xref: /openbmc/linux/arch/arm64/mm/init.c (revision 5f66f73b)
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/memblock.h>
20 #include <linux/sort.h>
21 #include <linux/of.h>
22 #include <linux/of_fdt.h>
23 #include <linux/dma-direct.h>
24 #include <linux/dma-map-ops.h>
25 #include <linux/efi.h>
26 #include <linux/swiotlb.h>
27 #include <linux/vmalloc.h>
28 #include <linux/mm.h>
29 #include <linux/kexec.h>
30 #include <linux/crash_dump.h>
31 #include <linux/hugetlb.h>
32 #include <linux/acpi_iort.h>
33 
34 #include <asm/boot.h>
35 #include <asm/fixmap.h>
36 #include <asm/kasan.h>
37 #include <asm/kernel-pgtable.h>
38 #include <asm/kvm_host.h>
39 #include <asm/memory.h>
40 #include <asm/numa.h>
41 #include <asm/sections.h>
42 #include <asm/setup.h>
43 #include <linux/sizes.h>
44 #include <asm/tlb.h>
45 #include <asm/alternative.h>
46 
47 /*
48  * We need to be able to catch inadvertent references to memstart_addr
49  * that occur (potentially in generic code) before arm64_memblock_init()
50  * executes, which assigns it its actual value. So use a default value
51  * that cannot be mistaken for a real physical address.
52  */
53 s64 memstart_addr __ro_after_init = -1;
54 EXPORT_SYMBOL(memstart_addr);
55 
56 /*
57  * If the corresponding config options are enabled, we create both ZONE_DMA
58  * and ZONE_DMA32. By default ZONE_DMA covers the 32-bit addressable memory
59  * unless restricted on specific platforms (e.g. 30-bit on Raspberry Pi 4).
60  * In such case, ZONE_DMA32 covers the rest of the 32-bit addressable memory,
61  * otherwise it is empty.
62  */
63 phys_addr_t arm64_dma_phys_limit __ro_after_init;
64 
65 #ifdef CONFIG_KEXEC_CORE
66 /*
67  * reserve_crashkernel() - reserves memory for crash kernel
68  *
69  * This function reserves memory area given in "crashkernel=" kernel command
70  * line parameter. The memory reserved is used by dump capture kernel when
71  * primary kernel is crashing.
72  */
73 static void __init reserve_crashkernel(void)
74 {
75 	unsigned long long crash_base, crash_size;
76 	int ret;
77 
78 	ret = parse_crashkernel(boot_command_line, memblock_phys_mem_size(),
79 				&crash_size, &crash_base);
80 	/* no crashkernel= or invalid value specified */
81 	if (ret || !crash_size)
82 		return;
83 
84 	crash_size = PAGE_ALIGN(crash_size);
85 
86 	if (crash_base == 0) {
87 		/* Current arm64 boot protocol requires 2MB alignment */
88 		crash_base = memblock_find_in_range(0, arm64_dma_phys_limit,
89 				crash_size, SZ_2M);
90 		if (crash_base == 0) {
91 			pr_warn("cannot allocate crashkernel (size:0x%llx)\n",
92 				crash_size);
93 			return;
94 		}
95 	} else {
96 		/* User specifies base address explicitly. */
97 		if (!memblock_is_region_memory(crash_base, crash_size)) {
98 			pr_warn("cannot reserve crashkernel: region is not memory\n");
99 			return;
100 		}
101 
102 		if (memblock_is_region_reserved(crash_base, crash_size)) {
103 			pr_warn("cannot reserve crashkernel: region overlaps reserved memory\n");
104 			return;
105 		}
106 
107 		if (!IS_ALIGNED(crash_base, SZ_2M)) {
108 			pr_warn("cannot reserve crashkernel: base address is not 2MB aligned\n");
109 			return;
110 		}
111 	}
112 	memblock_reserve(crash_base, crash_size);
113 
114 	pr_info("crashkernel reserved: 0x%016llx - 0x%016llx (%lld MB)\n",
115 		crash_base, crash_base + crash_size, crash_size >> 20);
116 
117 	crashk_res.start = crash_base;
118 	crashk_res.end = crash_base + crash_size - 1;
119 }
120 #else
121 static void __init reserve_crashkernel(void)
122 {
123 }
124 #endif /* CONFIG_KEXEC_CORE */
125 
126 #ifdef CONFIG_CRASH_DUMP
127 static int __init early_init_dt_scan_elfcorehdr(unsigned long node,
128 		const char *uname, int depth, void *data)
129 {
130 	const __be32 *reg;
131 	int len;
132 
133 	if (depth != 1 || strcmp(uname, "chosen") != 0)
134 		return 0;
135 
136 	reg = of_get_flat_dt_prop(node, "linux,elfcorehdr", &len);
137 	if (!reg || (len < (dt_root_addr_cells + dt_root_size_cells)))
138 		return 1;
139 
140 	elfcorehdr_addr = dt_mem_next_cell(dt_root_addr_cells, &reg);
141 	elfcorehdr_size = dt_mem_next_cell(dt_root_size_cells, &reg);
142 
143 	return 1;
144 }
145 
146 /*
147  * reserve_elfcorehdr() - reserves memory for elf core header
148  *
149  * This function reserves the memory occupied by an elf core header
150  * described in the device tree. This region contains all the
151  * information about primary kernel's core image and is used by a dump
152  * capture kernel to access the system memory on primary kernel.
153  */
154 static void __init reserve_elfcorehdr(void)
155 {
156 	of_scan_flat_dt(early_init_dt_scan_elfcorehdr, NULL);
157 
158 	if (!elfcorehdr_size)
159 		return;
160 
161 	if (memblock_is_region_reserved(elfcorehdr_addr, elfcorehdr_size)) {
162 		pr_warn("elfcorehdr is overlapped\n");
163 		return;
164 	}
165 
166 	memblock_reserve(elfcorehdr_addr, elfcorehdr_size);
167 
168 	pr_info("Reserving %lldKB of memory at 0x%llx for elfcorehdr\n",
169 		elfcorehdr_size >> 10, elfcorehdr_addr);
170 }
171 #else
172 static void __init reserve_elfcorehdr(void)
173 {
174 }
175 #endif /* CONFIG_CRASH_DUMP */
176 
177 /*
178  * Return the maximum physical address for a zone accessible by the given bits
179  * limit. If DRAM starts above 32-bit, expand the zone to the maximum
180  * available memory, otherwise cap it at 32-bit.
181  */
182 static phys_addr_t __init max_zone_phys(unsigned int zone_bits)
183 {
184 	phys_addr_t zone_mask = DMA_BIT_MASK(zone_bits);
185 	phys_addr_t phys_start = memblock_start_of_DRAM();
186 
187 	if (phys_start > U32_MAX)
188 		zone_mask = PHYS_ADDR_MAX;
189 	else if (phys_start > zone_mask)
190 		zone_mask = U32_MAX;
191 
192 	return min(zone_mask, memblock_end_of_DRAM() - 1) + 1;
193 }
194 
195 static void __init zone_sizes_init(unsigned long min, unsigned long max)
196 {
197 	unsigned long max_zone_pfns[MAX_NR_ZONES]  = {0};
198 	unsigned int __maybe_unused acpi_zone_dma_bits;
199 	unsigned int __maybe_unused dt_zone_dma_bits;
200 	phys_addr_t __maybe_unused dma32_phys_limit = max_zone_phys(32);
201 
202 #ifdef CONFIG_ZONE_DMA
203 	acpi_zone_dma_bits = fls64(acpi_iort_dma_get_max_cpu_address());
204 	dt_zone_dma_bits = fls64(of_dma_get_max_cpu_address(NULL));
205 	zone_dma_bits = min3(32U, dt_zone_dma_bits, acpi_zone_dma_bits);
206 	arm64_dma_phys_limit = max_zone_phys(zone_dma_bits);
207 	max_zone_pfns[ZONE_DMA] = PFN_DOWN(arm64_dma_phys_limit);
208 #endif
209 #ifdef CONFIG_ZONE_DMA32
210 	max_zone_pfns[ZONE_DMA32] = PFN_DOWN(dma32_phys_limit);
211 	if (!arm64_dma_phys_limit)
212 		arm64_dma_phys_limit = dma32_phys_limit;
213 #endif
214 	if (!arm64_dma_phys_limit)
215 		arm64_dma_phys_limit = PHYS_MASK + 1;
216 	max_zone_pfns[ZONE_NORMAL] = max;
217 
218 	free_area_init(max_zone_pfns);
219 }
220 
221 int pfn_valid(unsigned long pfn)
222 {
223 	phys_addr_t addr = PFN_PHYS(pfn);
224 
225 	/*
226 	 * Ensure the upper PAGE_SHIFT bits are clear in the
227 	 * pfn. Else it might lead to false positives when
228 	 * some of the upper bits are set, but the lower bits
229 	 * match a valid pfn.
230 	 */
231 	if (PHYS_PFN(addr) != pfn)
232 		return 0;
233 
234 #ifdef CONFIG_SPARSEMEM
235 {
236 	struct mem_section *ms;
237 
238 	if (pfn_to_section_nr(pfn) >= NR_MEM_SECTIONS)
239 		return 0;
240 
241 	ms = __pfn_to_section(pfn);
242 	if (!valid_section(ms))
243 		return 0;
244 
245 	/*
246 	 * ZONE_DEVICE memory does not have the memblock entries.
247 	 * memblock_is_map_memory() check for ZONE_DEVICE based
248 	 * addresses will always fail. Even the normal hotplugged
249 	 * memory will never have MEMBLOCK_NOMAP flag set in their
250 	 * memblock entries. Skip memblock search for all non early
251 	 * memory sections covering all of hotplug memory including
252 	 * both normal and ZONE_DEVICE based.
253 	 */
254 	if (!early_section(ms))
255 		return pfn_section_valid(ms, pfn);
256 }
257 #endif
258 	return memblock_is_map_memory(addr);
259 }
260 EXPORT_SYMBOL(pfn_valid);
261 
262 static phys_addr_t memory_limit = PHYS_ADDR_MAX;
263 
264 /*
265  * Limit the memory size that was specified via FDT.
266  */
267 static int __init early_mem(char *p)
268 {
269 	if (!p)
270 		return 1;
271 
272 	memory_limit = memparse(p, &p) & PAGE_MASK;
273 	pr_notice("Memory limited to %lldMB\n", memory_limit >> 20);
274 
275 	return 0;
276 }
277 early_param("mem", early_mem);
278 
279 static int __init early_init_dt_scan_usablemem(unsigned long node,
280 		const char *uname, int depth, void *data)
281 {
282 	struct memblock_region *usablemem = data;
283 	const __be32 *reg;
284 	int len;
285 
286 	if (depth != 1 || strcmp(uname, "chosen") != 0)
287 		return 0;
288 
289 	reg = of_get_flat_dt_prop(node, "linux,usable-memory-range", &len);
290 	if (!reg || (len < (dt_root_addr_cells + dt_root_size_cells)))
291 		return 1;
292 
293 	usablemem->base = dt_mem_next_cell(dt_root_addr_cells, &reg);
294 	usablemem->size = dt_mem_next_cell(dt_root_size_cells, &reg);
295 
296 	return 1;
297 }
298 
299 static void __init fdt_enforce_memory_region(void)
300 {
301 	struct memblock_region reg = {
302 		.size = 0,
303 	};
304 
305 	of_scan_flat_dt(early_init_dt_scan_usablemem, &reg);
306 
307 	if (reg.size)
308 		memblock_cap_memory_range(reg.base, reg.size);
309 }
310 
311 void __init arm64_memblock_init(void)
312 {
313 	const s64 linear_region_size = PAGE_END - _PAGE_OFFSET(vabits_actual);
314 
315 	/* Handle linux,usable-memory-range property */
316 	fdt_enforce_memory_region();
317 
318 	/* Remove memory above our supported physical address size */
319 	memblock_remove(1ULL << PHYS_MASK_SHIFT, ULLONG_MAX);
320 
321 	/*
322 	 * Select a suitable value for the base of physical memory.
323 	 */
324 	memstart_addr = round_down(memblock_start_of_DRAM(),
325 				   ARM64_MEMSTART_ALIGN);
326 
327 	if ((memblock_end_of_DRAM() - memstart_addr) > linear_region_size)
328 		pr_warn("Memory doesn't fit in the linear mapping, VA_BITS too small\n");
329 
330 	/*
331 	 * Remove the memory that we will not be able to cover with the
332 	 * linear mapping. Take care not to clip the kernel which may be
333 	 * high in memory.
334 	 */
335 	memblock_remove(max_t(u64, memstart_addr + linear_region_size,
336 			__pa_symbol(_end)), ULLONG_MAX);
337 	if (memstart_addr + linear_region_size < memblock_end_of_DRAM()) {
338 		/* ensure that memstart_addr remains sufficiently aligned */
339 		memstart_addr = round_up(memblock_end_of_DRAM() - linear_region_size,
340 					 ARM64_MEMSTART_ALIGN);
341 		memblock_remove(0, memstart_addr);
342 	}
343 
344 	/*
345 	 * If we are running with a 52-bit kernel VA config on a system that
346 	 * does not support it, we have to place the available physical
347 	 * memory in the 48-bit addressable part of the linear region, i.e.,
348 	 * we have to move it upward. Since memstart_addr represents the
349 	 * physical address of PAGE_OFFSET, we have to *subtract* from it.
350 	 */
351 	if (IS_ENABLED(CONFIG_ARM64_VA_BITS_52) && (vabits_actual != 52))
352 		memstart_addr -= _PAGE_OFFSET(48) - _PAGE_OFFSET(52);
353 
354 	/*
355 	 * Apply the memory limit if it was set. Since the kernel may be loaded
356 	 * high up in memory, add back the kernel region that must be accessible
357 	 * via the linear mapping.
358 	 */
359 	if (memory_limit != PHYS_ADDR_MAX) {
360 		memblock_mem_limit_remove_map(memory_limit);
361 		memblock_add(__pa_symbol(_text), (u64)(_end - _text));
362 	}
363 
364 	if (IS_ENABLED(CONFIG_BLK_DEV_INITRD) && phys_initrd_size) {
365 		/*
366 		 * Add back the memory we just removed if it results in the
367 		 * initrd to become inaccessible via the linear mapping.
368 		 * Otherwise, this is a no-op
369 		 */
370 		u64 base = phys_initrd_start & PAGE_MASK;
371 		u64 size = PAGE_ALIGN(phys_initrd_start + phys_initrd_size) - base;
372 
373 		/*
374 		 * We can only add back the initrd memory if we don't end up
375 		 * with more memory than we can address via the linear mapping.
376 		 * It is up to the bootloader to position the kernel and the
377 		 * initrd reasonably close to each other (i.e., within 32 GB of
378 		 * each other) so that all granule/#levels combinations can
379 		 * always access both.
380 		 */
381 		if (WARN(base < memblock_start_of_DRAM() ||
382 			 base + size > memblock_start_of_DRAM() +
383 				       linear_region_size,
384 			"initrd not fully accessible via the linear mapping -- please check your bootloader ...\n")) {
385 			phys_initrd_size = 0;
386 		} else {
387 			memblock_remove(base, size); /* clear MEMBLOCK_ flags */
388 			memblock_add(base, size);
389 			memblock_reserve(base, size);
390 		}
391 	}
392 
393 	if (IS_ENABLED(CONFIG_RANDOMIZE_BASE)) {
394 		extern u16 memstart_offset_seed;
395 		u64 mmfr0 = read_cpuid(ID_AA64MMFR0_EL1);
396 		int parange = cpuid_feature_extract_unsigned_field(
397 					mmfr0, ID_AA64MMFR0_PARANGE_SHIFT);
398 		s64 range = linear_region_size -
399 			    BIT(id_aa64mmfr0_parange_to_phys_shift(parange));
400 
401 		/*
402 		 * If the size of the linear region exceeds, by a sufficient
403 		 * margin, the size of the region that the physical memory can
404 		 * span, randomize the linear region as well.
405 		 */
406 		if (memstart_offset_seed > 0 && range >= (s64)ARM64_MEMSTART_ALIGN) {
407 			range /= ARM64_MEMSTART_ALIGN;
408 			memstart_addr -= ARM64_MEMSTART_ALIGN *
409 					 ((range * memstart_offset_seed) >> 16);
410 		}
411 	}
412 
413 	/*
414 	 * Register the kernel text, kernel data, initrd, and initial
415 	 * pagetables with memblock.
416 	 */
417 	memblock_reserve(__pa_symbol(_stext), _end - _stext);
418 	if (IS_ENABLED(CONFIG_BLK_DEV_INITRD) && phys_initrd_size) {
419 		/* the generic initrd code expects virtual addresses */
420 		initrd_start = __phys_to_virt(phys_initrd_start);
421 		initrd_end = initrd_start + phys_initrd_size;
422 	}
423 
424 	early_init_fdt_scan_reserved_mem();
425 
426 	reserve_elfcorehdr();
427 
428 	high_memory = __va(memblock_end_of_DRAM() - 1) + 1;
429 }
430 
431 void __init bootmem_init(void)
432 {
433 	unsigned long min, max;
434 
435 	min = PFN_UP(memblock_start_of_DRAM());
436 	max = PFN_DOWN(memblock_end_of_DRAM());
437 
438 	early_memtest(min << PAGE_SHIFT, max << PAGE_SHIFT);
439 
440 	max_pfn = max_low_pfn = max;
441 	min_low_pfn = min;
442 
443 	arch_numa_init();
444 
445 	/*
446 	 * must be done after arch_numa_init() which calls numa_init() to
447 	 * initialize node_online_map that gets used in hugetlb_cma_reserve()
448 	 * while allocating required CMA size across online nodes.
449 	 */
450 #if defined(CONFIG_HUGETLB_PAGE) && defined(CONFIG_CMA)
451 	arm64_hugetlb_cma_reserve();
452 #endif
453 
454 	dma_pernuma_cma_reserve();
455 
456 	kvm_hyp_reserve();
457 
458 	/*
459 	 * sparse_init() tries to allocate memory from memblock, so must be
460 	 * done after the fixed reservations
461 	 */
462 	sparse_init();
463 	zone_sizes_init(min, max);
464 
465 	/*
466 	 * Reserve the CMA area after arm64_dma_phys_limit was initialised.
467 	 */
468 	dma_contiguous_reserve(arm64_dma_phys_limit);
469 
470 	/*
471 	 * request_standard_resources() depends on crashkernel's memory being
472 	 * reserved, so do it here.
473 	 */
474 	reserve_crashkernel();
475 
476 	memblock_dump_all();
477 }
478 
479 /*
480  * mem_init() marks the free areas in the mem_map and tells us how much memory
481  * is free.  This is done after various parts of the system have claimed their
482  * memory after the kernel image.
483  */
484 void __init mem_init(void)
485 {
486 	if (swiotlb_force == SWIOTLB_FORCE ||
487 	    max_pfn > PFN_DOWN(arm64_dma_phys_limit))
488 		swiotlb_init(1);
489 	else
490 		swiotlb_force = SWIOTLB_NO_FORCE;
491 
492 	set_max_mapnr(max_pfn - PHYS_PFN_OFFSET);
493 
494 	/* this will put all unused low memory onto the freelists */
495 	memblock_free_all();
496 
497 	/*
498 	 * Check boundaries twice: Some fundamental inconsistencies can be
499 	 * detected at build time already.
500 	 */
501 #ifdef CONFIG_COMPAT
502 	BUILD_BUG_ON(TASK_SIZE_32 > DEFAULT_MAP_WINDOW_64);
503 #endif
504 
505 	if (PAGE_SIZE >= 16384 && get_num_physpages() <= 128) {
506 		extern int sysctl_overcommit_memory;
507 		/*
508 		 * On a machine this small we won't get anywhere without
509 		 * overcommit, so turn it on by default.
510 		 */
511 		sysctl_overcommit_memory = OVERCOMMIT_ALWAYS;
512 	}
513 }
514 
515 void free_initmem(void)
516 {
517 	free_reserved_area(lm_alias(__init_begin),
518 			   lm_alias(__init_end),
519 			   POISON_FREE_INITMEM, "unused kernel");
520 	/*
521 	 * Unmap the __init region but leave the VM area in place. This
522 	 * prevents the region from being reused for kernel modules, which
523 	 * is not supported by kallsyms.
524 	 */
525 	vunmap_range((u64)__init_begin, (u64)__init_end);
526 }
527 
528 void dump_mem_limit(void)
529 {
530 	if (memory_limit != PHYS_ADDR_MAX) {
531 		pr_emerg("Memory Limit: %llu MB\n", memory_limit >> 20);
532 	} else {
533 		pr_emerg("Memory Limit: none\n");
534 	}
535 }
536