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/memory.h>
39 #include <asm/numa.h>
40 #include <asm/sections.h>
41 #include <asm/setup.h>
42 #include <linux/sizes.h>
43 #include <asm/tlb.h>
44 #include <asm/alternative.h>
45
46 /*
47 * We need to be able to catch inadvertent references to memstart_addr
48 * that occur (potentially in generic code) before arm64_memblock_init()
49 * executes, which assigns it its actual value. So use a default value
50 * that cannot be mistaken for a real physical address.
51 */
52 s64 memstart_addr __ro_after_init = -1;
53 EXPORT_SYMBOL(memstart_addr);
54
55 /*
56 * If the corresponding config options are enabled, we create both ZONE_DMA
57 * and ZONE_DMA32. By default ZONE_DMA covers the 32-bit addressable memory
58 * unless restricted on specific platforms (e.g. 30-bit on Raspberry Pi 4).
59 * In such case, ZONE_DMA32 covers the rest of the 32-bit addressable memory,
60 * otherwise it is empty.
61 *
62 * Memory reservation for crash kernel either done early or deferred
63 * depending on DMA memory zones configs (ZONE_DMA) --
64 *
65 * In absence of ZONE_DMA configs arm64_dma_phys_limit initialized
66 * here instead of max_zone_phys(). This lets early reservation of
67 * crash kernel memory which has a dependency on arm64_dma_phys_limit.
68 * Reserving memory early for crash kernel allows linear creation of block
69 * mappings (greater than page-granularity) for all the memory bank rangs.
70 * In this scheme a comparatively quicker boot is observed.
71 *
72 * If ZONE_DMA configs are defined, crash kernel memory reservation
73 * is delayed until DMA zone memory range size initilazation performed in
74 * zone_sizes_init(). The defer is necessary to steer clear of DMA zone
75 * memory range to avoid overlap allocation. So crash kernel memory boundaries
76 * are not known when mapping all bank memory ranges, which otherwise means
77 * not possible to exclude crash kernel range from creating block mappings
78 * so page-granularity mappings are created for the entire memory range.
79 * Hence a slightly slower boot is observed.
80 *
81 * Note: Page-granularity mapppings are necessary for crash kernel memory
82 * range for shrinking its size via /sys/kernel/kexec_crash_size interface.
83 */
84 #if IS_ENABLED(CONFIG_ZONE_DMA) || IS_ENABLED(CONFIG_ZONE_DMA32)
85 phys_addr_t __ro_after_init arm64_dma_phys_limit;
86 #else
87 phys_addr_t __ro_after_init arm64_dma_phys_limit = PHYS_MASK + 1;
88 #endif
89
90 #ifdef CONFIG_KEXEC_CORE
91 /*
92 * reserve_crashkernel() - reserves memory for crash kernel
93 *
94 * This function reserves memory area given in "crashkernel=" kernel command
95 * line parameter. The memory reserved is used by dump capture kernel when
96 * primary kernel is crashing.
97 */
reserve_crashkernel(void)98 static void __init reserve_crashkernel(void)
99 {
100 unsigned long long crash_base, crash_size;
101 int ret;
102
103 ret = parse_crashkernel(boot_command_line, memblock_phys_mem_size(),
104 &crash_size, &crash_base);
105 /* no crashkernel= or invalid value specified */
106 if (ret || !crash_size)
107 return;
108
109 crash_size = PAGE_ALIGN(crash_size);
110
111 if (crash_base == 0) {
112 /* Current arm64 boot protocol requires 2MB alignment */
113 crash_base = memblock_find_in_range(0, arm64_dma_phys_limit,
114 crash_size, SZ_2M);
115 if (crash_base == 0) {
116 pr_warn("cannot allocate crashkernel (size:0x%llx)\n",
117 crash_size);
118 return;
119 }
120 } else {
121 /* User specifies base address explicitly. */
122 if (!memblock_is_region_memory(crash_base, crash_size)) {
123 pr_warn("cannot reserve crashkernel: region is not memory\n");
124 return;
125 }
126
127 if (memblock_is_region_reserved(crash_base, crash_size)) {
128 pr_warn("cannot reserve crashkernel: region overlaps reserved memory\n");
129 return;
130 }
131
132 if (!IS_ALIGNED(crash_base, SZ_2M)) {
133 pr_warn("cannot reserve crashkernel: base address is not 2MB aligned\n");
134 return;
135 }
136 }
137 memblock_reserve(crash_base, crash_size);
138
139 pr_info("crashkernel reserved: 0x%016llx - 0x%016llx (%lld MB)\n",
140 crash_base, crash_base + crash_size, crash_size >> 20);
141
142 crashk_res.start = crash_base;
143 crashk_res.end = crash_base + crash_size - 1;
144 }
145 #else
reserve_crashkernel(void)146 static void __init reserve_crashkernel(void)
147 {
148 }
149 #endif /* CONFIG_KEXEC_CORE */
150
151 #ifdef CONFIG_CRASH_DUMP
early_init_dt_scan_elfcorehdr(unsigned long node,const char * uname,int depth,void * data)152 static int __init early_init_dt_scan_elfcorehdr(unsigned long node,
153 const char *uname, int depth, void *data)
154 {
155 const __be32 *reg;
156 int len;
157
158 if (depth != 1 || strcmp(uname, "chosen") != 0)
159 return 0;
160
161 reg = of_get_flat_dt_prop(node, "linux,elfcorehdr", &len);
162 if (!reg || (len < (dt_root_addr_cells + dt_root_size_cells)))
163 return 1;
164
165 elfcorehdr_addr = dt_mem_next_cell(dt_root_addr_cells, ®);
166 elfcorehdr_size = dt_mem_next_cell(dt_root_size_cells, ®);
167
168 return 1;
169 }
170
171 /*
172 * reserve_elfcorehdr() - reserves memory for elf core header
173 *
174 * This function reserves the memory occupied by an elf core header
175 * described in the device tree. This region contains all the
176 * information about primary kernel's core image and is used by a dump
177 * capture kernel to access the system memory on primary kernel.
178 */
reserve_elfcorehdr(void)179 static void __init reserve_elfcorehdr(void)
180 {
181 of_scan_flat_dt(early_init_dt_scan_elfcorehdr, NULL);
182
183 if (!elfcorehdr_size)
184 return;
185
186 if (memblock_is_region_reserved(elfcorehdr_addr, elfcorehdr_size)) {
187 pr_warn("elfcorehdr is overlapped\n");
188 return;
189 }
190
191 memblock_reserve(elfcorehdr_addr, elfcorehdr_size);
192
193 pr_info("Reserving %lldKB of memory at 0x%llx for elfcorehdr\n",
194 elfcorehdr_size >> 10, elfcorehdr_addr);
195 }
196 #else
reserve_elfcorehdr(void)197 static void __init reserve_elfcorehdr(void)
198 {
199 }
200 #endif /* CONFIG_CRASH_DUMP */
201
202 /*
203 * Return the maximum physical address for a zone accessible by the given bits
204 * limit. If DRAM starts above 32-bit, expand the zone to the maximum
205 * available memory, otherwise cap it at 32-bit.
206 */
max_zone_phys(unsigned int zone_bits)207 static phys_addr_t __init max_zone_phys(unsigned int zone_bits)
208 {
209 phys_addr_t zone_mask = DMA_BIT_MASK(zone_bits);
210 phys_addr_t phys_start = memblock_start_of_DRAM();
211
212 if (phys_start > U32_MAX)
213 zone_mask = PHYS_ADDR_MAX;
214 else if (phys_start > zone_mask)
215 zone_mask = U32_MAX;
216
217 return min(zone_mask, memblock_end_of_DRAM() - 1) + 1;
218 }
219
zone_sizes_init(unsigned long min,unsigned long max)220 static void __init zone_sizes_init(unsigned long min, unsigned long max)
221 {
222 unsigned long max_zone_pfns[MAX_NR_ZONES] = {0};
223 unsigned int __maybe_unused acpi_zone_dma_bits;
224 unsigned int __maybe_unused dt_zone_dma_bits;
225 phys_addr_t __maybe_unused dma32_phys_limit = max_zone_phys(32);
226
227 #ifdef CONFIG_ZONE_DMA
228 acpi_zone_dma_bits = fls64(acpi_iort_dma_get_max_cpu_address());
229 dt_zone_dma_bits = fls64(of_dma_get_max_cpu_address(NULL));
230 zone_dma_bits = min3(32U, dt_zone_dma_bits, acpi_zone_dma_bits);
231 arm64_dma_phys_limit = max_zone_phys(zone_dma_bits);
232 max_zone_pfns[ZONE_DMA] = PFN_DOWN(arm64_dma_phys_limit);
233 #endif
234 #ifdef CONFIG_ZONE_DMA32
235 max_zone_pfns[ZONE_DMA32] = PFN_DOWN(dma32_phys_limit);
236 if (!arm64_dma_phys_limit)
237 arm64_dma_phys_limit = dma32_phys_limit;
238 #endif
239 max_zone_pfns[ZONE_NORMAL] = max;
240
241 free_area_init(max_zone_pfns);
242 }
243
pfn_valid(unsigned long pfn)244 int pfn_valid(unsigned long pfn)
245 {
246 phys_addr_t addr = pfn << PAGE_SHIFT;
247
248 if ((addr >> PAGE_SHIFT) != pfn)
249 return 0;
250
251 #ifdef CONFIG_SPARSEMEM
252 if (pfn_to_section_nr(pfn) >= NR_MEM_SECTIONS)
253 return 0;
254
255 if (!valid_section(__pfn_to_section(pfn)))
256 return 0;
257
258 /*
259 * ZONE_DEVICE memory does not have the memblock entries.
260 * memblock_is_map_memory() check for ZONE_DEVICE based
261 * addresses will always fail. Even the normal hotplugged
262 * memory will never have MEMBLOCK_NOMAP flag set in their
263 * memblock entries. Skip memblock search for all non early
264 * memory sections covering all of hotplug memory including
265 * both normal and ZONE_DEVICE based.
266 */
267 if (!early_section(__pfn_to_section(pfn)))
268 return pfn_section_valid(__pfn_to_section(pfn), pfn);
269 #endif
270 return memblock_is_map_memory(addr);
271 }
272 EXPORT_SYMBOL(pfn_valid);
273
274 static phys_addr_t memory_limit = PHYS_ADDR_MAX;
275
276 /*
277 * Limit the memory size that was specified via FDT.
278 */
early_mem(char * p)279 static int __init early_mem(char *p)
280 {
281 if (!p)
282 return 1;
283
284 memory_limit = memparse(p, &p) & PAGE_MASK;
285 pr_notice("Memory limited to %lldMB\n", memory_limit >> 20);
286
287 return 0;
288 }
289 early_param("mem", early_mem);
290
early_init_dt_scan_usablemem(unsigned long node,const char * uname,int depth,void * data)291 static int __init early_init_dt_scan_usablemem(unsigned long node,
292 const char *uname, int depth, void *data)
293 {
294 struct memblock_region *usablemem = data;
295 const __be32 *reg;
296 int len;
297
298 if (depth != 1 || strcmp(uname, "chosen") != 0)
299 return 0;
300
301 reg = of_get_flat_dt_prop(node, "linux,usable-memory-range", &len);
302 if (!reg || (len < (dt_root_addr_cells + dt_root_size_cells)))
303 return 1;
304
305 usablemem->base = dt_mem_next_cell(dt_root_addr_cells, ®);
306 usablemem->size = dt_mem_next_cell(dt_root_size_cells, ®);
307
308 return 1;
309 }
310
fdt_enforce_memory_region(void)311 static void __init fdt_enforce_memory_region(void)
312 {
313 struct memblock_region reg = {
314 .size = 0,
315 };
316
317 of_scan_flat_dt(early_init_dt_scan_usablemem, ®);
318
319 if (reg.size)
320 memblock_cap_memory_range(reg.base, reg.size);
321 }
322
arm64_memblock_init(void)323 void __init arm64_memblock_init(void)
324 {
325 const s64 linear_region_size = BIT(vabits_actual - 1);
326
327 /* Handle linux,usable-memory-range property */
328 fdt_enforce_memory_region();
329
330 /* Remove memory above our supported physical address size */
331 memblock_remove(1ULL << PHYS_MASK_SHIFT, ULLONG_MAX);
332
333 /*
334 * Select a suitable value for the base of physical memory.
335 */
336 memstart_addr = round_down(memblock_start_of_DRAM(),
337 ARM64_MEMSTART_ALIGN);
338
339 /*
340 * Remove the memory that we will not be able to cover with the
341 * linear mapping. Take care not to clip the kernel which may be
342 * high in memory.
343 */
344 memblock_remove(max_t(u64, memstart_addr + linear_region_size,
345 __pa_symbol(_end)), ULLONG_MAX);
346 if (memstart_addr + linear_region_size < memblock_end_of_DRAM()) {
347 /* ensure that memstart_addr remains sufficiently aligned */
348 memstart_addr = round_up(memblock_end_of_DRAM() - linear_region_size,
349 ARM64_MEMSTART_ALIGN);
350 memblock_remove(0, memstart_addr);
351 }
352
353 /*
354 * If we are running with a 52-bit kernel VA config on a system that
355 * does not support it, we have to place the available physical
356 * memory in the 48-bit addressable part of the linear region, i.e.,
357 * we have to move it upward. Since memstart_addr represents the
358 * physical address of PAGE_OFFSET, we have to *subtract* from it.
359 */
360 if (IS_ENABLED(CONFIG_ARM64_VA_BITS_52) && (vabits_actual != 52))
361 memstart_addr -= _PAGE_OFFSET(48) - _PAGE_OFFSET(52);
362
363 /*
364 * Apply the memory limit if it was set. Since the kernel may be loaded
365 * high up in memory, add back the kernel region that must be accessible
366 * via the linear mapping.
367 */
368 if (memory_limit != PHYS_ADDR_MAX) {
369 memblock_mem_limit_remove_map(memory_limit);
370 memblock_add(__pa_symbol(_text), (u64)(_end - _text));
371 }
372
373 if (IS_ENABLED(CONFIG_BLK_DEV_INITRD) && phys_initrd_size) {
374 /*
375 * Add back the memory we just removed if it results in the
376 * initrd to become inaccessible via the linear mapping.
377 * Otherwise, this is a no-op
378 */
379 u64 base = phys_initrd_start & PAGE_MASK;
380 u64 size = PAGE_ALIGN(phys_initrd_start + phys_initrd_size) - base;
381
382 /*
383 * We can only add back the initrd memory if we don't end up
384 * with more memory than we can address via the linear mapping.
385 * It is up to the bootloader to position the kernel and the
386 * initrd reasonably close to each other (i.e., within 32 GB of
387 * each other) so that all granule/#levels combinations can
388 * always access both.
389 */
390 if (WARN(base < memblock_start_of_DRAM() ||
391 base + size > memblock_start_of_DRAM() +
392 linear_region_size,
393 "initrd not fully accessible via the linear mapping -- please check your bootloader ...\n")) {
394 phys_initrd_size = 0;
395 } else {
396 memblock_remove(base, size); /* clear MEMBLOCK_ flags */
397 memblock_add(base, size);
398 memblock_reserve(base, size);
399 }
400 }
401
402 if (IS_ENABLED(CONFIG_RANDOMIZE_BASE)) {
403 extern u16 memstart_offset_seed;
404 u64 range = linear_region_size -
405 (memblock_end_of_DRAM() - memblock_start_of_DRAM());
406
407 /*
408 * If the size of the linear region exceeds, by a sufficient
409 * margin, the size of the region that the available physical
410 * memory spans, randomize the linear region as well.
411 */
412 if (memstart_offset_seed > 0 && range >= ARM64_MEMSTART_ALIGN) {
413 range /= ARM64_MEMSTART_ALIGN;
414 memstart_addr -= ARM64_MEMSTART_ALIGN *
415 ((range * memstart_offset_seed) >> 16);
416 }
417 }
418
419 /*
420 * Register the kernel text, kernel data, initrd, and initial
421 * pagetables with memblock.
422 */
423 memblock_reserve(__pa_symbol(_text), _end - _text);
424 if (IS_ENABLED(CONFIG_BLK_DEV_INITRD) && phys_initrd_size) {
425 /* the generic initrd code expects virtual addresses */
426 initrd_start = __phys_to_virt(phys_initrd_start);
427 initrd_end = initrd_start + phys_initrd_size;
428 }
429
430 early_init_fdt_scan_reserved_mem();
431
432 reserve_elfcorehdr();
433
434 if (!IS_ENABLED(CONFIG_ZONE_DMA) && !IS_ENABLED(CONFIG_ZONE_DMA32))
435 reserve_crashkernel();
436
437 high_memory = __va(memblock_end_of_DRAM() - 1) + 1;
438 }
439
bootmem_init(void)440 void __init bootmem_init(void)
441 {
442 unsigned long min, max;
443
444 min = PFN_UP(memblock_start_of_DRAM());
445 max = PFN_DOWN(memblock_end_of_DRAM());
446
447 early_memtest(min << PAGE_SHIFT, max << PAGE_SHIFT);
448
449 max_pfn = max_low_pfn = max;
450 min_low_pfn = min;
451
452 arm64_numa_init();
453
454 /*
455 * must be done after arm64_numa_init() which calls numa_init() to
456 * initialize node_online_map that gets used in hugetlb_cma_reserve()
457 * while allocating required CMA size across online nodes.
458 */
459 #if defined(CONFIG_HUGETLB_PAGE) && defined(CONFIG_CMA)
460 arm64_hugetlb_cma_reserve();
461 #endif
462
463 dma_pernuma_cma_reserve();
464
465 /*
466 * sparse_init() tries to allocate memory from memblock, so must be
467 * done after the fixed reservations
468 */
469 sparse_init();
470 zone_sizes_init(min, max);
471
472 /*
473 * Reserve the CMA area after arm64_dma_phys_limit was initialised.
474 */
475 dma_contiguous_reserve(arm64_dma_phys_limit);
476
477 /*
478 * request_standard_resources() depends on crashkernel's memory being
479 * reserved, so do it here.
480 */
481 if (IS_ENABLED(CONFIG_ZONE_DMA) || IS_ENABLED(CONFIG_ZONE_DMA32))
482 reserve_crashkernel();
483
484 memblock_dump_all();
485 }
486
487 #ifndef CONFIG_SPARSEMEM_VMEMMAP
free_memmap(unsigned long start_pfn,unsigned long end_pfn)488 static inline void free_memmap(unsigned long start_pfn, unsigned long end_pfn)
489 {
490 struct page *start_pg, *end_pg;
491 unsigned long pg, pgend;
492
493 /*
494 * Convert start_pfn/end_pfn to a struct page pointer.
495 */
496 start_pg = pfn_to_page(start_pfn - 1) + 1;
497 end_pg = pfn_to_page(end_pfn - 1) + 1;
498
499 /*
500 * Convert to physical addresses, and round start upwards and end
501 * downwards.
502 */
503 pg = (unsigned long)PAGE_ALIGN(__pa(start_pg));
504 pgend = (unsigned long)__pa(end_pg) & PAGE_MASK;
505
506 /*
507 * If there are free pages between these, free the section of the
508 * memmap array.
509 */
510 if (pg < pgend)
511 memblock_free(pg, pgend - pg);
512 }
513
514 /*
515 * The mem_map array can get very big. Free the unused area of the memory map.
516 */
free_unused_memmap(void)517 static void __init free_unused_memmap(void)
518 {
519 unsigned long start, end, prev_end = 0;
520 int i;
521
522 for_each_mem_pfn_range(i, MAX_NUMNODES, &start, &end, NULL) {
523 #ifdef CONFIG_SPARSEMEM
524 /*
525 * Take care not to free memmap entries that don't exist due
526 * to SPARSEMEM sections which aren't present.
527 */
528 start = min(start, ALIGN(prev_end, PAGES_PER_SECTION));
529 #endif
530 /*
531 * If we had a previous bank, and there is a space between the
532 * current bank and the previous, free it.
533 */
534 if (prev_end && prev_end < start)
535 free_memmap(prev_end, start);
536
537 /*
538 * Align up here since the VM subsystem insists that the
539 * memmap entries are valid from the bank end aligned to
540 * MAX_ORDER_NR_PAGES.
541 */
542 prev_end = ALIGN(end, MAX_ORDER_NR_PAGES);
543 }
544
545 #ifdef CONFIG_SPARSEMEM
546 if (!IS_ALIGNED(prev_end, PAGES_PER_SECTION))
547 free_memmap(prev_end, ALIGN(prev_end, PAGES_PER_SECTION));
548 #endif
549 }
550 #endif /* !CONFIG_SPARSEMEM_VMEMMAP */
551
552 /*
553 * mem_init() marks the free areas in the mem_map and tells us how much memory
554 * is free. This is done after various parts of the system have claimed their
555 * memory after the kernel image.
556 */
mem_init(void)557 void __init mem_init(void)
558 {
559 if (swiotlb_force == SWIOTLB_FORCE ||
560 max_pfn > PFN_DOWN(arm64_dma_phys_limit))
561 swiotlb_init(1);
562 else
563 swiotlb_force = SWIOTLB_NO_FORCE;
564
565 set_max_mapnr(max_pfn - PHYS_PFN_OFFSET);
566
567 #ifndef CONFIG_SPARSEMEM_VMEMMAP
568 free_unused_memmap();
569 #endif
570 /* this will put all unused low memory onto the freelists */
571 memblock_free_all();
572
573 mem_init_print_info(NULL);
574
575 /*
576 * Check boundaries twice: Some fundamental inconsistencies can be
577 * detected at build time already.
578 */
579 #ifdef CONFIG_COMPAT
580 BUILD_BUG_ON(TASK_SIZE_32 > DEFAULT_MAP_WINDOW_64);
581 #endif
582
583 if (PAGE_SIZE >= 16384 && get_num_physpages() <= 128) {
584 extern int sysctl_overcommit_memory;
585 /*
586 * On a machine this small we won't get anywhere without
587 * overcommit, so turn it on by default.
588 */
589 sysctl_overcommit_memory = OVERCOMMIT_ALWAYS;
590 }
591 }
592
free_initmem(void)593 void free_initmem(void)
594 {
595 free_reserved_area(lm_alias(__init_begin),
596 lm_alias(__init_end),
597 POISON_FREE_INITMEM, "unused kernel");
598 /*
599 * Unmap the __init region but leave the VM area in place. This
600 * prevents the region from being reused for kernel modules, which
601 * is not supported by kallsyms.
602 */
603 unmap_kernel_range((u64)__init_begin, (u64)(__init_end - __init_begin));
604 }
605
dump_mem_limit(void)606 void dump_mem_limit(void)
607 {
608 if (memory_limit != PHYS_ADDR_MAX) {
609 pr_emerg("Memory Limit: %llu MB\n", memory_limit >> 20);
610 } else {
611 pr_emerg("Memory Limit: none\n");
612 }
613 }
614