projects/linux·blob master · refs

929 lines · 24.9 KiB

1// SPDX-License-Identifier: GPL-2.0
2/*
3 * Virtual Memory Map support
4 *
5 * (C) 2007 sgi. Christoph Lameter.
6 *
7 * Virtual memory maps allow VM primitives pfn_to_page, page_to_pfn,
8 * virt_to_page, page_address() to be implemented as a base offset
9 * calculation without memory access.
10 *
11 * However, virtual mappings need a page table and TLBs. Many Linux
12 * architectures already map their physical space using 1-1 mappings
13 * via TLBs. For those arches the virtual memory map is essentially
14 * for free if we use the same page size as the 1-1 mappings. In that
15 * case the overhead consists of a few additional pages that are
16 * allocated to create a view of memory for vmemmap.
17 *
18 * The architecture is expected to provide a vmemmap_populate() function
19 * to instantiate the mapping.
20 */
21#include <linux/mm.h>
22#include <linux/mmzone.h>
23#include <linux/memblock.h>
24#include <linux/memremap.h>
25#include <linux/highmem.h>
26#include <linux/slab.h>
27#include <linux/spinlock.h>
28#include <linux/vmalloc.h>
29#include <linux/sched.h>
30#include <linux/pgalloc.h>
31
32#include <asm/dma.h>
33#include <asm/tlbflush.h>
34
35#include "hugetlb_vmemmap.h"
36
37/*
38 * Flags for vmemmap_populate_range and friends.
39 */
40/* Get a ref on the head page struct page, for ZONE_DEVICE compound pages */
41#define VMEMMAP_POPULATE_PAGEREF 0x0001
42
43#include "internal.h"
44#include "mm_init.h"
45#include "sparse.h"
46
47/*
48 * Allocate a block of memory to be used to back the virtual memory map
49 * or to back the page tables that are used to create the mapping.
50 * Uses the main allocators if they are available, else bootmem.
51 */
52
53static void * __ref __earlyonly_bootmem_alloc(int node,
54 unsigned long size,
55 unsigned long align,
56 unsigned long goal)
57{
58 return memmap_alloc(size, align, goal, node, false);
59}
60
61void * __meminit vmemmap_alloc_block(unsigned long size, int node)
62{
63 /* If the main allocator is up use that, fallback to bootmem. */
64 if (slab_is_available()) {
65 gfp_t gfp_mask = GFP_KERNEL|__GFP_RETRY_MAYFAIL|__GFP_NOWARN;
66 int order = get_order(size);
67 static bool warned __meminitdata;
68 struct page *page;
69
70 page = alloc_pages_node(node, gfp_mask, order);
71 if (page)
72 return page_address(page);
73
74 if (!warned) {
75 warn_alloc(gfp_mask & ~__GFP_NOWARN, NULL,
76 "vmemmap alloc failure: order:%u", order);
77 warned = true;
78 }
79 return NULL;
80 } else
81 return __earlyonly_bootmem_alloc(node, size, size,
82 __pa(MAX_DMA_ADDRESS));
83}
84
85static void * __meminit altmap_alloc_block_buf(unsigned long size,
86 struct vmem_altmap *altmap);
87
88/* need to make sure size is all the same during early stage */
89void * __meminit vmemmap_alloc_block_buf(unsigned long size, int node,
90 struct vmem_altmap *altmap)
91{
92 if (altmap)
93 return altmap_alloc_block_buf(size, altmap);
94
95 return vmemmap_alloc_block(size, node);
96}
97
98static unsigned long __meminit vmem_altmap_next_pfn(struct vmem_altmap *altmap)
99{
100 return altmap->base_pfn + altmap->reserve + altmap->alloc
101 + altmap->align;
102}
103
104static unsigned long __meminit vmem_altmap_nr_free(struct vmem_altmap *altmap)
105{
106 unsigned long allocated = altmap->alloc + altmap->align;
107
108 if (altmap->free > allocated)
109 return altmap->free - allocated;
110 return 0;
111}
112
113static void * __meminit altmap_alloc_block_buf(unsigned long size,
114 struct vmem_altmap *altmap)
115{
116 unsigned long pfn, nr_pfns, nr_align;
117
118 if (size & ~PAGE_MASK) {
119 pr_warn_once("%s: allocations must be multiple of PAGE_SIZE (%ld)\n",
120 __func__, size);
121 return NULL;
122 }
123
124 pfn = vmem_altmap_next_pfn(altmap);
125 nr_pfns = size >> PAGE_SHIFT;
126 nr_align = 1UL << find_first_bit(&nr_pfns, BITS_PER_LONG);
127 nr_align = ALIGN(pfn, nr_align) - pfn;
128 if (nr_pfns + nr_align > vmem_altmap_nr_free(altmap))
129 return NULL;
130
131 altmap->alloc += nr_pfns;
132 altmap->align += nr_align;
133 pfn += nr_align;
134
135 pr_debug("%s: pfn: %#lx alloc: %ld align: %ld nr: %#lx\n",
136 __func__, pfn, altmap->alloc, altmap->align, nr_pfns);
137 return __va(__pfn_to_phys(pfn));
138}
139
140void __meminit vmemmap_verify(pte_t *pte, int node,
141 unsigned long start, unsigned long end)
142{
143 unsigned long pfn = pte_pfn(ptep_get(pte));
144 int actual_node = early_pfn_to_nid(pfn);
145
146 if (node_distance(actual_node, node) > LOCAL_DISTANCE)
147 pr_warn_once("[%lx-%lx] potential offnode page_structs\n",
148 start, end - 1);
149}
150
151static pte_t * __meminit vmemmap_pte_populate(pmd_t *pmd, unsigned long addr, int node,
152 struct vmem_altmap *altmap,
153 unsigned long ptpfn, unsigned long flags)
154{
155 pte_t *pte = pte_offset_kernel(pmd, addr);
156 if (pte_none(ptep_get(pte))) {
157 pte_t entry;
158 void *p;
159
160 if (ptpfn == (unsigned long)-1) {
161 p = vmemmap_alloc_block_buf(PAGE_SIZE, node, altmap);
162 if (!p)
163 return NULL;
164 ptpfn = PHYS_PFN(__pa(p));
165 } else {
166 /*
167 * When a PTE/PMD entry is freed from the init_mm
168 * there's a free_pages() call to this page allocated
169 * above. Thus this get_page() is paired with the
170 * put_page_testzero() on the freeing path.
171 * This can only called by certain ZONE_DEVICE path,
172 * and through vmemmap_populate_compound_pages() when
173 * slab is available.
174 */
175 if (flags & VMEMMAP_POPULATE_PAGEREF)
176 get_page(pfn_to_page(ptpfn));
177 }
178 entry = pfn_pte(ptpfn, PAGE_KERNEL);
179 set_pte_at(&init_mm, addr, pte, entry);
180 }
181 return pte;
182}
183
184static void * __meminit vmemmap_alloc_block_zero(unsigned long size, int node)
185{
186 void *p = vmemmap_alloc_block(size, node);
187
188 if (!p)
189 return NULL;
190 memset(p, 0, size);
191
192 return p;
193}
194
195static pmd_t * __meminit vmemmap_pmd_populate(pud_t *pud, unsigned long addr, int node)
196{
197 pmd_t *pmd = pmd_offset(pud, addr);
198 if (pmd_none(*pmd)) {
199 void *p = vmemmap_alloc_block_zero(PAGE_SIZE, node);
200 if (!p)
201 return NULL;
202 kernel_pte_init(p);
203 pmd_populate_kernel(&init_mm, pmd, p);
204 }
205 return pmd;
206}
207
208static pud_t * __meminit vmemmap_pud_populate(p4d_t *p4d, unsigned long addr, int node)
209{
210 pud_t *pud = pud_offset(p4d, addr);
211 if (pud_none(*pud)) {
212 void *p = vmemmap_alloc_block_zero(PAGE_SIZE, node);
213 if (!p)
214 return NULL;
215 pmd_init(p);
216 pud_populate(&init_mm, pud, p);
217 }
218 return pud;
219}
220
221static p4d_t * __meminit vmemmap_p4d_populate(pgd_t *pgd, unsigned long addr, int node)
222{
223 p4d_t *p4d = p4d_offset(pgd, addr);
224 if (p4d_none(*p4d)) {
225 void *p = vmemmap_alloc_block_zero(PAGE_SIZE, node);
226 if (!p)
227 return NULL;
228 pud_init(p);
229 p4d_populate_kernel(addr, p4d, p);
230 }
231 return p4d;
232}
233
234static pgd_t * __meminit vmemmap_pgd_populate(unsigned long addr, int node)
235{
236 pgd_t *pgd = pgd_offset_k(addr);
237 if (pgd_none(*pgd)) {
238 void *p = vmemmap_alloc_block_zero(PAGE_SIZE, node);
239 if (!p)
240 return NULL;
241 pgd_populate_kernel(addr, pgd, p);
242 }
243 return pgd;
244}
245
246static pte_t * __meminit vmemmap_populate_address(unsigned long addr, int node,
247 struct vmem_altmap *altmap,
248 unsigned long ptpfn,
249 unsigned long flags)
250{
251 pgd_t *pgd;
252 p4d_t *p4d;
253 pud_t *pud;
254 pmd_t *pmd;
255 pte_t *pte;
256
257 pgd = vmemmap_pgd_populate(addr, node);
258 if (!pgd)
259 return NULL;
260 p4d = vmemmap_p4d_populate(pgd, addr, node);
261 if (!p4d)
262 return NULL;
263 pud = vmemmap_pud_populate(p4d, addr, node);
264 if (!pud)
265 return NULL;
266 pmd = vmemmap_pmd_populate(pud, addr, node);
267 if (!pmd)
268 return NULL;
269 pte = vmemmap_pte_populate(pmd, addr, node, altmap, ptpfn, flags);
270 if (!pte)
271 return NULL;
272 vmemmap_verify(pte, node, addr, addr + PAGE_SIZE);
273
274 return pte;
275}
276
277static int __meminit vmemmap_populate_range(unsigned long start,
278 unsigned long end, int node,
279 struct vmem_altmap *altmap,
280 unsigned long ptpfn,
281 unsigned long flags)
282{
283 unsigned long addr = start;
284 pte_t *pte;
285
286 for (; addr < end; addr += PAGE_SIZE) {
287 pte = vmemmap_populate_address(addr, node, altmap,
288 ptpfn, flags);
289 if (!pte)
290 return -ENOMEM;
291 }
292
293 return 0;
294}
295
296int __meminit vmemmap_populate_basepages(unsigned long start, unsigned long end,
297 int node, struct vmem_altmap *altmap)
298{
299 return vmemmap_populate_range(start, end, node, altmap, -1, 0);
300}
301
302/*
303 * Write protect the mirrored tail page structs for HVO. This will be
304 * called from the hugetlb code when gathering and initializing the
305 * memblock allocated gigantic pages. The write protect can't be
306 * done earlier, since it can't be guaranteed that the reserved
307 * page structures will not be written to during initialization,
308 * even if CONFIG_DEFERRED_STRUCT_PAGE_INIT is enabled.
309 *
310 * The PTEs are known to exist, and nothing else should be touching
311 * these pages. The caller is responsible for any TLB flushing.
312 */
313void vmemmap_wrprotect_hvo(unsigned long addr, unsigned long end,
314 int node, unsigned long headsize)
315{
316 unsigned long maddr;
317 pte_t *pte;
318
319 for (maddr = addr + headsize; maddr < end; maddr += PAGE_SIZE) {
320 pte = virt_to_kpte(maddr);
321 ptep_set_wrprotect(&init_mm, maddr, pte);
322 }
323}
324
325#ifdef CONFIG_HUGETLB_PAGE_OPTIMIZE_VMEMMAP
326static __meminit struct page *vmemmap_get_tail(unsigned int order, struct zone *zone)
327{
328 struct page *p, *tail;
329 unsigned int idx;
330 int node = zone_to_nid(zone);
331
332 if (WARN_ON_ONCE(order < VMEMMAP_TAIL_MIN_ORDER))
333 return NULL;
334 if (WARN_ON_ONCE(order > MAX_FOLIO_ORDER))
335 return NULL;
336
337 idx = order - VMEMMAP_TAIL_MIN_ORDER;
338 tail = zone->vmemmap_tails[idx];
339 if (tail)
340 return tail;
341
342 /*
343 * Only allocate the page, but do not initialize it.
344 *
345 * Any initialization done here will be overwritten by memmap_init().
346 *
347 * hugetlb_bootmem_struct_page_init() will take care of initialization
348 * after memmap_init().
349 */
350
351 p = vmemmap_alloc_block_zero(PAGE_SIZE, node);
352 if (!p)
353 return NULL;
354
355 tail = virt_to_page(p);
356 zone->vmemmap_tails[idx] = tail;
357
358 return tail;
359}
360
361int __meminit vmemmap_populate_hvo(unsigned long addr, unsigned long end,
362 unsigned int order, struct zone *zone,
363 unsigned long headsize)
364{
365 unsigned long maddr;
366 struct page *tail;
367 pte_t *pte;
368 int node = zone_to_nid(zone);
369
370 tail = vmemmap_get_tail(order, zone);
371 if (!tail)
372 return -ENOMEM;
373
374 for (maddr = addr; maddr < addr + headsize; maddr += PAGE_SIZE) {
375 pte = vmemmap_populate_address(maddr, node, NULL, -1, 0);
376 if (!pte)
377 return -ENOMEM;
378 }
379
380 /*
381 * Reuse the last page struct page mapped above for the rest.
382 */
383 return vmemmap_populate_range(maddr, end, node, NULL,
384 page_to_pfn(tail), 0);
385}
386#endif
387
388void __weak __meminit vmemmap_set_pmd(pmd_t *pmd, void *p, int node,
389 unsigned long addr, unsigned long next)
390{
391 WARN_ON_ONCE(!pmd_set_huge(pmd, virt_to_phys(p), PAGE_KERNEL));
392}
393
394int __weak __meminit vmemmap_check_pmd(pmd_t *pmd, int node,
395 unsigned long addr, unsigned long next)
396{
397 if (!pmd_leaf(pmdp_get(pmd)))
398 return 0;
399 vmemmap_verify((pte_t *)pmd, node, addr, next);
400
401 return 1;
402}
403
404int __meminit vmemmap_populate_hugepages(unsigned long start, unsigned long end,
405 int node, struct vmem_altmap *altmap)
406{
407 unsigned long addr;
408 unsigned long next;
409 pgd_t *pgd;
410 p4d_t *p4d;
411 pud_t *pud;
412 pmd_t *pmd;
413
414 for (addr = start; addr < end; addr = next) {
415 next = pmd_addr_end(addr, end);
416
417 pgd = vmemmap_pgd_populate(addr, node);
418 if (!pgd)
419 return -ENOMEM;
420
421 p4d = vmemmap_p4d_populate(pgd, addr, node);
422 if (!p4d)
423 return -ENOMEM;
424
425 pud = vmemmap_pud_populate(p4d, addr, node);
426 if (!pud)
427 return -ENOMEM;
428
429 pmd = pmd_offset(pud, addr);
430 if (pmd_none(pmdp_get(pmd))) {
431 void *p;
432
433 p = vmemmap_alloc_block_buf(PMD_SIZE, node, altmap);
434 if (p) {
435 vmemmap_set_pmd(pmd, p, node, addr, next);
436 continue;
437 } else if (altmap) {
438 /*
439 * No fallback: In any case we care about, the
440 * altmap should be reasonably sized and aligned
441 * such that vmemmap_alloc_block_buf() will always
442 * succeed. For consistency with the PTE case,
443 * return an error here as failure could indicate
444 * a configuration issue with the size of the altmap.
445 */
446 return -ENOMEM;
447 }
448 } else if (vmemmap_check_pmd(pmd, node, addr, next))
449 continue;
450 if (vmemmap_populate_basepages(addr, next, node, altmap))
451 return -ENOMEM;
452 }
453 return 0;
454}
455
456#ifndef vmemmap_populate_compound_pages
457/*
458 * For compound pages bigger than section size (e.g. x86 1G compound
459 * pages with 2M subsection size) fill the rest of sections as tail
460 * pages.
461 *
462 * Note that memremap_pages() resets @nr_range value and will increment
463 * it after each range successful onlining. Thus the value or @nr_range
464 * at section memmap populate corresponds to the in-progress range
465 * being onlined here.
466 */
467static bool __meminit reuse_compound_section(unsigned long start_pfn,
468 struct dev_pagemap *pgmap)
469{
470 unsigned long nr_pages = pgmap_vmemmap_nr(pgmap);
471 unsigned long offset = start_pfn -
472 PHYS_PFN(pgmap->ranges[pgmap->nr_range].start);
473
474 return !IS_ALIGNED(offset, nr_pages) && nr_pages > PAGES_PER_SUBSECTION;
475}
476
477static pte_t * __meminit compound_section_tail_page(unsigned long addr)
478{
479 pte_t *pte;
480
481 addr -= PAGE_SIZE;
482
483 /*
484 * Assuming sections are populated sequentially, the previous section's
485 * page data can be reused.
486 */
487 pte = pte_offset_kernel(pmd_off_k(addr), addr);
488 if (!pte)
489 return NULL;
490
491 return pte;
492}
493
494static int __meminit vmemmap_populate_compound_pages(unsigned long start_pfn,
495 unsigned long start,
496 unsigned long end, int node,
497 struct dev_pagemap *pgmap)
498{
499 unsigned long size, addr;
500 pte_t *pte;
501 int rc;
502
503 if (reuse_compound_section(start_pfn, pgmap)) {
504 pte = compound_section_tail_page(start);
505 if (!pte)
506 return -ENOMEM;
507
508 /*
509 * Reuse the page that was populated in the prior iteration
510 * with just tail struct pages.
511 */
512 return vmemmap_populate_range(start, end, node, NULL,
513 pte_pfn(ptep_get(pte)),
514 VMEMMAP_POPULATE_PAGEREF);
515 }
516
517 size = min(end - start, pgmap_vmemmap_nr(pgmap) * sizeof(struct page));
518 for (addr = start; addr < end; addr += size) {
519 unsigned long next, last = addr + size;
520
521 /* Populate the head page vmemmap page */
522 pte = vmemmap_populate_address(addr, node, NULL, -1, 0);
523 if (!pte)
524 return -ENOMEM;
525
526 /* Populate the tail pages vmemmap page */
527 next = addr + PAGE_SIZE;
528 pte = vmemmap_populate_address(next, node, NULL, -1, 0);
529 if (!pte)
530 return -ENOMEM;
531
532 /*
533 * Reuse the previous page for the rest of tail pages
534 * See layout diagram in Documentation/mm/vmemmap_dedup.rst
535 */
536 next += PAGE_SIZE;
537 rc = vmemmap_populate_range(next, last, node, NULL,
538 pte_pfn(ptep_get(pte)),
539 VMEMMAP_POPULATE_PAGEREF);
540 if (rc)
541 return -ENOMEM;
542 }
543
544 return 0;
545}
546
547#endif
548
549struct page * __meminit __populate_section_memmap(unsigned long pfn,
550 unsigned long nr_pages, int nid, struct vmem_altmap *altmap,
551 struct dev_pagemap *pgmap)
552{
553 unsigned long start = (unsigned long) pfn_to_page(pfn);
554 unsigned long end = start + nr_pages * sizeof(struct page);
555 int r;
556
557 if (WARN_ON_ONCE(!IS_ALIGNED(pfn, PAGES_PER_SUBSECTION) ||
558 !IS_ALIGNED(nr_pages, PAGES_PER_SUBSECTION)))
559 return NULL;
560
561 if (vmemmap_can_optimize(altmap, pgmap))
562 r = vmemmap_populate_compound_pages(pfn, start, end, nid, pgmap);
563 else
564 r = vmemmap_populate(start, end, nid, altmap);
565
566 if (r < 0)
567 return NULL;
568
569 flush_cache_vmap(start, end);
570
571 return pfn_to_page(pfn);
572}
573
574#ifdef CONFIG_SPARSEMEM_VMEMMAP_PREINIT
575/*
576 * This is called just before initializing sections for a NUMA node.
577 * Any special initialization that needs to be done before the
578 * generic initialization can be done from here. Sections that
579 * are initialized in hooks called from here will be skipped by
580 * the generic initialization.
581 */
582void __init sparse_vmemmap_init_nid_early(int nid)
583{
584 hugetlb_vmemmap_init_early(nid);
585}
586#endif
587
588static void subsection_mask_set(unsigned long *map, unsigned long pfn,
589 unsigned long nr_pages)
590{
591 int idx = subsection_map_index(pfn);
592 int end = subsection_map_index(pfn + nr_pages - 1);
593
594 bitmap_set(map, idx, end - idx + 1);
595}
596
597static void __init sparse_init_subsection_map_range(unsigned long pfn, unsigned long nr_pages)
598{
599 int end_sec_nr = pfn_to_section_nr(pfn + nr_pages - 1);
600 unsigned long nr, start_sec_nr = pfn_to_section_nr(pfn);
601
602 for (nr = start_sec_nr; nr <= end_sec_nr; nr++) {
603 struct mem_section *ms;
604 unsigned long pfns;
605
606 pfns = min(nr_pages, PAGES_PER_SECTION
607 - (pfn & ~PAGE_SECTION_MASK));
608 ms = __nr_to_section(nr);
609 subsection_mask_set(ms->usage->subsection_map, pfn, pfns);
610
611 pr_debug("%s: sec: %lu pfns: %lu set(%d, %d)\n", __func__, nr,
612 pfns, subsection_map_index(pfn),
613 subsection_map_index(pfn + pfns - 1));
614
615 pfn += pfns;
616 nr_pages -= pfns;
617 }
618}
619
620void __init sparse_init_subsection_map(void)
621{
622 int i, nid;
623 unsigned long start, end;
624
625 for_each_mem_pfn_range(i, MAX_NUMNODES, &start, &end, &nid)
626 sparse_init_subsection_map_range(start, end - start);
627}
628
629#ifdef CONFIG_MEMORY_HOTPLUG
630
631/* Mark all memory sections within the pfn range as online */
632void online_mem_sections(unsigned long start_pfn, unsigned long end_pfn)
633{
634 unsigned long pfn;
635
636 for (pfn = start_pfn; pfn < end_pfn; pfn += PAGES_PER_SECTION) {
637 unsigned long section_nr = pfn_to_section_nr(pfn);
638 struct mem_section *ms = __nr_to_section(section_nr);
639
640 ms->section_mem_map |= SECTION_IS_ONLINE;
641 }
642}
643
644/* Mark all memory sections within the pfn range as offline */
645void offline_mem_sections(unsigned long start_pfn, unsigned long end_pfn)
646{
647 unsigned long pfn;
648
649 for (pfn = start_pfn; pfn < end_pfn; pfn += PAGES_PER_SECTION) {
650 unsigned long section_nr = pfn_to_section_nr(pfn);
651 struct mem_section *ms = __nr_to_section(section_nr);
652
653 ms->section_mem_map &= ~SECTION_IS_ONLINE;
654 }
655}
656
657static int __meminit section_nr_vmemmap_pages(unsigned long pfn, unsigned long nr_pages,
658 struct vmem_altmap *altmap, struct dev_pagemap *pgmap)
659{
660 const unsigned int order = pgmap ? pgmap->vmemmap_shift : 0;
661 const unsigned long pages_per_compound = 1UL << order;
662
663 VM_WARN_ON_ONCE(!IS_ALIGNED(pfn | nr_pages, PAGES_PER_SUBSECTION));
664 VM_WARN_ON_ONCE(nr_pages > PAGES_PER_SECTION);
665
666 if (!vmemmap_can_optimize(altmap, pgmap))
667 return DIV_ROUND_UP(nr_pages * sizeof(struct page), PAGE_SIZE);
668
669 if (order < PFN_SECTION_SHIFT) {
670 VM_WARN_ON_ONCE(!IS_ALIGNED(pfn | nr_pages, pages_per_compound));
671 return VMEMMAP_RESERVE_NR * nr_pages / pages_per_compound;
672 }
673
674 VM_WARN_ON_ONCE(!IS_ALIGNED(pfn | nr_pages, PAGES_PER_SECTION));
675
676 if (IS_ALIGNED(pfn, pages_per_compound))
677 return VMEMMAP_RESERVE_NR;
678
679 return 0;
680}
681
682static struct page * __meminit populate_section_memmap(unsigned long pfn,
683 unsigned long nr_pages, int nid, struct vmem_altmap *altmap,
684 struct dev_pagemap *pgmap)
685{
686 struct page *page = __populate_section_memmap(pfn, nr_pages, nid, altmap,
687 pgmap);
688
689 memmap_pages_add(section_nr_vmemmap_pages(pfn, nr_pages, altmap, pgmap));
690
691 return page;
692}
693
694static void depopulate_section_memmap(unsigned long pfn, unsigned long nr_pages,
695 struct vmem_altmap *altmap, struct dev_pagemap *pgmap)
696{
697 unsigned long start = (unsigned long) pfn_to_page(pfn);
698 unsigned long end = start + nr_pages * sizeof(struct page);
699
700 memmap_pages_add(-section_nr_vmemmap_pages(pfn, nr_pages, altmap, pgmap));
701 vmemmap_free(start, end, altmap);
702}
703
704static void free_map_bootmem(struct page *memmap)
705{
706 unsigned long start = (unsigned long)memmap;
707 unsigned long end = (unsigned long)(memmap + PAGES_PER_SECTION);
708 unsigned long pfn = page_to_pfn(memmap);
709
710 memmap_boot_pages_add(-section_nr_vmemmap_pages(pfn, PAGES_PER_SECTION,
711 NULL, NULL));
712 vmemmap_free(start, end, NULL);
713}
714
715static int clear_subsection_map(unsigned long pfn, unsigned long nr_pages)
716{
717 DECLARE_BITMAP(map, SUBSECTIONS_PER_SECTION) = { 0 };
718 DECLARE_BITMAP(tmp, SUBSECTIONS_PER_SECTION) = { 0 };
719 struct mem_section *ms = __pfn_to_section(pfn);
720 unsigned long *subsection_map = ms->usage
721 ? &ms->usage->subsection_map[0] : NULL;
722
723 subsection_mask_set(map, pfn, nr_pages);
724 if (subsection_map)
725 bitmap_and(tmp, map, subsection_map, SUBSECTIONS_PER_SECTION);
726
727 if (WARN(!subsection_map || !bitmap_equal(tmp, map, SUBSECTIONS_PER_SECTION),
728 "section already deactivated (%#lx + %ld)\n",
729 pfn, nr_pages))
730 return -EINVAL;
731
732 bitmap_xor(subsection_map, map, subsection_map, SUBSECTIONS_PER_SECTION);
733 return 0;
734}
735
736static bool is_subsection_map_empty(struct mem_section *ms)
737{
738 return bitmap_empty(&ms->usage->subsection_map[0],
739 SUBSECTIONS_PER_SECTION);
740}
741
742static int fill_subsection_map(unsigned long pfn, unsigned long nr_pages)
743{
744 struct mem_section *ms = __pfn_to_section(pfn);
745 DECLARE_BITMAP(map, SUBSECTIONS_PER_SECTION) = { 0 };
746 unsigned long *subsection_map;
747 int rc = 0;
748
749 subsection_mask_set(map, pfn, nr_pages);
750
751 subsection_map = &ms->usage->subsection_map[0];
752
753 if (bitmap_empty(map, SUBSECTIONS_PER_SECTION))
754 rc = -EINVAL;
755 else if (bitmap_intersects(map, subsection_map, SUBSECTIONS_PER_SECTION))
756 rc = -EEXIST;
757 else
758 bitmap_or(subsection_map, map, subsection_map,
759 SUBSECTIONS_PER_SECTION);
760
761 return rc;
762}
763
764/*
765 * To deactivate a memory region, there are 3 cases to handle:
766 *
767 * 1. deactivation of a partial hot-added section:
768 * a) section was present at memory init.
769 * b) section was hot-added post memory init.
770 * 2. deactivation of a complete hot-added section.
771 * 3. deactivation of a complete section from memory init.
772 *
773 * For 1, when subsection_map does not empty we will not be freeing the
774 * usage map, but still need to free the vmemmap range.
775 */
776static void section_deactivate(unsigned long pfn, unsigned long nr_pages,
777 struct vmem_altmap *altmap, struct dev_pagemap *pgmap)
778{
779 struct mem_section *ms = __pfn_to_section(pfn);
780 bool section_is_early = early_section(ms);
781 struct page *memmap = NULL;
782 bool empty;
783
784 if (clear_subsection_map(pfn, nr_pages))
785 return;
786
787 empty = is_subsection_map_empty(ms);
788 if (empty) {
789 /*
790 * Mark the section invalid so that valid_section()
791 * return false. This prevents code from dereferencing
792 * ms->usage array.
793 */
794 ms->section_mem_map &= ~SECTION_HAS_MEM_MAP;
795
796 /*
797 * When removing an early section, the usage map is kept (as the
798 * usage maps of other sections fall into the same page). It
799 * will be re-used when re-adding the section - which is then no
800 * longer an early section. If the usage map is PageReserved, it
801 * was allocated during boot.
802 */
803 if (!PageReserved(virt_to_page(ms->usage))) {
804 kfree_rcu(ms->usage, rcu);
805 WRITE_ONCE(ms->usage, NULL);
806 }
807 memmap = pfn_to_page(SECTION_ALIGN_DOWN(pfn));
808 }
809
810 /*
811 * The memmap of early sections is always fully populated. See
812 * section_activate() and pfn_valid() .
813 */
814 if (!section_is_early)
815 depopulate_section_memmap(pfn, nr_pages, altmap, pgmap);
816 else if (memmap)
817 free_map_bootmem(memmap);
818
819 if (empty)
820 ms->section_mem_map = (unsigned long)NULL;
821}
822
823static struct page * __meminit section_activate(int nid, unsigned long pfn,
824 unsigned long nr_pages, struct vmem_altmap *altmap,
825 struct dev_pagemap *pgmap)
826{
827 struct mem_section *ms = __pfn_to_section(pfn);
828 struct mem_section_usage *usage = NULL;
829 struct page *memmap;
830 int rc;
831
832 if (!ms->usage) {
833 usage = kzalloc(mem_section_usage_size(), GFP_KERNEL);
834 if (!usage)
835 return ERR_PTR(-ENOMEM);
836 ms->usage = usage;
837 }
838
839 rc = fill_subsection_map(pfn, nr_pages);
840 if (rc) {
841 if (usage)
842 ms->usage = NULL;
843 kfree(usage);
844 return ERR_PTR(rc);
845 }
846
847 /*
848 * The early init code does not consider partially populated
849 * initial sections, it simply assumes that memory will never be
850 * referenced. If we hot-add memory into such a section then we
851 * do not need to populate the memmap and can simply reuse what
852 * is already there.
853 */
854 if (nr_pages < PAGES_PER_SECTION && early_section(ms))
855 return pfn_to_page(pfn);
856
857 memmap = populate_section_memmap(pfn, nr_pages, nid, altmap, pgmap);
858 if (!memmap) {
859 section_deactivate(pfn, nr_pages, altmap, pgmap);
860 return ERR_PTR(-ENOMEM);
861 }
862
863 return memmap;
864}
865
866/**
867 * sparse_add_section - add a memory section, or populate an existing one
868 * @nid: The node to add section on
869 * @start_pfn: start pfn of the memory range
870 * @nr_pages: number of pfns to add in the section
871 * @altmap: alternate pfns to allocate the memmap backing store
872 * @pgmap: alternate compound page geometry for devmap mappings
873 *
874 * This is only intended for hotplug.
875 *
876 * Note that only VMEMMAP supports sub-section aligned hotplug,
877 * the proper alignment and size are gated by check_pfn_span().
878 *
879 *
880 * Return:
881 * * 0 - On success.
882 * * -EEXIST - Section has been present.
883 * * -ENOMEM - Out of memory.
884 */
885int __meminit sparse_add_section(int nid, unsigned long start_pfn,
886 unsigned long nr_pages, struct vmem_altmap *altmap,
887 struct dev_pagemap *pgmap)
888{
889 unsigned long section_nr = pfn_to_section_nr(start_pfn);
890 struct mem_section *ms;
891 struct page *memmap;
892 int ret;
893
894 ret = sparse_index_init(section_nr, nid);
895 if (ret < 0)
896 return ret;
897
898 memmap = section_activate(nid, start_pfn, nr_pages, altmap, pgmap);
899 if (IS_ERR(memmap))
900 return PTR_ERR(memmap);
901
902 /*
903 * Poison uninitialized struct pages in order to catch invalid flags
904 * combinations.
905 */
906 page_init_poison(memmap, sizeof(struct page) * nr_pages);
907
908 ms = __nr_to_section(section_nr);
909 __section_mark_present(ms, section_nr);
910
911 /* Align memmap to section boundary in the subsection case */
912 if (section_nr_to_pfn(section_nr) != start_pfn)
913 memmap = pfn_to_page(section_nr_to_pfn(section_nr));
914 sparse_init_one_section(ms, section_nr, memmap, ms->usage, 0);
915
916 return 0;
917}
918
919void sparse_remove_section(unsigned long pfn, unsigned long nr_pages,
920 struct vmem_altmap *altmap, struct dev_pagemap *pgmap)
921{
922 struct mem_section *ms = __pfn_to_section(pfn);
923
924 if (WARN_ON_ONCE(!valid_section(ms)))
925 return;
926
927 section_deactivate(pfn, nr_pages, altmap, pgmap);
928}
929#endif /* CONFIG_MEMORY_HOTPLUG */