1 // SPDX-License-Identifier: GPL-2.0
2 /*
3 * sparse memory mappings.
4 */
5 #include <linux/mm.h>
6 #include <linux/slab.h>
7 #include <linux/mmzone.h>
8 #include <linux/memblock.h>
9 #include <linux/compiler.h>
10 #include <linux/highmem.h>
11 #include <linux/export.h>
12 #include <linux/spinlock.h>
13 #include <linux/vmalloc.h>
14 #include <linux/swap.h>
15 #include <linux/swapops.h>
16 #include <linux/bootmem_info.h>
17
18 #include "internal.h"
19 #include <asm/dma.h>
20
21 /*
22 * Permanent SPARSEMEM data:
23 *
24 * 1) mem_section - memory sections, mem_map's for valid memory
25 */
26 #ifdef CONFIG_SPARSEMEM_EXTREME
27 struct mem_section **mem_section;
28 #else
29 struct mem_section mem_section[NR_SECTION_ROOTS][SECTIONS_PER_ROOT]
30 ____cacheline_internodealigned_in_smp;
31 #endif
32 EXPORT_SYMBOL(mem_section);
33
34 #ifdef NODE_NOT_IN_PAGE_FLAGS
35 /*
36 * If we did not store the node number in the page then we have to
37 * do a lookup in the section_to_node_table in order to find which
38 * node the page belongs to.
39 */
40 #if MAX_NUMNODES <= 256
41 static u8 section_to_node_table[NR_MEM_SECTIONS] __cacheline_aligned;
42 #else
43 static u16 section_to_node_table[NR_MEM_SECTIONS] __cacheline_aligned;
44 #endif
45
page_to_nid(const struct page * page)46 int page_to_nid(const struct page *page)
47 {
48 return section_to_node_table[page_to_section(page)];
49 }
50 EXPORT_SYMBOL(page_to_nid);
51
set_section_nid(unsigned long section_nr,int nid)52 static void set_section_nid(unsigned long section_nr, int nid)
53 {
54 section_to_node_table[section_nr] = nid;
55 }
56 #else /* !NODE_NOT_IN_PAGE_FLAGS */
set_section_nid(unsigned long section_nr,int nid)57 static inline void set_section_nid(unsigned long section_nr, int nid)
58 {
59 }
60 #endif
61
62 #ifdef CONFIG_SPARSEMEM_EXTREME
sparse_index_alloc(int nid)63 static noinline struct mem_section __ref *sparse_index_alloc(int nid)
64 {
65 struct mem_section *section = NULL;
66 unsigned long array_size = SECTIONS_PER_ROOT *
67 sizeof(struct mem_section);
68
69 if (slab_is_available()) {
70 section = kzalloc_node(array_size, GFP_KERNEL, nid);
71 } else {
72 section = memblock_alloc_node(array_size, SMP_CACHE_BYTES,
73 nid);
74 if (!section)
75 panic("%s: Failed to allocate %lu bytes nid=%d\n",
76 __func__, array_size, nid);
77 }
78
79 return section;
80 }
81
sparse_index_init(unsigned long section_nr,int nid)82 static int __meminit sparse_index_init(unsigned long section_nr, int nid)
83 {
84 unsigned long root = SECTION_NR_TO_ROOT(section_nr);
85 struct mem_section *section;
86
87 /*
88 * An existing section is possible in the sub-section hotplug
89 * case. First hot-add instantiates, follow-on hot-add reuses
90 * the existing section.
91 *
92 * The mem_hotplug_lock resolves the apparent race below.
93 */
94 if (mem_section[root])
95 return 0;
96
97 section = sparse_index_alloc(nid);
98 if (!section)
99 return -ENOMEM;
100
101 mem_section[root] = section;
102
103 return 0;
104 }
105 #else /* !SPARSEMEM_EXTREME */
sparse_index_init(unsigned long section_nr,int nid)106 static inline int sparse_index_init(unsigned long section_nr, int nid)
107 {
108 return 0;
109 }
110 #endif
111
112 /*
113 * During early boot, before section_mem_map is used for an actual
114 * mem_map, we use section_mem_map to store the section's NUMA
115 * node. This keeps us from having to use another data structure. The
116 * node information is cleared just before we store the real mem_map.
117 */
sparse_encode_early_nid(int nid)118 static inline unsigned long sparse_encode_early_nid(int nid)
119 {
120 return ((unsigned long)nid << SECTION_NID_SHIFT);
121 }
122
sparse_early_nid(struct mem_section * section)123 static inline int sparse_early_nid(struct mem_section *section)
124 {
125 return (section->section_mem_map >> SECTION_NID_SHIFT);
126 }
127
128 /* Validate the physical addressing limitations of the model */
mminit_validate_memmodel_limits(unsigned long * start_pfn,unsigned long * end_pfn)129 static void __meminit mminit_validate_memmodel_limits(unsigned long *start_pfn,
130 unsigned long *end_pfn)
131 {
132 unsigned long max_sparsemem_pfn = 1UL << (MAX_PHYSMEM_BITS-PAGE_SHIFT);
133
134 /*
135 * Sanity checks - do not allow an architecture to pass
136 * in larger pfns than the maximum scope of sparsemem:
137 */
138 if (*start_pfn > max_sparsemem_pfn) {
139 mminit_dprintk(MMINIT_WARNING, "pfnvalidation",
140 "Start of range %lu -> %lu exceeds SPARSEMEM max %lu\n",
141 *start_pfn, *end_pfn, max_sparsemem_pfn);
142 WARN_ON_ONCE(1);
143 *start_pfn = max_sparsemem_pfn;
144 *end_pfn = max_sparsemem_pfn;
145 } else if (*end_pfn > max_sparsemem_pfn) {
146 mminit_dprintk(MMINIT_WARNING, "pfnvalidation",
147 "End of range %lu -> %lu exceeds SPARSEMEM max %lu\n",
148 *start_pfn, *end_pfn, max_sparsemem_pfn);
149 WARN_ON_ONCE(1);
150 *end_pfn = max_sparsemem_pfn;
151 }
152 }
153
154 /*
155 * There are a number of times that we loop over NR_MEM_SECTIONS,
156 * looking for section_present() on each. But, when we have very
157 * large physical address spaces, NR_MEM_SECTIONS can also be
158 * very large which makes the loops quite long.
159 *
160 * Keeping track of this gives us an easy way to break out of
161 * those loops early.
162 */
163 unsigned long __highest_present_section_nr;
__section_mark_present(struct mem_section * ms,unsigned long section_nr)164 static void __section_mark_present(struct mem_section *ms,
165 unsigned long section_nr)
166 {
167 if (section_nr > __highest_present_section_nr)
168 __highest_present_section_nr = section_nr;
169
170 ms->section_mem_map |= SECTION_MARKED_PRESENT;
171 }
172
173 #define for_each_present_section_nr(start, section_nr) \
174 for (section_nr = next_present_section_nr(start-1); \
175 section_nr != -1; \
176 section_nr = next_present_section_nr(section_nr))
177
first_present_section_nr(void)178 static inline unsigned long first_present_section_nr(void)
179 {
180 return next_present_section_nr(-1);
181 }
182
183 #ifdef CONFIG_SPARSEMEM_VMEMMAP
subsection_mask_set(unsigned long * map,unsigned long pfn,unsigned long nr_pages)184 static void subsection_mask_set(unsigned long *map, unsigned long pfn,
185 unsigned long nr_pages)
186 {
187 int idx = subsection_map_index(pfn);
188 int end = subsection_map_index(pfn + nr_pages - 1);
189
190 bitmap_set(map, idx, end - idx + 1);
191 }
192
subsection_map_init(unsigned long pfn,unsigned long nr_pages)193 void __init subsection_map_init(unsigned long pfn, unsigned long nr_pages)
194 {
195 int end_sec = pfn_to_section_nr(pfn + nr_pages - 1);
196 unsigned long nr, start_sec = pfn_to_section_nr(pfn);
197
198 if (!nr_pages)
199 return;
200
201 for (nr = start_sec; nr <= end_sec; nr++) {
202 struct mem_section *ms;
203 unsigned long pfns;
204
205 pfns = min(nr_pages, PAGES_PER_SECTION
206 - (pfn & ~PAGE_SECTION_MASK));
207 ms = __nr_to_section(nr);
208 subsection_mask_set(ms->usage->subsection_map, pfn, pfns);
209
210 pr_debug("%s: sec: %lu pfns: %lu set(%d, %d)\n", __func__, nr,
211 pfns, subsection_map_index(pfn),
212 subsection_map_index(pfn + pfns - 1));
213
214 pfn += pfns;
215 nr_pages -= pfns;
216 }
217 }
218 #else
subsection_map_init(unsigned long pfn,unsigned long nr_pages)219 void __init subsection_map_init(unsigned long pfn, unsigned long nr_pages)
220 {
221 }
222 #endif
223
224 /* Record a memory area against a node. */
memory_present(int nid,unsigned long start,unsigned long end)225 static void __init memory_present(int nid, unsigned long start, unsigned long end)
226 {
227 unsigned long pfn;
228
229 start &= PAGE_SECTION_MASK;
230 mminit_validate_memmodel_limits(&start, &end);
231 for (pfn = start; pfn < end; pfn += PAGES_PER_SECTION) {
232 unsigned long section = pfn_to_section_nr(pfn);
233 struct mem_section *ms;
234
235 sparse_index_init(section, nid);
236 set_section_nid(section, nid);
237
238 ms = __nr_to_section(section);
239 if (!ms->section_mem_map) {
240 ms->section_mem_map = sparse_encode_early_nid(nid) |
241 SECTION_IS_ONLINE;
242 __section_mark_present(ms, section);
243 }
244 }
245 }
246
247 /*
248 * Mark all memblocks as present using memory_present().
249 * This is a convenience function that is useful to mark all of the systems
250 * memory as present during initialization.
251 */
memblocks_present(void)252 static void __init memblocks_present(void)
253 {
254 unsigned long start, end;
255 int i, nid;
256
257 #ifdef CONFIG_SPARSEMEM_EXTREME
258 if (unlikely(!mem_section)) {
259 unsigned long size, align;
260
261 size = sizeof(struct mem_section *) * NR_SECTION_ROOTS;
262 align = 1 << (INTERNODE_CACHE_SHIFT);
263 mem_section = memblock_alloc(size, align);
264 if (!mem_section)
265 panic("%s: Failed to allocate %lu bytes align=0x%lx\n",
266 __func__, size, align);
267 }
268 #endif
269
270 for_each_mem_pfn_range(i, MAX_NUMNODES, &start, &end, &nid)
271 memory_present(nid, start, end);
272 }
273
274 /*
275 * Subtle, we encode the real pfn into the mem_map such that
276 * the identity pfn - section_mem_map will return the actual
277 * physical page frame number.
278 */
sparse_encode_mem_map(struct page * mem_map,unsigned long pnum)279 static unsigned long sparse_encode_mem_map(struct page *mem_map, unsigned long pnum)
280 {
281 unsigned long coded_mem_map =
282 (unsigned long)(mem_map - (section_nr_to_pfn(pnum)));
283 BUILD_BUG_ON(SECTION_MAP_LAST_BIT > PFN_SECTION_SHIFT);
284 BUG_ON(coded_mem_map & ~SECTION_MAP_MASK);
285 return coded_mem_map;
286 }
287
288 #ifdef CONFIG_MEMORY_HOTPLUG
289 /*
290 * Decode mem_map from the coded memmap
291 */
sparse_decode_mem_map(unsigned long coded_mem_map,unsigned long pnum)292 struct page *sparse_decode_mem_map(unsigned long coded_mem_map, unsigned long pnum)
293 {
294 /* mask off the extra low bits of information */
295 coded_mem_map &= SECTION_MAP_MASK;
296 return ((struct page *)coded_mem_map) + section_nr_to_pfn(pnum);
297 }
298 #endif /* CONFIG_MEMORY_HOTPLUG */
299
sparse_init_one_section(struct mem_section * ms,unsigned long pnum,struct page * mem_map,struct mem_section_usage * usage,unsigned long flags)300 static void __meminit sparse_init_one_section(struct mem_section *ms,
301 unsigned long pnum, struct page *mem_map,
302 struct mem_section_usage *usage, unsigned long flags)
303 {
304 ms->section_mem_map &= ~SECTION_MAP_MASK;
305 ms->section_mem_map |= sparse_encode_mem_map(mem_map, pnum)
306 | SECTION_HAS_MEM_MAP | flags;
307 ms->usage = usage;
308 }
309
usemap_size(void)310 static unsigned long usemap_size(void)
311 {
312 return BITS_TO_LONGS(SECTION_BLOCKFLAGS_BITS) * sizeof(unsigned long);
313 }
314
mem_section_usage_size(void)315 size_t mem_section_usage_size(void)
316 {
317 return sizeof(struct mem_section_usage) + usemap_size();
318 }
319
320 #ifdef CONFIG_MEMORY_HOTREMOVE
pgdat_to_phys(struct pglist_data * pgdat)321 static inline phys_addr_t pgdat_to_phys(struct pglist_data *pgdat)
322 {
323 #ifndef CONFIG_NUMA
324 VM_BUG_ON(pgdat != &contig_page_data);
325 return __pa_symbol(&contig_page_data);
326 #else
327 return __pa(pgdat);
328 #endif
329 }
330
331 static struct mem_section_usage * __init
sparse_early_usemaps_alloc_pgdat_section(struct pglist_data * pgdat,unsigned long size)332 sparse_early_usemaps_alloc_pgdat_section(struct pglist_data *pgdat,
333 unsigned long size)
334 {
335 struct mem_section_usage *usage;
336 unsigned long goal, limit;
337 int nid;
338 /*
339 * A page may contain usemaps for other sections preventing the
340 * page being freed and making a section unremovable while
341 * other sections referencing the usemap remain active. Similarly,
342 * a pgdat can prevent a section being removed. If section A
343 * contains a pgdat and section B contains the usemap, both
344 * sections become inter-dependent. This allocates usemaps
345 * from the same section as the pgdat where possible to avoid
346 * this problem.
347 */
348 goal = pgdat_to_phys(pgdat) & (PAGE_SECTION_MASK << PAGE_SHIFT);
349 limit = goal + (1UL << PA_SECTION_SHIFT);
350 nid = early_pfn_to_nid(goal >> PAGE_SHIFT);
351 again:
352 usage = memblock_alloc_try_nid(size, SMP_CACHE_BYTES, goal, limit, nid);
353 if (!usage && limit) {
354 limit = 0;
355 goto again;
356 }
357 return usage;
358 }
359
check_usemap_section_nr(int nid,struct mem_section_usage * usage)360 static void __init check_usemap_section_nr(int nid,
361 struct mem_section_usage *usage)
362 {
363 unsigned long usemap_snr, pgdat_snr;
364 static unsigned long old_usemap_snr;
365 static unsigned long old_pgdat_snr;
366 struct pglist_data *pgdat = NODE_DATA(nid);
367 int usemap_nid;
368
369 /* First call */
370 if (!old_usemap_snr) {
371 old_usemap_snr = NR_MEM_SECTIONS;
372 old_pgdat_snr = NR_MEM_SECTIONS;
373 }
374
375 usemap_snr = pfn_to_section_nr(__pa(usage) >> PAGE_SHIFT);
376 pgdat_snr = pfn_to_section_nr(pgdat_to_phys(pgdat) >> PAGE_SHIFT);
377 if (usemap_snr == pgdat_snr)
378 return;
379
380 if (old_usemap_snr == usemap_snr && old_pgdat_snr == pgdat_snr)
381 /* skip redundant message */
382 return;
383
384 old_usemap_snr = usemap_snr;
385 old_pgdat_snr = pgdat_snr;
386
387 usemap_nid = sparse_early_nid(__nr_to_section(usemap_snr));
388 if (usemap_nid != nid) {
389 pr_info("node %d must be removed before remove section %ld\n",
390 nid, usemap_snr);
391 return;
392 }
393 /*
394 * There is a circular dependency.
395 * Some platforms allow un-removable section because they will just
396 * gather other removable sections for dynamic partitioning.
397 * Just notify un-removable section's number here.
398 */
399 pr_info("Section %ld and %ld (node %d) have a circular dependency on usemap and pgdat allocations\n",
400 usemap_snr, pgdat_snr, nid);
401 }
402 #else
403 static struct mem_section_usage * __init
sparse_early_usemaps_alloc_pgdat_section(struct pglist_data * pgdat,unsigned long size)404 sparse_early_usemaps_alloc_pgdat_section(struct pglist_data *pgdat,
405 unsigned long size)
406 {
407 return memblock_alloc_node(size, SMP_CACHE_BYTES, pgdat->node_id);
408 }
409
check_usemap_section_nr(int nid,struct mem_section_usage * usage)410 static void __init check_usemap_section_nr(int nid,
411 struct mem_section_usage *usage)
412 {
413 }
414 #endif /* CONFIG_MEMORY_HOTREMOVE */
415
416 #ifdef CONFIG_SPARSEMEM_VMEMMAP
section_map_size(void)417 static unsigned long __init section_map_size(void)
418 {
419 return ALIGN(sizeof(struct page) * PAGES_PER_SECTION, PMD_SIZE);
420 }
421
422 #else
section_map_size(void)423 static unsigned long __init section_map_size(void)
424 {
425 return PAGE_ALIGN(sizeof(struct page) * PAGES_PER_SECTION);
426 }
427
__populate_section_memmap(unsigned long pfn,unsigned long nr_pages,int nid,struct vmem_altmap * altmap,struct dev_pagemap * pgmap)428 struct page __init *__populate_section_memmap(unsigned long pfn,
429 unsigned long nr_pages, int nid, struct vmem_altmap *altmap,
430 struct dev_pagemap *pgmap)
431 {
432 unsigned long size = section_map_size();
433 struct page *map = sparse_buffer_alloc(size);
434 phys_addr_t addr = __pa(MAX_DMA_ADDRESS);
435
436 if (map)
437 return map;
438
439 map = memmap_alloc(size, size, addr, nid, false);
440 if (!map)
441 panic("%s: Failed to allocate %lu bytes align=0x%lx nid=%d from=%pa\n",
442 __func__, size, PAGE_SIZE, nid, &addr);
443
444 return map;
445 }
446 #endif /* !CONFIG_SPARSEMEM_VMEMMAP */
447
448 static void *sparsemap_buf __meminitdata;
449 static void *sparsemap_buf_end __meminitdata;
450
sparse_buffer_free(unsigned long size)451 static inline void __meminit sparse_buffer_free(unsigned long size)
452 {
453 WARN_ON(!sparsemap_buf || size == 0);
454 memblock_free(sparsemap_buf, size);
455 }
456
sparse_buffer_init(unsigned long size,int nid)457 static void __init sparse_buffer_init(unsigned long size, int nid)
458 {
459 phys_addr_t addr = __pa(MAX_DMA_ADDRESS);
460 WARN_ON(sparsemap_buf); /* forgot to call sparse_buffer_fini()? */
461 /*
462 * Pre-allocated buffer is mainly used by __populate_section_memmap
463 * and we want it to be properly aligned to the section size - this is
464 * especially the case for VMEMMAP which maps memmap to PMDs
465 */
466 sparsemap_buf = memmap_alloc(size, section_map_size(), addr, nid, true);
467 sparsemap_buf_end = sparsemap_buf + size;
468 }
469
sparse_buffer_fini(void)470 static void __init sparse_buffer_fini(void)
471 {
472 unsigned long size = sparsemap_buf_end - sparsemap_buf;
473
474 if (sparsemap_buf && size > 0)
475 sparse_buffer_free(size);
476 sparsemap_buf = NULL;
477 }
478
sparse_buffer_alloc(unsigned long size)479 void * __meminit sparse_buffer_alloc(unsigned long size)
480 {
481 void *ptr = NULL;
482
483 if (sparsemap_buf) {
484 ptr = (void *) roundup((unsigned long)sparsemap_buf, size);
485 if (ptr + size > sparsemap_buf_end)
486 ptr = NULL;
487 else {
488 /* Free redundant aligned space */
489 if ((unsigned long)(ptr - sparsemap_buf) > 0)
490 sparse_buffer_free((unsigned long)(ptr - sparsemap_buf));
491 sparsemap_buf = ptr + size;
492 }
493 }
494 return ptr;
495 }
496
vmemmap_populate_print_last(void)497 void __weak __meminit vmemmap_populate_print_last(void)
498 {
499 }
500
501 /*
502 * Initialize sparse on a specific node. The node spans [pnum_begin, pnum_end)
503 * And number of present sections in this node is map_count.
504 */
sparse_init_nid(int nid,unsigned long pnum_begin,unsigned long pnum_end,unsigned long map_count)505 static void __init sparse_init_nid(int nid, unsigned long pnum_begin,
506 unsigned long pnum_end,
507 unsigned long map_count)
508 {
509 struct mem_section_usage *usage;
510 unsigned long pnum;
511 struct page *map;
512
513 usage = sparse_early_usemaps_alloc_pgdat_section(NODE_DATA(nid),
514 mem_section_usage_size() * map_count);
515 if (!usage) {
516 pr_err("%s: node[%d] usemap allocation failed", __func__, nid);
517 goto failed;
518 }
519 sparse_buffer_init(map_count * section_map_size(), nid);
520 for_each_present_section_nr(pnum_begin, pnum) {
521 unsigned long pfn = section_nr_to_pfn(pnum);
522
523 if (pnum >= pnum_end)
524 break;
525
526 map = __populate_section_memmap(pfn, PAGES_PER_SECTION,
527 nid, NULL, NULL);
528 if (!map) {
529 pr_err("%s: node[%d] memory map backing failed. Some memory will not be available.",
530 __func__, nid);
531 pnum_begin = pnum;
532 sparse_buffer_fini();
533 goto failed;
534 }
535 check_usemap_section_nr(nid, usage);
536 sparse_init_one_section(__nr_to_section(pnum), pnum, map, usage,
537 SECTION_IS_EARLY);
538 usage = (void *) usage + mem_section_usage_size();
539 }
540 sparse_buffer_fini();
541 return;
542 failed:
543 /* We failed to allocate, mark all the following pnums as not present */
544 for_each_present_section_nr(pnum_begin, pnum) {
545 struct mem_section *ms;
546
547 if (pnum >= pnum_end)
548 break;
549 ms = __nr_to_section(pnum);
550 ms->section_mem_map = 0;
551 }
552 }
553
554 /*
555 * Allocate the accumulated non-linear sections, allocate a mem_map
556 * for each and record the physical to section mapping.
557 */
sparse_init(void)558 void __init sparse_init(void)
559 {
560 unsigned long pnum_end, pnum_begin, map_count = 1;
561 int nid_begin;
562
563 /* see include/linux/mmzone.h 'struct mem_section' definition */
564 BUILD_BUG_ON(!is_power_of_2(sizeof(struct mem_section)));
565 memblocks_present();
566
567 pnum_begin = first_present_section_nr();
568 nid_begin = sparse_early_nid(__nr_to_section(pnum_begin));
569
570 /* Setup pageblock_order for HUGETLB_PAGE_SIZE_VARIABLE */
571 set_pageblock_order();
572
573 for_each_present_section_nr(pnum_begin + 1, pnum_end) {
574 int nid = sparse_early_nid(__nr_to_section(pnum_end));
575
576 if (nid == nid_begin) {
577 map_count++;
578 continue;
579 }
580 /* Init node with sections in range [pnum_begin, pnum_end) */
581 sparse_init_nid(nid_begin, pnum_begin, pnum_end, map_count);
582 nid_begin = nid;
583 pnum_begin = pnum_end;
584 map_count = 1;
585 }
586 /* cover the last node */
587 sparse_init_nid(nid_begin, pnum_begin, pnum_end, map_count);
588 vmemmap_populate_print_last();
589 }
590
591 #ifdef CONFIG_MEMORY_HOTPLUG
592
593 /* Mark all memory sections within the pfn range as online */
online_mem_sections(unsigned long start_pfn,unsigned long end_pfn)594 void online_mem_sections(unsigned long start_pfn, unsigned long end_pfn)
595 {
596 unsigned long pfn;
597
598 for (pfn = start_pfn; pfn < end_pfn; pfn += PAGES_PER_SECTION) {
599 unsigned long section_nr = pfn_to_section_nr(pfn);
600 struct mem_section *ms;
601
602 /* onlining code should never touch invalid ranges */
603 if (WARN_ON(!valid_section_nr(section_nr)))
604 continue;
605
606 ms = __nr_to_section(section_nr);
607 ms->section_mem_map |= SECTION_IS_ONLINE;
608 }
609 }
610
611 /* Mark all memory sections within the pfn range as offline */
offline_mem_sections(unsigned long start_pfn,unsigned long end_pfn)612 void offline_mem_sections(unsigned long start_pfn, unsigned long end_pfn)
613 {
614 unsigned long pfn;
615
616 for (pfn = start_pfn; pfn < end_pfn; pfn += PAGES_PER_SECTION) {
617 unsigned long section_nr = pfn_to_section_nr(pfn);
618 struct mem_section *ms;
619
620 /*
621 * TODO this needs some double checking. Offlining code makes
622 * sure to check pfn_valid but those checks might be just bogus
623 */
624 if (WARN_ON(!valid_section_nr(section_nr)))
625 continue;
626
627 ms = __nr_to_section(section_nr);
628 ms->section_mem_map &= ~SECTION_IS_ONLINE;
629 }
630 }
631
632 #ifdef CONFIG_SPARSEMEM_VMEMMAP
populate_section_memmap(unsigned long pfn,unsigned long nr_pages,int nid,struct vmem_altmap * altmap,struct dev_pagemap * pgmap)633 static struct page * __meminit populate_section_memmap(unsigned long pfn,
634 unsigned long nr_pages, int nid, struct vmem_altmap *altmap,
635 struct dev_pagemap *pgmap)
636 {
637 return __populate_section_memmap(pfn, nr_pages, nid, altmap, pgmap);
638 }
639
depopulate_section_memmap(unsigned long pfn,unsigned long nr_pages,struct vmem_altmap * altmap)640 static void depopulate_section_memmap(unsigned long pfn, unsigned long nr_pages,
641 struct vmem_altmap *altmap)
642 {
643 unsigned long start = (unsigned long) pfn_to_page(pfn);
644 unsigned long end = start + nr_pages * sizeof(struct page);
645
646 vmemmap_free(start, end, altmap);
647 }
free_map_bootmem(struct page * memmap)648 static void free_map_bootmem(struct page *memmap)
649 {
650 unsigned long start = (unsigned long)memmap;
651 unsigned long end = (unsigned long)(memmap + PAGES_PER_SECTION);
652
653 vmemmap_free(start, end, NULL);
654 }
655
clear_subsection_map(unsigned long pfn,unsigned long nr_pages)656 static int clear_subsection_map(unsigned long pfn, unsigned long nr_pages)
657 {
658 DECLARE_BITMAP(map, SUBSECTIONS_PER_SECTION) = { 0 };
659 DECLARE_BITMAP(tmp, SUBSECTIONS_PER_SECTION) = { 0 };
660 struct mem_section *ms = __pfn_to_section(pfn);
661 unsigned long *subsection_map = ms->usage
662 ? &ms->usage->subsection_map[0] : NULL;
663
664 subsection_mask_set(map, pfn, nr_pages);
665 if (subsection_map)
666 bitmap_and(tmp, map, subsection_map, SUBSECTIONS_PER_SECTION);
667
668 if (WARN(!subsection_map || !bitmap_equal(tmp, map, SUBSECTIONS_PER_SECTION),
669 "section already deactivated (%#lx + %ld)\n",
670 pfn, nr_pages))
671 return -EINVAL;
672
673 bitmap_xor(subsection_map, map, subsection_map, SUBSECTIONS_PER_SECTION);
674 return 0;
675 }
676
is_subsection_map_empty(struct mem_section * ms)677 static bool is_subsection_map_empty(struct mem_section *ms)
678 {
679 return bitmap_empty(&ms->usage->subsection_map[0],
680 SUBSECTIONS_PER_SECTION);
681 }
682
fill_subsection_map(unsigned long pfn,unsigned long nr_pages)683 static int fill_subsection_map(unsigned long pfn, unsigned long nr_pages)
684 {
685 struct mem_section *ms = __pfn_to_section(pfn);
686 DECLARE_BITMAP(map, SUBSECTIONS_PER_SECTION) = { 0 };
687 unsigned long *subsection_map;
688 int rc = 0;
689
690 subsection_mask_set(map, pfn, nr_pages);
691
692 subsection_map = &ms->usage->subsection_map[0];
693
694 if (bitmap_empty(map, SUBSECTIONS_PER_SECTION))
695 rc = -EINVAL;
696 else if (bitmap_intersects(map, subsection_map, SUBSECTIONS_PER_SECTION))
697 rc = -EEXIST;
698 else
699 bitmap_or(subsection_map, map, subsection_map,
700 SUBSECTIONS_PER_SECTION);
701
702 return rc;
703 }
704 #else
populate_section_memmap(unsigned long pfn,unsigned long nr_pages,int nid,struct vmem_altmap * altmap,struct dev_pagemap * pgmap)705 static struct page * __meminit populate_section_memmap(unsigned long pfn,
706 unsigned long nr_pages, int nid, struct vmem_altmap *altmap,
707 struct dev_pagemap *pgmap)
708 {
709 return kvmalloc_node(array_size(sizeof(struct page),
710 PAGES_PER_SECTION), GFP_KERNEL, nid);
711 }
712
depopulate_section_memmap(unsigned long pfn,unsigned long nr_pages,struct vmem_altmap * altmap)713 static void depopulate_section_memmap(unsigned long pfn, unsigned long nr_pages,
714 struct vmem_altmap *altmap)
715 {
716 kvfree(pfn_to_page(pfn));
717 }
718
free_map_bootmem(struct page * memmap)719 static void free_map_bootmem(struct page *memmap)
720 {
721 unsigned long maps_section_nr, removing_section_nr, i;
722 unsigned long magic, nr_pages;
723 struct page *page = virt_to_page(memmap);
724
725 nr_pages = PAGE_ALIGN(PAGES_PER_SECTION * sizeof(struct page))
726 >> PAGE_SHIFT;
727
728 for (i = 0; i < nr_pages; i++, page++) {
729 magic = page->index;
730
731 BUG_ON(magic == NODE_INFO);
732
733 maps_section_nr = pfn_to_section_nr(page_to_pfn(page));
734 removing_section_nr = page_private(page);
735
736 /*
737 * When this function is called, the removing section is
738 * logical offlined state. This means all pages are isolated
739 * from page allocator. If removing section's memmap is placed
740 * on the same section, it must not be freed.
741 * If it is freed, page allocator may allocate it which will
742 * be removed physically soon.
743 */
744 if (maps_section_nr != removing_section_nr)
745 put_page_bootmem(page);
746 }
747 }
748
clear_subsection_map(unsigned long pfn,unsigned long nr_pages)749 static int clear_subsection_map(unsigned long pfn, unsigned long nr_pages)
750 {
751 return 0;
752 }
753
is_subsection_map_empty(struct mem_section * ms)754 static bool is_subsection_map_empty(struct mem_section *ms)
755 {
756 return true;
757 }
758
fill_subsection_map(unsigned long pfn,unsigned long nr_pages)759 static int fill_subsection_map(unsigned long pfn, unsigned long nr_pages)
760 {
761 return 0;
762 }
763 #endif /* CONFIG_SPARSEMEM_VMEMMAP */
764
765 /*
766 * To deactivate a memory region, there are 3 cases to handle across
767 * two configurations (SPARSEMEM_VMEMMAP={y,n}):
768 *
769 * 1. deactivation of a partial hot-added section (only possible in
770 * the SPARSEMEM_VMEMMAP=y case).
771 * a) section was present at memory init.
772 * b) section was hot-added post memory init.
773 * 2. deactivation of a complete hot-added section.
774 * 3. deactivation of a complete section from memory init.
775 *
776 * For 1, when subsection_map does not empty we will not be freeing the
777 * usage map, but still need to free the vmemmap range.
778 *
779 * For 2 and 3, the SPARSEMEM_VMEMMAP={y,n} cases are unified
780 */
section_deactivate(unsigned long pfn,unsigned long nr_pages,struct vmem_altmap * altmap)781 static void section_deactivate(unsigned long pfn, unsigned long nr_pages,
782 struct vmem_altmap *altmap)
783 {
784 struct mem_section *ms = __pfn_to_section(pfn);
785 bool section_is_early = early_section(ms);
786 struct page *memmap = NULL;
787 bool empty;
788
789 if (clear_subsection_map(pfn, nr_pages))
790 return;
791
792 empty = is_subsection_map_empty(ms);
793 if (empty) {
794 unsigned long section_nr = pfn_to_section_nr(pfn);
795
796 /*
797 * Mark the section invalid so that valid_section()
798 * return false. This prevents code from dereferencing
799 * ms->usage array.
800 */
801 ms->section_mem_map &= ~SECTION_HAS_MEM_MAP;
802
803 /*
804 * When removing an early section, the usage map is kept (as the
805 * usage maps of other sections fall into the same page). It
806 * will be re-used when re-adding the section - which is then no
807 * longer an early section. If the usage map is PageReserved, it
808 * was allocated during boot.
809 */
810 if (!PageReserved(virt_to_page(ms->usage))) {
811 kfree_rcu(ms->usage, rcu);
812 WRITE_ONCE(ms->usage, NULL);
813 }
814 memmap = sparse_decode_mem_map(ms->section_mem_map, section_nr);
815 }
816
817 /*
818 * The memmap of early sections is always fully populated. See
819 * section_activate() and pfn_valid() .
820 */
821 if (!section_is_early)
822 depopulate_section_memmap(pfn, nr_pages, altmap);
823 else if (memmap)
824 free_map_bootmem(memmap);
825
826 if (empty)
827 ms->section_mem_map = (unsigned long)NULL;
828 }
829
section_activate(int nid,unsigned long pfn,unsigned long nr_pages,struct vmem_altmap * altmap,struct dev_pagemap * pgmap)830 static struct page * __meminit section_activate(int nid, unsigned long pfn,
831 unsigned long nr_pages, struct vmem_altmap *altmap,
832 struct dev_pagemap *pgmap)
833 {
834 struct mem_section *ms = __pfn_to_section(pfn);
835 struct mem_section_usage *usage = NULL;
836 struct page *memmap;
837 int rc;
838
839 if (!ms->usage) {
840 usage = kzalloc(mem_section_usage_size(), GFP_KERNEL);
841 if (!usage)
842 return ERR_PTR(-ENOMEM);
843 ms->usage = usage;
844 }
845
846 rc = fill_subsection_map(pfn, nr_pages);
847 if (rc) {
848 if (usage)
849 ms->usage = NULL;
850 kfree(usage);
851 return ERR_PTR(rc);
852 }
853
854 /*
855 * The early init code does not consider partially populated
856 * initial sections, it simply assumes that memory will never be
857 * referenced. If we hot-add memory into such a section then we
858 * do not need to populate the memmap and can simply reuse what
859 * is already there.
860 */
861 if (nr_pages < PAGES_PER_SECTION && early_section(ms))
862 return pfn_to_page(pfn);
863
864 memmap = populate_section_memmap(pfn, nr_pages, nid, altmap, pgmap);
865 if (!memmap) {
866 section_deactivate(pfn, nr_pages, altmap);
867 return ERR_PTR(-ENOMEM);
868 }
869
870 return memmap;
871 }
872
873 /**
874 * sparse_add_section - add a memory section, or populate an existing one
875 * @nid: The node to add section on
876 * @start_pfn: start pfn of the memory range
877 * @nr_pages: number of pfns to add in the section
878 * @altmap: alternate pfns to allocate the memmap backing store
879 * @pgmap: alternate compound page geometry for devmap mappings
880 *
881 * This is only intended for hotplug.
882 *
883 * Note that only VMEMMAP supports sub-section aligned hotplug,
884 * the proper alignment and size are gated by check_pfn_span().
885 *
886 *
887 * Return:
888 * * 0 - On success.
889 * * -EEXIST - Section has been present.
890 * * -ENOMEM - Out of memory.
891 */
sparse_add_section(int nid,unsigned long start_pfn,unsigned long nr_pages,struct vmem_altmap * altmap,struct dev_pagemap * pgmap)892 int __meminit sparse_add_section(int nid, unsigned long start_pfn,
893 unsigned long nr_pages, struct vmem_altmap *altmap,
894 struct dev_pagemap *pgmap)
895 {
896 unsigned long section_nr = pfn_to_section_nr(start_pfn);
897 struct mem_section *ms;
898 struct page *memmap;
899 int ret;
900
901 ret = sparse_index_init(section_nr, nid);
902 if (ret < 0)
903 return ret;
904
905 memmap = section_activate(nid, start_pfn, nr_pages, altmap, pgmap);
906 if (IS_ERR(memmap))
907 return PTR_ERR(memmap);
908
909 /*
910 * Poison uninitialized struct pages in order to catch invalid flags
911 * combinations.
912 */
913 if (!altmap || !altmap->inaccessible)
914 page_init_poison(memmap, sizeof(struct page) * nr_pages);
915
916 ms = __nr_to_section(section_nr);
917 set_section_nid(section_nr, nid);
918 __section_mark_present(ms, section_nr);
919
920 /* Align memmap to section boundary in the subsection case */
921 if (section_nr_to_pfn(section_nr) != start_pfn)
922 memmap = pfn_to_page(section_nr_to_pfn(section_nr));
923 sparse_init_one_section(ms, section_nr, memmap, ms->usage, 0);
924
925 return 0;
926 }
927
sparse_remove_section(unsigned long pfn,unsigned long nr_pages,struct vmem_altmap * altmap)928 void sparse_remove_section(unsigned long pfn, unsigned long nr_pages,
929 struct vmem_altmap *altmap)
930 {
931 struct mem_section *ms = __pfn_to_section(pfn);
932
933 if (WARN_ON_ONCE(!valid_section(ms)))
934 return;
935
936 section_deactivate(pfn, nr_pages, altmap);
937 }
938 #endif /* CONFIG_MEMORY_HOTPLUG */
939