]>
Commit | Line | Data |
---|---|---|
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 | ||
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 | ||
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 */ | |
57 | static inline void set_section_nid(unsigned long section_nr, int nid) | |
58 | { | |
59 | } | |
60 | #endif | |
61 | ||
62 | #ifdef CONFIG_SPARSEMEM_EXTREME | |
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 | ||
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 */ | |
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 | */ | |
118 | static inline unsigned long sparse_encode_early_nid(int nid) | |
119 | { | |
120 | return ((unsigned long)nid << SECTION_NID_SHIFT); | |
121 | } | |
122 | ||
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 */ | |
129 | 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; | |
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 <= __highest_present_section_nr)); \ | |
177 | section_nr = next_present_section_nr(section_nr)) | |
178 | ||
179 | static inline unsigned long first_present_section_nr(void) | |
180 | { | |
181 | return next_present_section_nr(-1); | |
182 | } | |
183 | ||
184 | #ifdef CONFIG_SPARSEMEM_VMEMMAP | |
185 | static void subsection_mask_set(unsigned long *map, unsigned long pfn, | |
186 | unsigned long nr_pages) | |
187 | { | |
188 | int idx = subsection_map_index(pfn); | |
189 | int end = subsection_map_index(pfn + nr_pages - 1); | |
190 | ||
191 | bitmap_set(map, idx, end - idx + 1); | |
192 | } | |
193 | ||
194 | void __init subsection_map_init(unsigned long pfn, unsigned long nr_pages) | |
195 | { | |
196 | int end_sec = pfn_to_section_nr(pfn + nr_pages - 1); | |
197 | unsigned long nr, start_sec = pfn_to_section_nr(pfn); | |
198 | ||
199 | if (!nr_pages) | |
200 | return; | |
201 | ||
202 | for (nr = start_sec; nr <= end_sec; nr++) { | |
203 | struct mem_section *ms; | |
204 | unsigned long pfns; | |
205 | ||
206 | pfns = min(nr_pages, PAGES_PER_SECTION | |
207 | - (pfn & ~PAGE_SECTION_MASK)); | |
208 | ms = __nr_to_section(nr); | |
209 | subsection_mask_set(ms->usage->subsection_map, pfn, pfns); | |
210 | ||
211 | pr_debug("%s: sec: %lu pfns: %lu set(%d, %d)\n", __func__, nr, | |
212 | pfns, subsection_map_index(pfn), | |
213 | subsection_map_index(pfn + pfns - 1)); | |
214 | ||
215 | pfn += pfns; | |
216 | nr_pages -= pfns; | |
217 | } | |
218 | } | |
219 | #else | |
220 | void __init subsection_map_init(unsigned long pfn, unsigned long nr_pages) | |
221 | { | |
222 | } | |
223 | #endif | |
224 | ||
225 | /* Record a memory area against a node. */ | |
226 | static void __init memory_present(int nid, unsigned long start, unsigned long end) | |
227 | { | |
228 | unsigned long pfn; | |
229 | ||
230 | #ifdef CONFIG_SPARSEMEM_EXTREME | |
231 | if (unlikely(!mem_section)) { | |
232 | unsigned long size, align; | |
233 | ||
234 | size = sizeof(struct mem_section *) * NR_SECTION_ROOTS; | |
235 | align = 1 << (INTERNODE_CACHE_SHIFT); | |
236 | mem_section = memblock_alloc(size, align); | |
237 | if (!mem_section) | |
238 | panic("%s: Failed to allocate %lu bytes align=0x%lx\n", | |
239 | __func__, size, align); | |
240 | } | |
241 | #endif | |
242 | ||
243 | start &= PAGE_SECTION_MASK; | |
244 | mminit_validate_memmodel_limits(&start, &end); | |
245 | for (pfn = start; pfn < end; pfn += PAGES_PER_SECTION) { | |
246 | unsigned long section = pfn_to_section_nr(pfn); | |
247 | struct mem_section *ms; | |
248 | ||
249 | sparse_index_init(section, nid); | |
250 | set_section_nid(section, nid); | |
251 | ||
252 | ms = __nr_to_section(section); | |
253 | if (!ms->section_mem_map) { | |
254 | ms->section_mem_map = sparse_encode_early_nid(nid) | | |
255 | SECTION_IS_ONLINE; | |
256 | __section_mark_present(ms, section); | |
257 | } | |
258 | } | |
259 | } | |
260 | ||
261 | /* | |
262 | * Mark all memblocks as present using memory_present(). | |
263 | * This is a convenience function that is useful to mark all of the systems | |
264 | * memory as present during initialization. | |
265 | */ | |
266 | static void __init memblocks_present(void) | |
267 | { | |
268 | unsigned long start, end; | |
269 | int i, nid; | |
270 | ||
271 | for_each_mem_pfn_range(i, MAX_NUMNODES, &start, &end, &nid) | |
272 | memory_present(nid, start, end); | |
273 | } | |
274 | ||
275 | /* | |
276 | * Subtle, we encode the real pfn into the mem_map such that | |
277 | * the identity pfn - section_mem_map will return the actual | |
278 | * physical page frame number. | |
279 | */ | |
280 | static unsigned long sparse_encode_mem_map(struct page *mem_map, unsigned long pnum) | |
281 | { | |
282 | unsigned long coded_mem_map = | |
283 | (unsigned long)(mem_map - (section_nr_to_pfn(pnum))); | |
284 | BUILD_BUG_ON(SECTION_MAP_LAST_BIT > (1UL<<PFN_SECTION_SHIFT)); | |
285 | BUG_ON(coded_mem_map & ~SECTION_MAP_MASK); | |
286 | return coded_mem_map; | |
287 | } | |
288 | ||
289 | #ifdef CONFIG_MEMORY_HOTPLUG | |
290 | /* | |
291 | * Decode mem_map from the coded memmap | |
292 | */ | |
293 | struct page *sparse_decode_mem_map(unsigned long coded_mem_map, unsigned long pnum) | |
294 | { | |
295 | /* mask off the extra low bits of information */ | |
296 | coded_mem_map &= SECTION_MAP_MASK; | |
297 | return ((struct page *)coded_mem_map) + section_nr_to_pfn(pnum); | |
298 | } | |
299 | #endif /* CONFIG_MEMORY_HOTPLUG */ | |
300 | ||
301 | static void __meminit sparse_init_one_section(struct mem_section *ms, | |
302 | unsigned long pnum, struct page *mem_map, | |
303 | struct mem_section_usage *usage, unsigned long flags) | |
304 | { | |
305 | ms->section_mem_map &= ~SECTION_MAP_MASK; | |
306 | ms->section_mem_map |= sparse_encode_mem_map(mem_map, pnum) | |
307 | | SECTION_HAS_MEM_MAP | flags; | |
308 | ms->usage = usage; | |
309 | } | |
310 | ||
311 | static unsigned long usemap_size(void) | |
312 | { | |
313 | return BITS_TO_LONGS(SECTION_BLOCKFLAGS_BITS) * sizeof(unsigned long); | |
314 | } | |
315 | ||
316 | size_t mem_section_usage_size(void) | |
317 | { | |
318 | return sizeof(struct mem_section_usage) + usemap_size(); | |
319 | } | |
320 | ||
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 | #ifdef CONFIG_MEMORY_HOTREMOVE | |
332 | static struct mem_section_usage * __init | |
333 | sparse_early_usemaps_alloc_pgdat_section(struct pglist_data *pgdat, | |
334 | unsigned long size) | |
335 | { | |
336 | struct mem_section_usage *usage; | |
337 | unsigned long goal, limit; | |
338 | int nid; | |
339 | /* | |
340 | * A page may contain usemaps for other sections preventing the | |
341 | * page being freed and making a section unremovable while | |
342 | * other sections referencing the usemap remain active. Similarly, | |
343 | * a pgdat can prevent a section being removed. If section A | |
344 | * contains a pgdat and section B contains the usemap, both | |
345 | * sections become inter-dependent. This allocates usemaps | |
346 | * from the same section as the pgdat where possible to avoid | |
347 | * this problem. | |
348 | */ | |
349 | goal = pgdat_to_phys(pgdat) & (PAGE_SECTION_MASK << PAGE_SHIFT); | |
350 | limit = goal + (1UL << PA_SECTION_SHIFT); | |
351 | nid = early_pfn_to_nid(goal >> PAGE_SHIFT); | |
352 | again: | |
353 | usage = memblock_alloc_try_nid(size, SMP_CACHE_BYTES, goal, limit, nid); | |
354 | if (!usage && limit) { | |
355 | limit = 0; | |
356 | goto again; | |
357 | } | |
358 | return usage; | |
359 | } | |
360 | ||
361 | static void __init check_usemap_section_nr(int nid, | |
362 | struct mem_section_usage *usage) | |
363 | { | |
364 | unsigned long usemap_snr, pgdat_snr; | |
365 | static unsigned long old_usemap_snr; | |
366 | static unsigned long old_pgdat_snr; | |
367 | struct pglist_data *pgdat = NODE_DATA(nid); | |
368 | int usemap_nid; | |
369 | ||
370 | /* First call */ | |
371 | if (!old_usemap_snr) { | |
372 | old_usemap_snr = NR_MEM_SECTIONS; | |
373 | old_pgdat_snr = NR_MEM_SECTIONS; | |
374 | } | |
375 | ||
376 | usemap_snr = pfn_to_section_nr(__pa(usage) >> PAGE_SHIFT); | |
377 | pgdat_snr = pfn_to_section_nr(pgdat_to_phys(pgdat) >> PAGE_SHIFT); | |
378 | if (usemap_snr == pgdat_snr) | |
379 | return; | |
380 | ||
381 | if (old_usemap_snr == usemap_snr && old_pgdat_snr == pgdat_snr) | |
382 | /* skip redundant message */ | |
383 | return; | |
384 | ||
385 | old_usemap_snr = usemap_snr; | |
386 | old_pgdat_snr = pgdat_snr; | |
387 | ||
388 | usemap_nid = sparse_early_nid(__nr_to_section(usemap_snr)); | |
389 | if (usemap_nid != nid) { | |
390 | pr_info("node %d must be removed before remove section %ld\n", | |
391 | nid, usemap_snr); | |
392 | return; | |
393 | } | |
394 | /* | |
395 | * There is a circular dependency. | |
396 | * Some platforms allow un-removable section because they will just | |
397 | * gather other removable sections for dynamic partitioning. | |
398 | * Just notify un-removable section's number here. | |
399 | */ | |
400 | pr_info("Section %ld and %ld (node %d) have a circular dependency on usemap and pgdat allocations\n", | |
401 | usemap_snr, pgdat_snr, nid); | |
402 | } | |
403 | #else | |
404 | static struct mem_section_usage * __init | |
405 | sparse_early_usemaps_alloc_pgdat_section(struct pglist_data *pgdat, | |
406 | unsigned long size) | |
407 | { | |
408 | return memblock_alloc_node(size, SMP_CACHE_BYTES, pgdat->node_id); | |
409 | } | |
410 | ||
411 | static void __init check_usemap_section_nr(int nid, | |
412 | struct mem_section_usage *usage) | |
413 | { | |
414 | } | |
415 | #endif /* CONFIG_MEMORY_HOTREMOVE */ | |
416 | ||
417 | #ifdef CONFIG_SPARSEMEM_VMEMMAP | |
418 | static unsigned long __init section_map_size(void) | |
419 | { | |
420 | return ALIGN(sizeof(struct page) * PAGES_PER_SECTION, PMD_SIZE); | |
421 | } | |
422 | ||
423 | #else | |
424 | static unsigned long __init section_map_size(void) | |
425 | { | |
426 | return PAGE_ALIGN(sizeof(struct page) * PAGES_PER_SECTION); | |
427 | } | |
428 | ||
429 | struct page __init *__populate_section_memmap(unsigned long pfn, | |
430 | unsigned long nr_pages, int nid, struct vmem_altmap *altmap) | |
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 | ||
451 | static inline void __meminit sparse_buffer_free(unsigned long size) | |
452 | { | |
453 | WARN_ON(!sparsemap_buf || size == 0); | |
454 | memblock_free_early(__pa(sparsemap_buf), size); | |
455 | } | |
456 | ||
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 | ||
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 | ||
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 | ||
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 | */ | |
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); | |
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 | */ | |
558 | void __init sparse_init(void) | |
559 | { | |
560 | unsigned long pnum_end, pnum_begin, map_count = 1; | |
561 | int nid_begin; | |
562 | ||
563 | memblocks_present(); | |
564 | ||
565 | pnum_begin = first_present_section_nr(); | |
566 | nid_begin = sparse_early_nid(__nr_to_section(pnum_begin)); | |
567 | ||
568 | /* Setup pageblock_order for HUGETLB_PAGE_SIZE_VARIABLE */ | |
569 | set_pageblock_order(); | |
570 | ||
571 | for_each_present_section_nr(pnum_begin + 1, pnum_end) { | |
572 | int nid = sparse_early_nid(__nr_to_section(pnum_end)); | |
573 | ||
574 | if (nid == nid_begin) { | |
575 | map_count++; | |
576 | continue; | |
577 | } | |
578 | /* Init node with sections in range [pnum_begin, pnum_end) */ | |
579 | sparse_init_nid(nid_begin, pnum_begin, pnum_end, map_count); | |
580 | nid_begin = nid; | |
581 | pnum_begin = pnum_end; | |
582 | map_count = 1; | |
583 | } | |
584 | /* cover the last node */ | |
585 | sparse_init_nid(nid_begin, pnum_begin, pnum_end, map_count); | |
586 | vmemmap_populate_print_last(); | |
587 | } | |
588 | ||
589 | #ifdef CONFIG_MEMORY_HOTPLUG | |
590 | ||
591 | /* Mark all memory sections within the pfn range as online */ | |
592 | void online_mem_sections(unsigned long start_pfn, unsigned long end_pfn) | |
593 | { | |
594 | unsigned long pfn; | |
595 | ||
596 | for (pfn = start_pfn; pfn < end_pfn; pfn += PAGES_PER_SECTION) { | |
597 | unsigned long section_nr = pfn_to_section_nr(pfn); | |
598 | struct mem_section *ms; | |
599 | ||
600 | /* onlining code should never touch invalid ranges */ | |
601 | if (WARN_ON(!valid_section_nr(section_nr))) | |
602 | continue; | |
603 | ||
604 | ms = __nr_to_section(section_nr); | |
605 | ms->section_mem_map |= SECTION_IS_ONLINE; | |
606 | } | |
607 | } | |
608 | ||
609 | /* Mark all memory sections within the pfn range as offline */ | |
610 | void offline_mem_sections(unsigned long start_pfn, unsigned long end_pfn) | |
611 | { | |
612 | unsigned long pfn; | |
613 | ||
614 | for (pfn = start_pfn; pfn < end_pfn; pfn += PAGES_PER_SECTION) { | |
615 | unsigned long section_nr = pfn_to_section_nr(pfn); | |
616 | struct mem_section *ms; | |
617 | ||
618 | /* | |
619 | * TODO this needs some double checking. Offlining code makes | |
620 | * sure to check pfn_valid but those checks might be just bogus | |
621 | */ | |
622 | if (WARN_ON(!valid_section_nr(section_nr))) | |
623 | continue; | |
624 | ||
625 | ms = __nr_to_section(section_nr); | |
626 | ms->section_mem_map &= ~SECTION_IS_ONLINE; | |
627 | } | |
628 | } | |
629 | ||
630 | #ifdef CONFIG_SPARSEMEM_VMEMMAP | |
631 | static struct page * __meminit populate_section_memmap(unsigned long pfn, | |
632 | unsigned long nr_pages, int nid, struct vmem_altmap *altmap) | |
633 | { | |
634 | return __populate_section_memmap(pfn, nr_pages, nid, altmap); | |
635 | } | |
636 | ||
637 | static void depopulate_section_memmap(unsigned long pfn, unsigned long nr_pages, | |
638 | struct vmem_altmap *altmap) | |
639 | { | |
640 | unsigned long start = (unsigned long) pfn_to_page(pfn); | |
641 | unsigned long end = start + nr_pages * sizeof(struct page); | |
642 | ||
643 | vmemmap_free(start, end, altmap); | |
644 | } | |
645 | static void free_map_bootmem(struct page *memmap) | |
646 | { | |
647 | unsigned long start = (unsigned long)memmap; | |
648 | unsigned long end = (unsigned long)(memmap + PAGES_PER_SECTION); | |
649 | ||
650 | vmemmap_free(start, end, NULL); | |
651 | } | |
652 | ||
653 | static int clear_subsection_map(unsigned long pfn, unsigned long nr_pages) | |
654 | { | |
655 | DECLARE_BITMAP(map, SUBSECTIONS_PER_SECTION) = { 0 }; | |
656 | DECLARE_BITMAP(tmp, SUBSECTIONS_PER_SECTION) = { 0 }; | |
657 | struct mem_section *ms = __pfn_to_section(pfn); | |
658 | unsigned long *subsection_map = ms->usage | |
659 | ? &ms->usage->subsection_map[0] : NULL; | |
660 | ||
661 | subsection_mask_set(map, pfn, nr_pages); | |
662 | if (subsection_map) | |
663 | bitmap_and(tmp, map, subsection_map, SUBSECTIONS_PER_SECTION); | |
664 | ||
665 | if (WARN(!subsection_map || !bitmap_equal(tmp, map, SUBSECTIONS_PER_SECTION), | |
666 | "section already deactivated (%#lx + %ld)\n", | |
667 | pfn, nr_pages)) | |
668 | return -EINVAL; | |
669 | ||
670 | bitmap_xor(subsection_map, map, subsection_map, SUBSECTIONS_PER_SECTION); | |
671 | return 0; | |
672 | } | |
673 | ||
674 | static bool is_subsection_map_empty(struct mem_section *ms) | |
675 | { | |
676 | return bitmap_empty(&ms->usage->subsection_map[0], | |
677 | SUBSECTIONS_PER_SECTION); | |
678 | } | |
679 | ||
680 | static int fill_subsection_map(unsigned long pfn, unsigned long nr_pages) | |
681 | { | |
682 | struct mem_section *ms = __pfn_to_section(pfn); | |
683 | DECLARE_BITMAP(map, SUBSECTIONS_PER_SECTION) = { 0 }; | |
684 | unsigned long *subsection_map; | |
685 | int rc = 0; | |
686 | ||
687 | subsection_mask_set(map, pfn, nr_pages); | |
688 | ||
689 | subsection_map = &ms->usage->subsection_map[0]; | |
690 | ||
691 | if (bitmap_empty(map, SUBSECTIONS_PER_SECTION)) | |
692 | rc = -EINVAL; | |
693 | else if (bitmap_intersects(map, subsection_map, SUBSECTIONS_PER_SECTION)) | |
694 | rc = -EEXIST; | |
695 | else | |
696 | bitmap_or(subsection_map, map, subsection_map, | |
697 | SUBSECTIONS_PER_SECTION); | |
698 | ||
699 | return rc; | |
700 | } | |
701 | #else | |
702 | struct page * __meminit populate_section_memmap(unsigned long pfn, | |
703 | unsigned long nr_pages, int nid, struct vmem_altmap *altmap) | |
704 | { | |
705 | return kvmalloc_node(array_size(sizeof(struct page), | |
706 | PAGES_PER_SECTION), GFP_KERNEL, nid); | |
707 | } | |
708 | ||
709 | static void depopulate_section_memmap(unsigned long pfn, unsigned long nr_pages, | |
710 | struct vmem_altmap *altmap) | |
711 | { | |
712 | kvfree(pfn_to_page(pfn)); | |
713 | } | |
714 | ||
715 | static void free_map_bootmem(struct page *memmap) | |
716 | { | |
717 | unsigned long maps_section_nr, removing_section_nr, i; | |
718 | unsigned long magic, nr_pages; | |
719 | struct page *page = virt_to_page(memmap); | |
720 | ||
721 | nr_pages = PAGE_ALIGN(PAGES_PER_SECTION * sizeof(struct page)) | |
722 | >> PAGE_SHIFT; | |
723 | ||
724 | for (i = 0; i < nr_pages; i++, page++) { | |
725 | magic = (unsigned long) page->freelist; | |
726 | ||
727 | BUG_ON(magic == NODE_INFO); | |
728 | ||
729 | maps_section_nr = pfn_to_section_nr(page_to_pfn(page)); | |
730 | removing_section_nr = page_private(page); | |
731 | ||
732 | /* | |
733 | * When this function is called, the removing section is | |
734 | * logical offlined state. This means all pages are isolated | |
735 | * from page allocator. If removing section's memmap is placed | |
736 | * on the same section, it must not be freed. | |
737 | * If it is freed, page allocator may allocate it which will | |
738 | * be removed physically soon. | |
739 | */ | |
740 | if (maps_section_nr != removing_section_nr) | |
741 | put_page_bootmem(page); | |
742 | } | |
743 | } | |
744 | ||
745 | static int clear_subsection_map(unsigned long pfn, unsigned long nr_pages) | |
746 | { | |
747 | return 0; | |
748 | } | |
749 | ||
750 | static bool is_subsection_map_empty(struct mem_section *ms) | |
751 | { | |
752 | return true; | |
753 | } | |
754 | ||
755 | static int fill_subsection_map(unsigned long pfn, unsigned long nr_pages) | |
756 | { | |
757 | return 0; | |
758 | } | |
759 | #endif /* CONFIG_SPARSEMEM_VMEMMAP */ | |
760 | ||
761 | /* | |
762 | * To deactivate a memory region, there are 3 cases to handle across | |
763 | * two configurations (SPARSEMEM_VMEMMAP={y,n}): | |
764 | * | |
765 | * 1. deactivation of a partial hot-added section (only possible in | |
766 | * the SPARSEMEM_VMEMMAP=y case). | |
767 | * a) section was present at memory init. | |
768 | * b) section was hot-added post memory init. | |
769 | * 2. deactivation of a complete hot-added section. | |
770 | * 3. deactivation of a complete section from memory init. | |
771 | * | |
772 | * For 1, when subsection_map does not empty we will not be freeing the | |
773 | * usage map, but still need to free the vmemmap range. | |
774 | * | |
775 | * For 2 and 3, the SPARSEMEM_VMEMMAP={y,n} cases are unified | |
776 | */ | |
777 | static void section_deactivate(unsigned long pfn, unsigned long nr_pages, | |
778 | struct vmem_altmap *altmap) | |
779 | { | |
780 | struct mem_section *ms = __pfn_to_section(pfn); | |
781 | bool section_is_early = early_section(ms); | |
782 | struct page *memmap = NULL; | |
783 | bool empty; | |
784 | ||
785 | if (clear_subsection_map(pfn, nr_pages)) | |
786 | return; | |
787 | ||
788 | empty = is_subsection_map_empty(ms); | |
789 | if (empty) { | |
790 | unsigned long section_nr = pfn_to_section_nr(pfn); | |
791 | ||
792 | /* | |
793 | * When removing an early section, the usage map is kept (as the | |
794 | * usage maps of other sections fall into the same page). It | |
795 | * will be re-used when re-adding the section - which is then no | |
796 | * longer an early section. If the usage map is PageReserved, it | |
797 | * was allocated during boot. | |
798 | */ | |
799 | if (!PageReserved(virt_to_page(ms->usage))) { | |
800 | kfree(ms->usage); | |
801 | ms->usage = NULL; | |
802 | } | |
803 | memmap = sparse_decode_mem_map(ms->section_mem_map, section_nr); | |
804 | /* | |
805 | * Mark the section invalid so that valid_section() | |
806 | * return false. This prevents code from dereferencing | |
807 | * ms->usage array. | |
808 | */ | |
809 | ms->section_mem_map &= ~SECTION_HAS_MEM_MAP; | |
810 | } | |
811 | ||
812 | /* | |
813 | * The memmap of early sections is always fully populated. See | |
814 | * section_activate() and pfn_valid() . | |
815 | */ | |
816 | if (!section_is_early) | |
817 | depopulate_section_memmap(pfn, nr_pages, altmap); | |
818 | else if (memmap) | |
819 | free_map_bootmem(memmap); | |
820 | ||
821 | if (empty) | |
822 | ms->section_mem_map = (unsigned long)NULL; | |
823 | } | |
824 | ||
825 | static struct page * __meminit section_activate(int nid, unsigned long pfn, | |
826 | unsigned long nr_pages, struct vmem_altmap *altmap) | |
827 | { | |
828 | struct mem_section *ms = __pfn_to_section(pfn); | |
829 | struct mem_section_usage *usage = NULL; | |
830 | struct page *memmap; | |
831 | int rc = 0; | |
832 | ||
833 | if (!ms->usage) { | |
834 | usage = kzalloc(mem_section_usage_size(), GFP_KERNEL); | |
835 | if (!usage) | |
836 | return ERR_PTR(-ENOMEM); | |
837 | ms->usage = usage; | |
838 | } | |
839 | ||
840 | rc = fill_subsection_map(pfn, nr_pages); | |
841 | if (rc) { | |
842 | if (usage) | |
843 | ms->usage = NULL; | |
844 | kfree(usage); | |
845 | return ERR_PTR(rc); | |
846 | } | |
847 | ||
848 | /* | |
849 | * The early init code does not consider partially populated | |
850 | * initial sections, it simply assumes that memory will never be | |
851 | * referenced. If we hot-add memory into such a section then we | |
852 | * do not need to populate the memmap and can simply reuse what | |
853 | * is already there. | |
854 | */ | |
855 | if (nr_pages < PAGES_PER_SECTION && early_section(ms)) | |
856 | return pfn_to_page(pfn); | |
857 | ||
858 | memmap = populate_section_memmap(pfn, nr_pages, nid, altmap); | |
859 | if (!memmap) { | |
860 | section_deactivate(pfn, nr_pages, altmap); | |
861 | return ERR_PTR(-ENOMEM); | |
862 | } | |
863 | ||
864 | return memmap; | |
865 | } | |
866 | ||
867 | /** | |
868 | * sparse_add_section - add a memory section, or populate an existing one | |
869 | * @nid: The node to add section on | |
870 | * @start_pfn: start pfn of the memory range | |
871 | * @nr_pages: number of pfns to add in the section | |
872 | * @altmap: device page map | |
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 | */ | |
885 | int __meminit sparse_add_section(int nid, unsigned long start_pfn, | |
886 | unsigned long nr_pages, struct vmem_altmap *altmap) | |
887 | { | |
888 | unsigned long section_nr = pfn_to_section_nr(start_pfn); | |
889 | struct mem_section *ms; | |
890 | struct page *memmap; | |
891 | int ret; | |
892 | ||
893 | ret = sparse_index_init(section_nr, nid); | |
894 | if (ret < 0) | |
895 | return ret; | |
896 | ||
897 | memmap = section_activate(nid, start_pfn, nr_pages, altmap); | |
898 | if (IS_ERR(memmap)) | |
899 | return PTR_ERR(memmap); | |
900 | ||
901 | /* | |
902 | * Poison uninitialized struct pages in order to catch invalid flags | |
903 | * combinations. | |
904 | */ | |
905 | page_init_poison(memmap, sizeof(struct page) * nr_pages); | |
906 | ||
907 | ms = __nr_to_section(section_nr); | |
908 | set_section_nid(section_nr, nid); | |
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 | ||
919 | #ifdef CONFIG_MEMORY_FAILURE | |
920 | static void clear_hwpoisoned_pages(struct page *memmap, int nr_pages) | |
921 | { | |
922 | int i; | |
923 | ||
924 | /* | |
925 | * A further optimization is to have per section refcounted | |
926 | * num_poisoned_pages. But that would need more space per memmap, so | |
927 | * for now just do a quick global check to speed up this routine in the | |
928 | * absence of bad pages. | |
929 | */ | |
930 | if (atomic_long_read(&num_poisoned_pages) == 0) | |
931 | return; | |
932 | ||
933 | for (i = 0; i < nr_pages; i++) { | |
934 | if (PageHWPoison(&memmap[i])) { | |
935 | num_poisoned_pages_dec(); | |
936 | ClearPageHWPoison(&memmap[i]); | |
937 | } | |
938 | } | |
939 | } | |
940 | #else | |
941 | static inline void clear_hwpoisoned_pages(struct page *memmap, int nr_pages) | |
942 | { | |
943 | } | |
944 | #endif | |
945 | ||
946 | void sparse_remove_section(struct mem_section *ms, unsigned long pfn, | |
947 | unsigned long nr_pages, unsigned long map_offset, | |
948 | struct vmem_altmap *altmap) | |
949 | { | |
950 | clear_hwpoisoned_pages(pfn_to_page(pfn) + map_offset, | |
951 | nr_pages - map_offset); | |
952 | section_deactivate(pfn, nr_pages, altmap); | |
953 | } | |
954 | #endif /* CONFIG_MEMORY_HOTPLUG */ |