543 	/*
544 	 * Where possible, prefer to allocate within direct branch range of the
545 	 * kernel such that no PLTs are necessary.
546 	 */
547 	if (module_direct_base) {
548 		start = module_direct_base;
549 		end = module_direct_base + SZ_128M;
550 
551 		if (module_plt_base) {
552 			fallback_start = module_plt_base;
553 			fallback_end = module_plt_base + SZ_2G;
554 		}
555 	} else if (module_plt_base) {
556 		start = module_plt_base;
557 		end = module_plt_base + SZ_2G;
558 	}
559 
560 	execmem_info = (struct execmem_info){
561 		.ranges = {
562 			[EXECMEM_DEFAULT] = {
563 				.start	= start,
564 				.end	= end,
565 				.pgprot	= PAGE_KERNEL,
566 				.alignment = 1,
567 				.fallback_start	= fallback_start,
568 				.fallback_end	= fallback_end,
569 			},
570 			[EXECMEM_KPROBES] = {
571 				.start	= VMALLOC_START,
572 				.end	= VMALLOC_END,
573 				.pgprot	= PAGE_KERNEL_ROX,
574 				.alignment = 1,
575 			},
576 			[EXECMEM_BPF] = {
577 				.start	= VMALLOC_START,
578 				.end	= VMALLOC_END,
579 				.pgprot	= PAGE_KERNEL,
580 				.alignment = 1,
581 			},
582 		},
583 	};
584 
585 	return &execmem_info;
586 }
587 #endif /* CONFIG_EXECMEM */
588