2 * Procedures for maintaining information about logical memory blocks.
4 * Peter Bergner, IBM Corp. June 2001.
5 * Copyright (C) 2001 Peter Bergner.
7 * This program is free software; you can redistribute it and/or
8 * modify it under the terms of the GNU General Public License
9 * as published by the Free Software Foundation; either version
10 * 2 of the License, or (at your option) any later version.
13 #include <linux/kernel.h>
14 #include <linux/slab.h>
15 #include <linux/init.h>
16 #include <linux/bitops.h>
17 #include <linux/poison.h>
18 #include <linux/pfn.h>
19 #include <linux/debugfs.h>
20 #include <linux/seq_file.h>
21 #include <linux/memblock.h>
23 struct memblock memblock __initdata_memblock;
25 int memblock_debug __initdata_memblock;
26 int memblock_can_resize __initdata_memblock;
27 static struct memblock_region memblock_memory_init_regions[INIT_MEMBLOCK_REGIONS + 1] __initdata_memblock;
28 static struct memblock_region memblock_reserved_init_regions[INIT_MEMBLOCK_REGIONS + 1] __initdata_memblock;
30 /* inline so we don't get a warning when pr_debug is compiled out */
31 static inline const char *memblock_type_name(struct memblock_type *type)
33 if (type == &memblock.memory)
35 else if (type == &memblock.reserved)
42 * Address comparison utilities
45 static phys_addr_t __init_memblock memblock_align_down(phys_addr_t addr, phys_addr_t size)
47 return addr & ~(size - 1);
50 static phys_addr_t __init_memblock memblock_align_up(phys_addr_t addr, phys_addr_t size)
52 return (addr + (size - 1)) & ~(size - 1);
55 static unsigned long __init_memblock memblock_addrs_overlap(phys_addr_t base1, phys_addr_t size1,
56 phys_addr_t base2, phys_addr_t size2)
58 return ((base1 < (base2 + size2)) && (base2 < (base1 + size1)));
61 static long __init_memblock memblock_addrs_adjacent(phys_addr_t base1, phys_addr_t size1,
62 phys_addr_t base2, phys_addr_t size2)
64 if (base2 == base1 + size1)
66 else if (base1 == base2 + size2)
72 static long __init_memblock memblock_regions_adjacent(struct memblock_type *type,
73 unsigned long r1, unsigned long r2)
75 phys_addr_t base1 = type->regions[r1].base;
76 phys_addr_t size1 = type->regions[r1].size;
77 phys_addr_t base2 = type->regions[r2].base;
78 phys_addr_t size2 = type->regions[r2].size;
80 return memblock_addrs_adjacent(base1, size1, base2, size2);
83 long __init_memblock memblock_overlaps_region(struct memblock_type *type, phys_addr_t base, phys_addr_t size)
87 for (i = 0; i < type->cnt; i++) {
88 phys_addr_t rgnbase = type->regions[i].base;
89 phys_addr_t rgnsize = type->regions[i].size;
90 if (memblock_addrs_overlap(base, size, rgnbase, rgnsize))
94 return (i < type->cnt) ? i : -1;
98 * Find, allocate, deallocate or reserve unreserved regions. All allocations
102 static phys_addr_t __init memblock_find_region(phys_addr_t start, phys_addr_t end,
103 phys_addr_t size, phys_addr_t align)
105 phys_addr_t base, res_base;
108 /* Prevent allocations returning 0 as it's also used to
109 * indicate an allocation failure
114 base = memblock_align_down((end - size), align);
115 while (start <= base) {
116 j = memblock_overlaps_region(&memblock.reserved, base, size);
119 res_base = memblock.reserved.regions[j].base;
122 base = memblock_align_down(res_base - size, align);
125 return MEMBLOCK_ERROR;
128 static phys_addr_t __init_memblock memblock_find_base(phys_addr_t size,
129 phys_addr_t align, phys_addr_t start, phys_addr_t end)
135 size = memblock_align_up(size, align);
137 /* Pump up max_addr */
138 if (end == MEMBLOCK_ALLOC_ACCESSIBLE)
139 end = memblock.current_limit;
141 /* We do a top-down search, this tends to limit memory
142 * fragmentation by keeping early boot allocs near the
145 for (i = memblock.memory.cnt - 1; i >= 0; i--) {
146 phys_addr_t memblockbase = memblock.memory.regions[i].base;
147 phys_addr_t memblocksize = memblock.memory.regions[i].size;
148 phys_addr_t bottom, top, found;
150 if (memblocksize < size)
152 if ((memblockbase + memblocksize) <= start)
154 bottom = max(memblockbase, start);
155 top = min(memblockbase + memblocksize, end);
158 found = memblock_find_region(bottom, top, size, align);
159 if (found != MEMBLOCK_ERROR)
162 return MEMBLOCK_ERROR;
166 * Find a free area with specified alignment in a specific range.
168 u64 __init_memblock memblock_find_in_range(u64 start, u64 end, u64 size, u64 align)
170 return memblock_find_base(size, align, start, end);
174 * Free memblock.reserved.regions
176 int __init_memblock memblock_free_reserved_regions(void)
178 if (memblock.reserved.regions == memblock_reserved_init_regions)
181 return memblock_free(__pa(memblock.reserved.regions),
182 sizeof(struct memblock_region) * memblock.reserved.max);
186 * Reserve memblock.reserved.regions
188 int __init_memblock memblock_reserve_reserved_regions(void)
190 if (memblock.reserved.regions == memblock_reserved_init_regions)
193 return memblock_reserve(__pa(memblock.reserved.regions),
194 sizeof(struct memblock_region) * memblock.reserved.max);
197 static void __init_memblock memblock_remove_region(struct memblock_type *type, unsigned long r)
201 for (i = r; i < type->cnt - 1; i++) {
202 type->regions[i].base = type->regions[i + 1].base;
203 type->regions[i].size = type->regions[i + 1].size;
208 /* Assumption: base addr of region 1 < base addr of region 2 */
209 static void __init_memblock memblock_coalesce_regions(struct memblock_type *type,
210 unsigned long r1, unsigned long r2)
212 type->regions[r1].size += type->regions[r2].size;
213 memblock_remove_region(type, r2);
216 /* Defined below but needed now */
217 static long memblock_add_region(struct memblock_type *type, phys_addr_t base, phys_addr_t size);
219 static int __init_memblock memblock_double_array(struct memblock_type *type)
221 struct memblock_region *new_array, *old_array;
222 phys_addr_t old_size, new_size, addr;
223 int use_slab = slab_is_available();
225 /* We don't allow resizing until we know about the reserved regions
226 * of memory that aren't suitable for allocation
228 if (!memblock_can_resize)
231 /* Calculate new doubled size */
232 old_size = type->max * sizeof(struct memblock_region);
233 new_size = old_size << 1;
235 /* Try to find some space for it.
237 * WARNING: We assume that either slab_is_available() and we use it or
238 * we use MEMBLOCK for allocations. That means that this is unsafe to use
239 * when bootmem is currently active (unless bootmem itself is implemented
240 * on top of MEMBLOCK which isn't the case yet)
242 * This should however not be an issue for now, as we currently only
243 * call into MEMBLOCK while it's still active, or much later when slab is
244 * active for memory hotplug operations
247 new_array = kmalloc(new_size, GFP_KERNEL);
248 addr = new_array == NULL ? MEMBLOCK_ERROR : __pa(new_array);
250 addr = memblock_find_base(new_size, sizeof(phys_addr_t), 0, MEMBLOCK_ALLOC_ACCESSIBLE);
251 if (addr == MEMBLOCK_ERROR) {
252 pr_err("memblock: Failed to double %s array from %ld to %ld entries !\n",
253 memblock_type_name(type), type->max, type->max * 2);
256 new_array = __va(addr);
258 memblock_dbg("memblock: %s array is doubled to %ld at [%#010llx-%#010llx]",
259 memblock_type_name(type), type->max * 2, (u64)addr, (u64)addr + new_size - 1);
261 /* Found space, we now need to move the array over before
262 * we add the reserved region since it may be our reserved
263 * array itself that is full.
265 memcpy(new_array, type->regions, old_size);
266 memset(new_array + type->max, 0, old_size);
267 old_array = type->regions;
268 type->regions = new_array;
271 /* If we use SLAB that's it, we are done */
275 /* Add the new reserved region now. Should not fail ! */
276 BUG_ON(memblock_add_region(&memblock.reserved, addr, new_size) < 0);
278 /* If the array wasn't our static init one, then free it. We only do
279 * that before SLAB is available as later on, we don't know whether
280 * to use kfree or free_bootmem_pages(). Shouldn't be a big deal
283 if (old_array != memblock_memory_init_regions &&
284 old_array != memblock_reserved_init_regions)
285 memblock_free(__pa(old_array), old_size);
290 extern int __init_memblock __weak memblock_memory_can_coalesce(phys_addr_t addr1, phys_addr_t size1,
291 phys_addr_t addr2, phys_addr_t size2)
296 static long __init_memblock memblock_add_region(struct memblock_type *type, phys_addr_t base, phys_addr_t size)
298 unsigned long coalesced = 0;
301 if ((type->cnt == 1) && (type->regions[0].size == 0)) {
302 type->regions[0].base = base;
303 type->regions[0].size = size;
307 /* First try and coalesce this MEMBLOCK with another. */
308 for (i = 0; i < type->cnt; i++) {
309 phys_addr_t rgnbase = type->regions[i].base;
310 phys_addr_t rgnsize = type->regions[i].size;
312 if ((rgnbase == base) && (rgnsize == size))
313 /* Already have this region, so we're done */
316 adjacent = memblock_addrs_adjacent(base, size, rgnbase, rgnsize);
317 /* Check if arch allows coalescing */
318 if (adjacent != 0 && type == &memblock.memory &&
319 !memblock_memory_can_coalesce(base, size, rgnbase, rgnsize))
322 type->regions[i].base -= size;
323 type->regions[i].size += size;
326 } else if (adjacent < 0) {
327 type->regions[i].size += size;
333 /* If we plugged a hole, we may want to also coalesce with the
336 if ((i < type->cnt - 1) && memblock_regions_adjacent(type, i, i+1) &&
337 ((type != &memblock.memory || memblock_memory_can_coalesce(type->regions[i].base,
338 type->regions[i].size,
339 type->regions[i+1].base,
340 type->regions[i+1].size)))) {
341 memblock_coalesce_regions(type, i, i+1);
348 /* If we are out of space, we fail. It's too late to resize the array
349 * but then this shouldn't have happened in the first place.
351 if (WARN_ON(type->cnt >= type->max))
354 /* Couldn't coalesce the MEMBLOCK, so add it to the sorted table. */
355 for (i = type->cnt - 1; i >= 0; i--) {
356 if (base < type->regions[i].base) {
357 type->regions[i+1].base = type->regions[i].base;
358 type->regions[i+1].size = type->regions[i].size;
360 type->regions[i+1].base = base;
361 type->regions[i+1].size = size;
366 if (base < type->regions[0].base) {
367 type->regions[0].base = base;
368 type->regions[0].size = size;
372 /* The array is full ? Try to resize it. If that fails, we undo
373 * our allocation and return an error
375 if (type->cnt == type->max && memblock_double_array(type)) {
383 long __init_memblock memblock_add(phys_addr_t base, phys_addr_t size)
385 return memblock_add_region(&memblock.memory, base, size);
389 static long __init_memblock __memblock_remove(struct memblock_type *type, phys_addr_t base, phys_addr_t size)
391 phys_addr_t rgnbegin, rgnend;
392 phys_addr_t end = base + size;
395 rgnbegin = rgnend = 0; /* supress gcc warnings */
397 /* Find the region where (base, size) belongs to */
398 for (i=0; i < type->cnt; i++) {
399 rgnbegin = type->regions[i].base;
400 rgnend = rgnbegin + type->regions[i].size;
402 if ((rgnbegin <= base) && (end <= rgnend))
406 /* Didn't find the region */
410 /* Check to see if we are removing entire region */
411 if ((rgnbegin == base) && (rgnend == end)) {
412 memblock_remove_region(type, i);
416 /* Check to see if region is matching at the front */
417 if (rgnbegin == base) {
418 type->regions[i].base = end;
419 type->regions[i].size -= size;
423 /* Check to see if the region is matching at the end */
425 type->regions[i].size -= size;
430 * We need to split the entry - adjust the current one to the
431 * beginging of the hole and add the region after hole.
433 type->regions[i].size = base - type->regions[i].base;
434 return memblock_add_region(type, end, rgnend - end);
437 long __init_memblock memblock_remove(phys_addr_t base, phys_addr_t size)
439 return __memblock_remove(&memblock.memory, base, size);
442 long __init_memblock memblock_free(phys_addr_t base, phys_addr_t size)
444 return __memblock_remove(&memblock.reserved, base, size);
447 long __init_memblock memblock_reserve(phys_addr_t base, phys_addr_t size)
449 struct memblock_type *_rgn = &memblock.reserved;
453 return memblock_add_region(_rgn, base, size);
456 phys_addr_t __init __memblock_alloc_base(phys_addr_t size, phys_addr_t align, phys_addr_t max_addr)
460 /* We align the size to limit fragmentation. Without this, a lot of
461 * small allocs quickly eat up the whole reserve array on sparc
463 size = memblock_align_up(size, align);
465 found = memblock_find_base(size, align, 0, max_addr);
466 if (found != MEMBLOCK_ERROR &&
467 memblock_add_region(&memblock.reserved, found, size) >= 0)
473 phys_addr_t __init memblock_alloc_base(phys_addr_t size, phys_addr_t align, phys_addr_t max_addr)
477 alloc = __memblock_alloc_base(size, align, max_addr);
480 panic("ERROR: Failed to allocate 0x%llx bytes below 0x%llx.\n",
481 (unsigned long long) size, (unsigned long long) max_addr);
486 phys_addr_t __init memblock_alloc(phys_addr_t size, phys_addr_t align)
488 return memblock_alloc_base(size, align, MEMBLOCK_ALLOC_ACCESSIBLE);
493 * Additional node-local allocators. Search for node memory is bottom up
494 * and walks memblock regions within that node bottom-up as well, but allocation
495 * within an memblock region is top-down. XXX I plan to fix that at some stage
497 * WARNING: Only available after early_node_map[] has been populated,
498 * on some architectures, that is after all the calls to add_active_range()
499 * have been done to populate it.
502 phys_addr_t __weak __init memblock_nid_range(phys_addr_t start, phys_addr_t end, int *nid)
504 #ifdef CONFIG_ARCH_POPULATES_NODE_MAP
506 * This code originates from sparc which really wants use to walk by addresses
507 * and returns the nid. This is not very convenient for early_pfn_map[] users
508 * as the map isn't sorted yet, and it really wants to be walked by nid.
510 * For now, I implement the inefficient method below which walks the early
511 * map multiple times. Eventually we may want to use an ARCH config option
512 * to implement a completely different method for both case.
514 unsigned long start_pfn, end_pfn;
517 for (i = 0; i < MAX_NUMNODES; i++) {
518 get_pfn_range_for_nid(i, &start_pfn, &end_pfn);
519 if (start < PFN_PHYS(start_pfn) || start >= PFN_PHYS(end_pfn))
522 return min(end, PFN_PHYS(end_pfn));
530 static phys_addr_t __init memblock_alloc_nid_region(struct memblock_region *mp,
532 phys_addr_t align, int nid)
534 phys_addr_t start, end;
537 end = start + mp->size;
539 start = memblock_align_up(start, align);
540 while (start < end) {
541 phys_addr_t this_end;
544 this_end = memblock_nid_range(start, end, &this_nid);
545 if (this_nid == nid) {
546 phys_addr_t ret = memblock_find_region(start, this_end, size, align);
547 if (ret != MEMBLOCK_ERROR &&
548 memblock_add_region(&memblock.reserved, ret, size) >= 0)
554 return MEMBLOCK_ERROR;
557 phys_addr_t __init memblock_alloc_nid(phys_addr_t size, phys_addr_t align, int nid)
559 struct memblock_type *mem = &memblock.memory;
564 /* We align the size to limit fragmentation. Without this, a lot of
565 * small allocs quickly eat up the whole reserve array on sparc
567 size = memblock_align_up(size, align);
569 /* We do a bottom-up search for a region with the right
570 * nid since that's easier considering how memblock_nid_range()
573 for (i = 0; i < mem->cnt; i++) {
574 phys_addr_t ret = memblock_alloc_nid_region(&mem->regions[i],
576 if (ret != MEMBLOCK_ERROR)
583 phys_addr_t __init memblock_alloc_try_nid(phys_addr_t size, phys_addr_t align, int nid)
585 phys_addr_t res = memblock_alloc_nid(size, align, nid);
589 return memblock_alloc_base(size, align, MEMBLOCK_ALLOC_ANYWHERE);
594 * Remaining API functions
597 /* You must call memblock_analyze() before this. */
598 phys_addr_t __init memblock_phys_mem_size(void)
600 return memblock.memory_size;
603 phys_addr_t __init_memblock memblock_end_of_DRAM(void)
605 int idx = memblock.memory.cnt - 1;
607 return (memblock.memory.regions[idx].base + memblock.memory.regions[idx].size);
610 /* You must call memblock_analyze() after this. */
611 void __init memblock_enforce_memory_limit(phys_addr_t memory_limit)
615 struct memblock_region *p;
620 /* Truncate the memblock regions to satisfy the memory limit. */
621 limit = memory_limit;
622 for (i = 0; i < memblock.memory.cnt; i++) {
623 if (limit > memblock.memory.regions[i].size) {
624 limit -= memblock.memory.regions[i].size;
628 memblock.memory.regions[i].size = limit;
629 memblock.memory.cnt = i + 1;
633 memory_limit = memblock_end_of_DRAM();
635 /* And truncate any reserves above the limit also. */
636 for (i = 0; i < memblock.reserved.cnt; i++) {
637 p = &memblock.reserved.regions[i];
639 if (p->base > memory_limit)
641 else if ((p->base + p->size) > memory_limit)
642 p->size = memory_limit - p->base;
645 memblock_remove_region(&memblock.reserved, i);
651 static int memblock_search(struct memblock_type *type, phys_addr_t addr)
653 unsigned int left = 0, right = type->cnt;
656 unsigned int mid = (right + left) / 2;
658 if (addr < type->regions[mid].base)
660 else if (addr >= (type->regions[mid].base +
661 type->regions[mid].size))
665 } while (left < right);
669 int __init memblock_is_reserved(phys_addr_t addr)
671 return memblock_search(&memblock.reserved, addr) != -1;
674 int __init_memblock memblock_is_memory(phys_addr_t addr)
676 return memblock_search(&memblock.memory, addr) != -1;
679 int __init_memblock memblock_is_region_memory(phys_addr_t base, phys_addr_t size)
681 int idx = memblock_search(&memblock.reserved, base);
685 return memblock.reserved.regions[idx].base <= base &&
686 (memblock.reserved.regions[idx].base +
687 memblock.reserved.regions[idx].size) >= (base + size);
690 int __init_memblock memblock_is_region_reserved(phys_addr_t base, phys_addr_t size)
692 return memblock_overlaps_region(&memblock.reserved, base, size) >= 0;
696 void __init_memblock memblock_set_current_limit(phys_addr_t limit)
698 memblock.current_limit = limit;
701 static void __init_memblock memblock_dump(struct memblock_type *region, char *name)
703 unsigned long long base, size;
706 pr_info(" %s.cnt = 0x%lx\n", name, region->cnt);
708 for (i = 0; i < region->cnt; i++) {
709 base = region->regions[i].base;
710 size = region->regions[i].size;
712 pr_info(" %s[%#x]\t[%#016llx-%#016llx], %#llx bytes\n",
713 name, i, base, base + size - 1, size);
717 void __init_memblock memblock_dump_all(void)
722 pr_info("MEMBLOCK configuration:\n");
723 pr_info(" memory size = 0x%llx\n", (unsigned long long)memblock.memory_size);
725 memblock_dump(&memblock.memory, "memory");
726 memblock_dump(&memblock.reserved, "reserved");
729 void __init memblock_analyze(void)
733 /* Check marker in the unused last array entry */
734 WARN_ON(memblock_memory_init_regions[INIT_MEMBLOCK_REGIONS].base
735 != (phys_addr_t)RED_INACTIVE);
736 WARN_ON(memblock_reserved_init_regions[INIT_MEMBLOCK_REGIONS].base
737 != (phys_addr_t)RED_INACTIVE);
739 memblock.memory_size = 0;
741 for (i = 0; i < memblock.memory.cnt; i++)
742 memblock.memory_size += memblock.memory.regions[i].size;
744 /* We allow resizing from there */
745 memblock_can_resize = 1;
748 void __init memblock_init(void)
750 /* Hookup the initial arrays */
751 memblock.memory.regions = memblock_memory_init_regions;
752 memblock.memory.max = INIT_MEMBLOCK_REGIONS;
753 memblock.reserved.regions = memblock_reserved_init_regions;
754 memblock.reserved.max = INIT_MEMBLOCK_REGIONS;
756 /* Write a marker in the unused last array entry */
757 memblock.memory.regions[INIT_MEMBLOCK_REGIONS].base = (phys_addr_t)RED_INACTIVE;
758 memblock.reserved.regions[INIT_MEMBLOCK_REGIONS].base = (phys_addr_t)RED_INACTIVE;
760 /* Create a dummy zero size MEMBLOCK which will get coalesced away later.
761 * This simplifies the memblock_add() code below...
763 memblock.memory.regions[0].base = 0;
764 memblock.memory.regions[0].size = 0;
765 memblock.memory.cnt = 1;
768 memblock.reserved.regions[0].base = 0;
769 memblock.reserved.regions[0].size = 0;
770 memblock.reserved.cnt = 1;
772 memblock.current_limit = MEMBLOCK_ALLOC_ANYWHERE;
775 static int __init early_memblock(char *p)
777 if (p && strstr(p, "debug"))
781 early_param("memblock", early_memblock);
783 #if defined(CONFIG_DEBUG_FS) && !defined(ARCH_DISCARD_MEMBLOCK)
785 static int memblock_debug_show(struct seq_file *m, void *private)
787 struct memblock_type *type = m->private;
788 struct memblock_region *reg;
791 for (i = 0; i < type->cnt; i++) {
792 reg = &type->regions[i];
793 seq_printf(m, "%4d: ", i);
794 if (sizeof(phys_addr_t) == 4)
795 seq_printf(m, "0x%08lx..0x%08lx\n",
796 (unsigned long)reg->base,
797 (unsigned long)(reg->base + reg->size - 1));
799 seq_printf(m, "0x%016llx..0x%016llx\n",
800 (unsigned long long)reg->base,
801 (unsigned long long)(reg->base + reg->size - 1));
807 static int memblock_debug_open(struct inode *inode, struct file *file)
809 return single_open(file, memblock_debug_show, inode->i_private);
812 static const struct file_operations memblock_debug_fops = {
813 .open = memblock_debug_open,
816 .release = single_release,
819 static int __init memblock_init_debugfs(void)
821 struct dentry *root = debugfs_create_dir("memblock", NULL);
824 debugfs_create_file("memory", S_IRUGO, root, &memblock.memory, &memblock_debug_fops);
825 debugfs_create_file("reserved", S_IRUGO, root, &memblock.reserved, &memblock_debug_fops);
829 __initcall(memblock_init_debugfs);
831 #endif /* CONFIG_DEBUG_FS */