init.c

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