]>
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 | ||
17 | #include "internal.h" | |
18 | #include <asm/dma.h> | |
19 | ||
20 | /* | |
21 | * Permanent SPARSEMEM data: | |
22 | * | |
23 | * 1) mem_section - memory sections, mem_map's for valid memory | |
24 | */ | |
25 | #ifdef CONFIG_SPARSEMEM_EXTREME | |
26 | struct mem_section **mem_section; | |
27 | #else | |
28 | struct mem_section mem_section[NR_SECTION_ROOTS][SECTIONS_PER_ROOT] | |
29 | ____cacheline_internodealigned_in_smp; | |
30 | #endif | |
31 | EXPORT_SYMBOL(mem_section); | |
32 | ||
33 | #ifdef NODE_NOT_IN_PAGE_FLAGS | |
34 | /* | |
35 | * If we did not store the node number in the page then we have to | |
36 | * do a lookup in the section_to_node_table in order to find which | |
37 | * node the page belongs to. | |
38 | */ | |
39 | #if MAX_NUMNODES <= 256 | |
40 | static u8 section_to_node_table[NR_MEM_SECTIONS] __cacheline_aligned; | |
41 | #else | |
42 | static u16 section_to_node_table[NR_MEM_SECTIONS] __cacheline_aligned; | |
43 | #endif | |
44 | ||
45 | int page_to_nid(const struct page *page) | |
46 | { | |
47 | return section_to_node_table[page_to_section(page)]; | |
48 | } | |
49 | EXPORT_SYMBOL(page_to_nid); | |
50 | ||
51 | static void set_section_nid(unsigned long section_nr, int nid) | |
52 | { | |
53 | section_to_node_table[section_nr] = nid; | |
54 | } | |
55 | #else /* !NODE_NOT_IN_PAGE_FLAGS */ | |
56 | static inline void set_section_nid(unsigned long section_nr, int nid) | |
57 | { | |
58 | } | |
59 | #endif | |
60 | ||
61 | #ifdef CONFIG_SPARSEMEM_EXTREME | |
62 | static noinline struct mem_section __ref *sparse_index_alloc(int nid) | |
63 | { | |
64 | struct mem_section *section = NULL; | |
65 | unsigned long array_size = SECTIONS_PER_ROOT * | |
66 | sizeof(struct mem_section); | |
67 | ||
68 | if (slab_is_available()) { | |
69 | section = kzalloc_node(array_size, GFP_KERNEL, nid); | |
70 | } else { | |
71 | section = memblock_alloc_node(array_size, SMP_CACHE_BYTES, | |
72 | nid); | |
73 | if (!section) | |
74 | panic("%s: Failed to allocate %lu bytes nid=%d\n", | |
75 | __func__, array_size, nid); | |
76 | } | |
77 | ||
78 | return section; | |
79 | } | |
80 | ||
81 | static int __meminit sparse_index_init(unsigned long section_nr, int nid) | |
82 | { | |
83 | unsigned long root = SECTION_NR_TO_ROOT(section_nr); | |
84 | struct mem_section *section; | |
85 | ||
86 | /* | |
87 | * An existing section is possible in the sub-section hotplug | |
88 | * case. First hot-add instantiates, follow-on hot-add reuses | |
89 | * the existing section. | |
90 | * | |
91 | * The mem_hotplug_lock resolves the apparent race below. | |
92 | */ | |
93 | if (mem_section[root]) | |
94 | return 0; | |
95 | ||
96 | section = sparse_index_alloc(nid); | |
97 | if (!section) | |
98 | return -ENOMEM; | |
99 | ||
100 | mem_section[root] = section; | |
101 | ||
102 | return 0; | |
103 | } | |
104 | #else /* !SPARSEMEM_EXTREME */ | |
105 | static inline int sparse_index_init(unsigned long section_nr, int nid) | |
106 | { | |
107 | return 0; | |
108 | } | |
109 | #endif | |
110 | ||
111 | #ifdef CONFIG_SPARSEMEM_EXTREME | |
112 | unsigned long __section_nr(struct mem_section *ms) | |
113 | { | |
114 | unsigned long root_nr; | |
115 | struct mem_section *root = NULL; | |
116 | ||
117 | for (root_nr = 0; root_nr < NR_SECTION_ROOTS; root_nr++) { | |
118 | root = __nr_to_section(root_nr * SECTIONS_PER_ROOT); | |
119 | if (!root) | |
120 | continue; | |
121 | ||
122 | if ((ms >= root) && (ms < (root + SECTIONS_PER_ROOT))) | |
123 | break; | |
124 | } | |
125 | ||
126 | VM_BUG_ON(!root); | |
127 | ||
128 | return (root_nr * SECTIONS_PER_ROOT) + (ms - root); | |
129 | } | |
130 | #else | |
131 | unsigned long __section_nr(struct mem_section *ms) | |
132 | { | |
133 | return (unsigned long)(ms - mem_section[0]); | |
134 | } | |
135 | #endif | |
136 | ||
137 | /* | |
138 | * During early boot, before section_mem_map is used for an actual | |
139 | * mem_map, we use section_mem_map to store the section's NUMA | |
140 | * node. This keeps us from having to use another data structure. The | |
141 | * node information is cleared just before we store the real mem_map. | |
142 | */ | |
143 | static inline unsigned long sparse_encode_early_nid(int nid) | |
144 | { | |
145 | return (nid << SECTION_NID_SHIFT); | |
146 | } | |
147 | ||
148 | static inline int sparse_early_nid(struct mem_section *section) | |
149 | { | |
150 | return (section->section_mem_map >> SECTION_NID_SHIFT); | |
151 | } | |
152 | ||
153 | /* Validate the physical addressing limitations of the model */ | |
154 | void __meminit mminit_validate_memmodel_limits(unsigned long *start_pfn, | |
155 | unsigned long *end_pfn) | |
156 | { | |
157 | unsigned long max_sparsemem_pfn = 1UL << (MAX_PHYSMEM_BITS-PAGE_SHIFT); | |
158 | ||
159 | /* | |
160 | * Sanity checks - do not allow an architecture to pass | |
161 | * in larger pfns than the maximum scope of sparsemem: | |
162 | */ | |
163 | if (*start_pfn > max_sparsemem_pfn) { | |
164 | mminit_dprintk(MMINIT_WARNING, "pfnvalidation", | |
165 | "Start of range %lu -> %lu exceeds SPARSEMEM max %lu\n", | |
166 | *start_pfn, *end_pfn, max_sparsemem_pfn); | |
167 | WARN_ON_ONCE(1); | |
168 | *start_pfn = max_sparsemem_pfn; | |
169 | *end_pfn = max_sparsemem_pfn; | |
170 | } else if (*end_pfn > max_sparsemem_pfn) { | |
171 | mminit_dprintk(MMINIT_WARNING, "pfnvalidation", | |
172 | "End of range %lu -> %lu exceeds SPARSEMEM max %lu\n", | |
173 | *start_pfn, *end_pfn, max_sparsemem_pfn); | |
174 | WARN_ON_ONCE(1); | |
175 | *end_pfn = max_sparsemem_pfn; | |
176 | } | |
177 | } | |
178 | ||
179 | /* | |
180 | * There are a number of times that we loop over NR_MEM_SECTIONS, | |
181 | * looking for section_present() on each. But, when we have very | |
182 | * large physical address spaces, NR_MEM_SECTIONS can also be | |
183 | * very large which makes the loops quite long. | |
184 | * | |
185 | * Keeping track of this gives us an easy way to break out of | |
186 | * those loops early. | |
187 | */ | |
188 | unsigned long __highest_present_section_nr; | |
189 | static void section_mark_present(struct mem_section *ms) | |
190 | { | |
191 | unsigned long section_nr = __section_nr(ms); | |
192 | ||
193 | if (section_nr > __highest_present_section_nr) | |
194 | __highest_present_section_nr = section_nr; | |
195 | ||
196 | ms->section_mem_map |= SECTION_MARKED_PRESENT; | |
197 | } | |
198 | ||
199 | #define for_each_present_section_nr(start, section_nr) \ | |
200 | for (section_nr = next_present_section_nr(start-1); \ | |
201 | ((section_nr != -1) && \ | |
202 | (section_nr <= __highest_present_section_nr)); \ | |
203 | section_nr = next_present_section_nr(section_nr)) | |
204 | ||
205 | static inline unsigned long first_present_section_nr(void) | |
206 | { | |
207 | return next_present_section_nr(-1); | |
208 | } | |
209 | ||
210 | #ifdef CONFIG_SPARSEMEM_VMEMMAP | |
211 | static void subsection_mask_set(unsigned long *map, unsigned long pfn, | |
212 | unsigned long nr_pages) | |
213 | { | |
214 | int idx = subsection_map_index(pfn); | |
215 | int end = subsection_map_index(pfn + nr_pages - 1); | |
216 | ||
217 | bitmap_set(map, idx, end - idx + 1); | |
218 | } | |
219 | ||
220 | void __init subsection_map_init(unsigned long pfn, unsigned long nr_pages) | |
221 | { | |
222 | int end_sec = pfn_to_section_nr(pfn + nr_pages - 1); | |
223 | unsigned long nr, start_sec = pfn_to_section_nr(pfn); | |
224 | ||
225 | if (!nr_pages) | |
226 | return; | |
227 | ||
228 | for (nr = start_sec; nr <= end_sec; nr++) { | |
229 | struct mem_section *ms; | |
230 | unsigned long pfns; | |
231 | ||
232 | pfns = min(nr_pages, PAGES_PER_SECTION | |
233 | - (pfn & ~PAGE_SECTION_MASK)); | |
234 | ms = __nr_to_section(nr); | |
235 | subsection_mask_set(ms->usage->subsection_map, pfn, pfns); | |
236 | ||
237 | pr_debug("%s: sec: %lu pfns: %lu set(%d, %d)\n", __func__, nr, | |
238 | pfns, subsection_map_index(pfn), | |
239 | subsection_map_index(pfn + pfns - 1)); | |
240 | ||
241 | pfn += pfns; | |
242 | nr_pages -= pfns; | |
243 | } | |
244 | } | |
245 | #else | |
246 | void __init subsection_map_init(unsigned long pfn, unsigned long nr_pages) | |
247 | { | |
248 | } | |
249 | #endif | |
250 | ||
251 | /* Record a memory area against a node. */ | |
252 | static void __init memory_present(int nid, unsigned long start, unsigned long end) | |
253 | { | |
254 | unsigned long pfn; | |
255 | ||
256 | #ifdef CONFIG_SPARSEMEM_EXTREME | |
257 | if (unlikely(!mem_section)) { | |
258 | unsigned long size, align; | |
259 | ||
260 | size = sizeof(struct mem_section*) * NR_SECTION_ROOTS; | |
261 | align = 1 << (INTERNODE_CACHE_SHIFT); | |
262 | mem_section = memblock_alloc(size, align); | |
263 | if (!mem_section) | |
264 | panic("%s: Failed to allocate %lu bytes align=0x%lx\n", | |
265 | __func__, size, align); | |
266 | } | |
267 | #endif | |
268 | ||
269 | start &= PAGE_SECTION_MASK; | |
270 | mminit_validate_memmodel_limits(&start, &end); | |
271 | for (pfn = start; pfn < end; pfn += PAGES_PER_SECTION) { | |
272 | unsigned long section = pfn_to_section_nr(pfn); | |
273 | struct mem_section *ms; | |
274 | ||
275 | sparse_index_init(section, nid); | |
276 | set_section_nid(section, nid); | |
277 | ||
278 | ms = __nr_to_section(section); | |
279 | if (!ms->section_mem_map) { | |
280 | ms->section_mem_map = sparse_encode_early_nid(nid) | | |
281 | SECTION_IS_ONLINE; | |
282 | section_mark_present(ms); | |
283 | } | |
284 | } | |
285 | } | |
286 | ||
287 | /* | |
288 | * Mark all memblocks as present using memory_present(). | |
289 | * This is a convenience function that is useful to mark all of the systems | |
290 | * memory as present during initialization. | |
291 | */ | |
292 | static void __init memblocks_present(void) | |
293 | { | |
294 | unsigned long start, end; | |
295 | int i, nid; | |
296 | ||
297 | for_each_mem_pfn_range(i, MAX_NUMNODES, &start, &end, &nid) | |
298 | memory_present(nid, start, end); | |
299 | } | |
300 | ||
301 | /* | |
302 | * Subtle, we encode the real pfn into the mem_map such that | |
303 | * the identity pfn - section_mem_map will return the actual | |
304 | * physical page frame number. | |
305 | */ | |
306 | static unsigned long sparse_encode_mem_map(struct page *mem_map, unsigned long pnum) | |
307 | { | |
308 | unsigned long coded_mem_map = | |
309 | (unsigned long)(mem_map - (section_nr_to_pfn(pnum))); | |
310 | BUILD_BUG_ON(SECTION_MAP_LAST_BIT > (1UL<<PFN_SECTION_SHIFT)); | |
311 | BUG_ON(coded_mem_map & ~SECTION_MAP_MASK); | |
312 | return coded_mem_map; | |
313 | } | |
314 | ||
315 | #ifdef CONFIG_MEMORY_HOTPLUG | |
316 | /* | |
317 | * Decode mem_map from the coded memmap | |
318 | */ | |
319 | struct page *sparse_decode_mem_map(unsigned long coded_mem_map, unsigned long pnum) | |
320 | { | |
321 | /* mask off the extra low bits of information */ | |
322 | coded_mem_map &= SECTION_MAP_MASK; | |
323 | return ((struct page *)coded_mem_map) + section_nr_to_pfn(pnum); | |
324 | } | |
325 | #endif /* CONFIG_MEMORY_HOTPLUG */ | |
326 | ||
327 | static void __meminit sparse_init_one_section(struct mem_section *ms, | |
328 | unsigned long pnum, struct page *mem_map, | |
329 | struct mem_section_usage *usage, unsigned long flags) | |
330 | { | |
331 | ms->section_mem_map &= ~SECTION_MAP_MASK; | |
332 | ms->section_mem_map |= sparse_encode_mem_map(mem_map, pnum) | |
333 | | SECTION_HAS_MEM_MAP | flags; | |
334 | ms->usage = usage; | |
335 | } | |
336 | ||
337 | static unsigned long usemap_size(void) | |
338 | { | |
339 | return BITS_TO_LONGS(SECTION_BLOCKFLAGS_BITS) * sizeof(unsigned long); | |
340 | } | |
341 | ||
342 | size_t mem_section_usage_size(void) | |
343 | { | |
344 | return sizeof(struct mem_section_usage) + usemap_size(); | |
345 | } | |
346 | ||
347 | #ifdef CONFIG_MEMORY_HOTREMOVE | |
348 | static struct mem_section_usage * __init | |
349 | sparse_early_usemaps_alloc_pgdat_section(struct pglist_data *pgdat, | |
350 | unsigned long size) | |
351 | { | |
352 | struct mem_section_usage *usage; | |
353 | unsigned long goal, limit; | |
354 | int nid; | |
355 | /* | |
356 | * A page may contain usemaps for other sections preventing the | |
357 | * page being freed and making a section unremovable while | |
358 | * other sections referencing the usemap remain active. Similarly, | |
359 | * a pgdat can prevent a section being removed. If section A | |
360 | * contains a pgdat and section B contains the usemap, both | |
361 | * sections become inter-dependent. This allocates usemaps | |
362 | * from the same section as the pgdat where possible to avoid | |
363 | * this problem. | |
364 | */ | |
365 | goal = __pa(pgdat) & (PAGE_SECTION_MASK << PAGE_SHIFT); | |
366 | limit = goal + (1UL << PA_SECTION_SHIFT); | |
367 | nid = early_pfn_to_nid(goal >> PAGE_SHIFT); | |
368 | again: | |
369 | usage = memblock_alloc_try_nid(size, SMP_CACHE_BYTES, goal, limit, nid); | |
370 | if (!usage && limit) { | |
371 | limit = 0; | |
372 | goto again; | |
373 | } | |
374 | return usage; | |
375 | } | |
376 | ||
377 | static void __init check_usemap_section_nr(int nid, | |
378 | struct mem_section_usage *usage) | |
379 | { | |
380 | unsigned long usemap_snr, pgdat_snr; | |
381 | static unsigned long old_usemap_snr; | |
382 | static unsigned long old_pgdat_snr; | |
383 | struct pglist_data *pgdat = NODE_DATA(nid); | |
384 | int usemap_nid; | |
385 | ||
386 | /* First call */ | |
387 | if (!old_usemap_snr) { | |
388 | old_usemap_snr = NR_MEM_SECTIONS; | |
389 | old_pgdat_snr = NR_MEM_SECTIONS; | |
390 | } | |
391 | ||
392 | usemap_snr = pfn_to_section_nr(__pa(usage) >> PAGE_SHIFT); | |
393 | pgdat_snr = pfn_to_section_nr(__pa(pgdat) >> PAGE_SHIFT); | |
394 | if (usemap_snr == pgdat_snr) | |
395 | return; | |
396 | ||
397 | if (old_usemap_snr == usemap_snr && old_pgdat_snr == pgdat_snr) | |
398 | /* skip redundant message */ | |
399 | return; | |
400 | ||
401 | old_usemap_snr = usemap_snr; | |
402 | old_pgdat_snr = pgdat_snr; | |
403 | ||
404 | usemap_nid = sparse_early_nid(__nr_to_section(usemap_snr)); | |
405 | if (usemap_nid != nid) { | |
406 | pr_info("node %d must be removed before remove section %ld\n", | |
407 | nid, usemap_snr); | |
408 | return; | |
409 | } | |
410 | /* | |
411 | * There is a circular dependency. | |
412 | * Some platforms allow un-removable section because they will just | |
413 | * gather other removable sections for dynamic partitioning. | |
414 | * Just notify un-removable section's number here. | |
415 | */ | |
416 | pr_info("Section %ld and %ld (node %d) have a circular dependency on usemap and pgdat allocations\n", | |
417 | usemap_snr, pgdat_snr, nid); | |
418 | } | |
419 | #else | |
420 | static struct mem_section_usage * __init | |
421 | sparse_early_usemaps_alloc_pgdat_section(struct pglist_data *pgdat, | |
422 | unsigned long size) | |
423 | { | |
424 | return memblock_alloc_node(size, SMP_CACHE_BYTES, pgdat->node_id); | |
425 | } | |
426 | ||
427 | static void __init check_usemap_section_nr(int nid, | |
428 | struct mem_section_usage *usage) | |
429 | { | |
430 | } | |
431 | #endif /* CONFIG_MEMORY_HOTREMOVE */ | |
432 | ||
433 | #ifdef CONFIG_SPARSEMEM_VMEMMAP | |
434 | static unsigned long __init section_map_size(void) | |
435 | { | |
436 | return ALIGN(sizeof(struct page) * PAGES_PER_SECTION, PMD_SIZE); | |
437 | } | |
438 | ||
439 | #else | |
440 | static unsigned long __init section_map_size(void) | |
441 | { | |
442 | return PAGE_ALIGN(sizeof(struct page) * PAGES_PER_SECTION); | |
443 | } | |
444 | ||
445 | struct page __init *__populate_section_memmap(unsigned long pfn, | |
446 | unsigned long nr_pages, int nid, struct vmem_altmap *altmap) | |
447 | { | |
448 | unsigned long size = section_map_size(); | |
449 | struct page *map = sparse_buffer_alloc(size); | |
450 | phys_addr_t addr = __pa(MAX_DMA_ADDRESS); | |
451 | ||
452 | if (map) | |
453 | return map; | |
454 | ||
455 | map = memblock_alloc_try_nid_raw(size, size, addr, | |
456 | MEMBLOCK_ALLOC_ACCESSIBLE, nid); | |
457 | if (!map) | |
458 | panic("%s: Failed to allocate %lu bytes align=0x%lx nid=%d from=%pa\n", | |
459 | __func__, size, PAGE_SIZE, nid, &addr); | |
460 | ||
461 | return map; | |
462 | } | |
463 | #endif /* !CONFIG_SPARSEMEM_VMEMMAP */ | |
464 | ||
465 | static void *sparsemap_buf __meminitdata; | |
466 | static void *sparsemap_buf_end __meminitdata; | |
467 | ||
468 | static inline void __meminit sparse_buffer_free(unsigned long size) | |
469 | { | |
470 | WARN_ON(!sparsemap_buf || size == 0); | |
471 | memblock_free_early(__pa(sparsemap_buf), size); | |
472 | } | |
473 | ||
474 | static void __init sparse_buffer_init(unsigned long size, int nid) | |
475 | { | |
476 | phys_addr_t addr = __pa(MAX_DMA_ADDRESS); | |
477 | WARN_ON(sparsemap_buf); /* forgot to call sparse_buffer_fini()? */ | |
478 | /* | |
479 | * Pre-allocated buffer is mainly used by __populate_section_memmap | |
480 | * and we want it to be properly aligned to the section size - this is | |
481 | * especially the case for VMEMMAP which maps memmap to PMDs | |
482 | */ | |
483 | sparsemap_buf = memblock_alloc_exact_nid_raw(size, section_map_size(), | |
484 | addr, MEMBLOCK_ALLOC_ACCESSIBLE, nid); | |
485 | sparsemap_buf_end = sparsemap_buf + size; | |
486 | } | |
487 | ||
488 | static void __init sparse_buffer_fini(void) | |
489 | { | |
490 | unsigned long size = sparsemap_buf_end - sparsemap_buf; | |
491 | ||
492 | if (sparsemap_buf && size > 0) | |
493 | sparse_buffer_free(size); | |
494 | sparsemap_buf = NULL; | |
495 | } | |
496 | ||
497 | void * __meminit sparse_buffer_alloc(unsigned long size) | |
498 | { | |
499 | void *ptr = NULL; | |
500 | ||
501 | if (sparsemap_buf) { | |
502 | ptr = (void *) roundup((unsigned long)sparsemap_buf, size); | |
503 | if (ptr + size > sparsemap_buf_end) | |
504 | ptr = NULL; | |
505 | else { | |
506 | /* Free redundant aligned space */ | |
507 | if ((unsigned long)(ptr - sparsemap_buf) > 0) | |
508 | sparse_buffer_free((unsigned long)(ptr - sparsemap_buf)); | |
509 | sparsemap_buf = ptr + size; | |
510 | } | |
511 | } | |
512 | return ptr; | |
513 | } | |
514 | ||
515 | void __weak __meminit vmemmap_populate_print_last(void) | |
516 | { | |
517 | } | |
518 | ||
519 | /* | |
520 | * Initialize sparse on a specific node. The node spans [pnum_begin, pnum_end) | |
521 | * And number of present sections in this node is map_count. | |
522 | */ | |
523 | static void __init sparse_init_nid(int nid, unsigned long pnum_begin, | |
524 | unsigned long pnum_end, | |
525 | unsigned long map_count) | |
526 | { | |
527 | struct mem_section_usage *usage; | |
528 | unsigned long pnum; | |
529 | struct page *map; | |
530 | ||
531 | usage = sparse_early_usemaps_alloc_pgdat_section(NODE_DATA(nid), | |
532 | mem_section_usage_size() * map_count); | |
533 | if (!usage) { | |
534 | pr_err("%s: node[%d] usemap allocation failed", __func__, nid); | |
535 | goto failed; | |
536 | } | |
537 | sparse_buffer_init(map_count * section_map_size(), nid); | |
538 | for_each_present_section_nr(pnum_begin, pnum) { | |
539 | unsigned long pfn = section_nr_to_pfn(pnum); | |
540 | ||
541 | if (pnum >= pnum_end) | |
542 | break; | |
543 | ||
544 | map = __populate_section_memmap(pfn, PAGES_PER_SECTION, | |
545 | nid, NULL); | |
546 | if (!map) { | |
547 | pr_err("%s: node[%d] memory map backing failed. Some memory will not be available.", | |
548 | __func__, nid); | |
549 | pnum_begin = pnum; | |
550 | sparse_buffer_fini(); | |
551 | goto failed; | |
552 | } | |
553 | check_usemap_section_nr(nid, usage); | |
554 | sparse_init_one_section(__nr_to_section(pnum), pnum, map, usage, | |
555 | SECTION_IS_EARLY); | |
556 | usage = (void *) usage + mem_section_usage_size(); | |
557 | } | |
558 | sparse_buffer_fini(); | |
559 | return; | |
560 | failed: | |
561 | /* We failed to allocate, mark all the following pnums as not present */ | |
562 | for_each_present_section_nr(pnum_begin, pnum) { | |
563 | struct mem_section *ms; | |
564 | ||
565 | if (pnum >= pnum_end) | |
566 | break; | |
567 | ms = __nr_to_section(pnum); | |
568 | ms->section_mem_map = 0; | |
569 | } | |
570 | } | |
571 | ||
572 | /* | |
573 | * Allocate the accumulated non-linear sections, allocate a mem_map | |
574 | * for each and record the physical to section mapping. | |
575 | */ | |
576 | void __init sparse_init(void) | |
577 | { | |
578 | unsigned long pnum_end, pnum_begin, map_count = 1; | |
579 | int nid_begin; | |
580 | ||
581 | memblocks_present(); | |
582 | ||
583 | pnum_begin = first_present_section_nr(); | |
584 | nid_begin = sparse_early_nid(__nr_to_section(pnum_begin)); | |
585 | ||
586 | /* Setup pageblock_order for HUGETLB_PAGE_SIZE_VARIABLE */ | |
587 | set_pageblock_order(); | |
588 | ||
589 | for_each_present_section_nr(pnum_begin + 1, pnum_end) { | |
590 | int nid = sparse_early_nid(__nr_to_section(pnum_end)); | |
591 | ||
592 | if (nid == nid_begin) { | |
593 | map_count++; | |
594 | continue; | |
595 | } | |
596 | /* Init node with sections in range [pnum_begin, pnum_end) */ | |
597 | sparse_init_nid(nid_begin, pnum_begin, pnum_end, map_count); | |
598 | nid_begin = nid; | |
599 | pnum_begin = pnum_end; | |
600 | map_count = 1; | |
601 | } | |
602 | /* cover the last node */ | |
603 | sparse_init_nid(nid_begin, pnum_begin, pnum_end, map_count); | |
604 | vmemmap_populate_print_last(); | |
605 | } | |
606 | ||
607 | #ifdef CONFIG_MEMORY_HOTPLUG | |
608 | ||
609 | /* Mark all memory sections within the pfn range as online */ | |
610 | void online_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 | /* onlining code should never touch invalid ranges */ | |
619 | if (WARN_ON(!valid_section_nr(section_nr))) | |
620 | continue; | |
621 | ||
622 | ms = __nr_to_section(section_nr); | |
623 | ms->section_mem_map |= SECTION_IS_ONLINE; | |
624 | } | |
625 | } | |
626 | ||
627 | #ifdef CONFIG_MEMORY_HOTREMOVE | |
628 | /* Mark all memory sections within the pfn range as offline */ | |
629 | void offline_mem_sections(unsigned long start_pfn, unsigned long end_pfn) | |
630 | { | |
631 | unsigned long pfn; | |
632 | ||
633 | for (pfn = start_pfn; pfn < end_pfn; pfn += PAGES_PER_SECTION) { | |
634 | unsigned long section_nr = pfn_to_section_nr(pfn); | |
635 | struct mem_section *ms; | |
636 | ||
637 | /* | |
638 | * TODO this needs some double checking. Offlining code makes | |
639 | * sure to check pfn_valid but those checks might be just bogus | |
640 | */ | |
641 | if (WARN_ON(!valid_section_nr(section_nr))) | |
642 | continue; | |
643 | ||
644 | ms = __nr_to_section(section_nr); | |
645 | ms->section_mem_map &= ~SECTION_IS_ONLINE; | |
646 | } | |
647 | } | |
648 | #endif | |
649 | ||
650 | #ifdef CONFIG_SPARSEMEM_VMEMMAP | |
651 | static struct page * __meminit populate_section_memmap(unsigned long pfn, | |
652 | unsigned long nr_pages, int nid, struct vmem_altmap *altmap) | |
653 | { | |
654 | return __populate_section_memmap(pfn, nr_pages, nid, altmap); | |
655 | } | |
656 | ||
657 | static void depopulate_section_memmap(unsigned long pfn, unsigned long nr_pages, | |
658 | struct vmem_altmap *altmap) | |
659 | { | |
660 | unsigned long start = (unsigned long) pfn_to_page(pfn); | |
661 | unsigned long end = start + nr_pages * sizeof(struct page); | |
662 | ||
663 | vmemmap_free(start, end, altmap); | |
664 | } | |
665 | static void free_map_bootmem(struct page *memmap) | |
666 | { | |
667 | unsigned long start = (unsigned long)memmap; | |
668 | unsigned long end = (unsigned long)(memmap + PAGES_PER_SECTION); | |
669 | ||
670 | vmemmap_free(start, end, NULL); | |
671 | } | |
672 | ||
673 | static int clear_subsection_map(unsigned long pfn, unsigned long nr_pages) | |
674 | { | |
675 | DECLARE_BITMAP(map, SUBSECTIONS_PER_SECTION) = { 0 }; | |
676 | DECLARE_BITMAP(tmp, SUBSECTIONS_PER_SECTION) = { 0 }; | |
677 | struct mem_section *ms = __pfn_to_section(pfn); | |
678 | unsigned long *subsection_map = ms->usage | |
679 | ? &ms->usage->subsection_map[0] : NULL; | |
680 | ||
681 | subsection_mask_set(map, pfn, nr_pages); | |
682 | if (subsection_map) | |
683 | bitmap_and(tmp, map, subsection_map, SUBSECTIONS_PER_SECTION); | |
684 | ||
685 | if (WARN(!subsection_map || !bitmap_equal(tmp, map, SUBSECTIONS_PER_SECTION), | |
686 | "section already deactivated (%#lx + %ld)\n", | |
687 | pfn, nr_pages)) | |
688 | return -EINVAL; | |
689 | ||
690 | bitmap_xor(subsection_map, map, subsection_map, SUBSECTIONS_PER_SECTION); | |
691 | return 0; | |
692 | } | |
693 | ||
694 | static bool is_subsection_map_empty(struct mem_section *ms) | |
695 | { | |
696 | return bitmap_empty(&ms->usage->subsection_map[0], | |
697 | SUBSECTIONS_PER_SECTION); | |
698 | } | |
699 | ||
700 | static int fill_subsection_map(unsigned long pfn, unsigned long nr_pages) | |
701 | { | |
702 | struct mem_section *ms = __pfn_to_section(pfn); | |
703 | DECLARE_BITMAP(map, SUBSECTIONS_PER_SECTION) = { 0 }; | |
704 | unsigned long *subsection_map; | |
705 | int rc = 0; | |
706 | ||
707 | subsection_mask_set(map, pfn, nr_pages); | |
708 | ||
709 | subsection_map = &ms->usage->subsection_map[0]; | |
710 | ||
711 | if (bitmap_empty(map, SUBSECTIONS_PER_SECTION)) | |
712 | rc = -EINVAL; | |
713 | else if (bitmap_intersects(map, subsection_map, SUBSECTIONS_PER_SECTION)) | |
714 | rc = -EEXIST; | |
715 | else | |
716 | bitmap_or(subsection_map, map, subsection_map, | |
717 | SUBSECTIONS_PER_SECTION); | |
718 | ||
719 | return rc; | |
720 | } | |
721 | #else | |
722 | struct page * __meminit populate_section_memmap(unsigned long pfn, | |
723 | unsigned long nr_pages, int nid, struct vmem_altmap *altmap) | |
724 | { | |
725 | return kvmalloc_node(array_size(sizeof(struct page), | |
726 | PAGES_PER_SECTION), GFP_KERNEL, nid); | |
727 | } | |
728 | ||
729 | static void depopulate_section_memmap(unsigned long pfn, unsigned long nr_pages, | |
730 | struct vmem_altmap *altmap) | |
731 | { | |
732 | kvfree(pfn_to_page(pfn)); | |
733 | } | |
734 | ||
735 | static void free_map_bootmem(struct page *memmap) | |
736 | { | |
737 | unsigned long maps_section_nr, removing_section_nr, i; | |
738 | unsigned long magic, nr_pages; | |
739 | struct page *page = virt_to_page(memmap); | |
740 | ||
741 | nr_pages = PAGE_ALIGN(PAGES_PER_SECTION * sizeof(struct page)) | |
742 | >> PAGE_SHIFT; | |
743 | ||
744 | for (i = 0; i < nr_pages; i++, page++) { | |
745 | magic = (unsigned long) page->freelist; | |
746 | ||
747 | BUG_ON(magic == NODE_INFO); | |
748 | ||
749 | maps_section_nr = pfn_to_section_nr(page_to_pfn(page)); | |
750 | removing_section_nr = page_private(page); | |
751 | ||
752 | /* | |
753 | * When this function is called, the removing section is | |
754 | * logical offlined state. This means all pages are isolated | |
755 | * from page allocator. If removing section's memmap is placed | |
756 | * on the same section, it must not be freed. | |
757 | * If it is freed, page allocator may allocate it which will | |
758 | * be removed physically soon. | |
759 | */ | |
760 | if (maps_section_nr != removing_section_nr) | |
761 | put_page_bootmem(page); | |
762 | } | |
763 | } | |
764 | ||
765 | static int clear_subsection_map(unsigned long pfn, unsigned long nr_pages) | |
766 | { | |
767 | return 0; | |
768 | } | |
769 | ||
770 | static bool is_subsection_map_empty(struct mem_section *ms) | |
771 | { | |
772 | return true; | |
773 | } | |
774 | ||
775 | static int fill_subsection_map(unsigned long pfn, unsigned long nr_pages) | |
776 | { | |
777 | return 0; | |
778 | } | |
779 | #endif /* CONFIG_SPARSEMEM_VMEMMAP */ | |
780 | ||
781 | /* | |
782 | * To deactivate a memory region, there are 3 cases to handle across | |
783 | * two configurations (SPARSEMEM_VMEMMAP={y,n}): | |
784 | * | |
785 | * 1. deactivation of a partial hot-added section (only possible in | |
786 | * the SPARSEMEM_VMEMMAP=y case). | |
787 | * a) section was present at memory init. | |
788 | * b) section was hot-added post memory init. | |
789 | * 2. deactivation of a complete hot-added section. | |
790 | * 3. deactivation of a complete section from memory init. | |
791 | * | |
792 | * For 1, when subsection_map does not empty we will not be freeing the | |
793 | * usage map, but still need to free the vmemmap range. | |
794 | * | |
795 | * For 2 and 3, the SPARSEMEM_VMEMMAP={y,n} cases are unified | |
796 | */ | |
797 | static void section_deactivate(unsigned long pfn, unsigned long nr_pages, | |
798 | struct vmem_altmap *altmap) | |
799 | { | |
800 | struct mem_section *ms = __pfn_to_section(pfn); | |
801 | bool section_is_early = early_section(ms); | |
802 | struct page *memmap = NULL; | |
803 | bool empty; | |
804 | ||
805 | if (clear_subsection_map(pfn, nr_pages)) | |
806 | return; | |
807 | ||
808 | empty = is_subsection_map_empty(ms); | |
809 | if (empty) { | |
810 | unsigned long section_nr = pfn_to_section_nr(pfn); | |
811 | ||
812 | /* | |
813 | * When removing an early section, the usage map is kept (as the | |
814 | * usage maps of other sections fall into the same page). It | |
815 | * will be re-used when re-adding the section - which is then no | |
816 | * longer an early section. If the usage map is PageReserved, it | |
817 | * was allocated during boot. | |
818 | */ | |
819 | if (!PageReserved(virt_to_page(ms->usage))) { | |
820 | kfree(ms->usage); | |
821 | ms->usage = NULL; | |
822 | } | |
823 | memmap = sparse_decode_mem_map(ms->section_mem_map, section_nr); | |
824 | /* | |
825 | * Mark the section invalid so that valid_section() | |
826 | * return false. This prevents code from dereferencing | |
827 | * ms->usage array. | |
828 | */ | |
829 | ms->section_mem_map &= ~SECTION_HAS_MEM_MAP; | |
830 | } | |
831 | ||
832 | /* | |
833 | * The memmap of early sections is always fully populated. See | |
834 | * section_activate() and pfn_valid() . | |
835 | */ | |
836 | if (!section_is_early) | |
837 | depopulate_section_memmap(pfn, nr_pages, altmap); | |
838 | else if (memmap) | |
839 | free_map_bootmem(memmap); | |
840 | ||
841 | if (empty) | |
842 | ms->section_mem_map = (unsigned long)NULL; | |
843 | } | |
844 | ||
845 | static struct page * __meminit section_activate(int nid, unsigned long pfn, | |
846 | unsigned long nr_pages, struct vmem_altmap *altmap) | |
847 | { | |
848 | struct mem_section *ms = __pfn_to_section(pfn); | |
849 | struct mem_section_usage *usage = NULL; | |
850 | struct page *memmap; | |
851 | int rc = 0; | |
852 | ||
853 | if (!ms->usage) { | |
854 | usage = kzalloc(mem_section_usage_size(), GFP_KERNEL); | |
855 | if (!usage) | |
856 | return ERR_PTR(-ENOMEM); | |
857 | ms->usage = usage; | |
858 | } | |
859 | ||
860 | rc = fill_subsection_map(pfn, nr_pages); | |
861 | if (rc) { | |
862 | if (usage) | |
863 | ms->usage = NULL; | |
864 | kfree(usage); | |
865 | return ERR_PTR(rc); | |
866 | } | |
867 | ||
868 | /* | |
869 | * The early init code does not consider partially populated | |
870 | * initial sections, it simply assumes that memory will never be | |
871 | * referenced. If we hot-add memory into such a section then we | |
872 | * do not need to populate the memmap and can simply reuse what | |
873 | * is already there. | |
874 | */ | |
875 | if (nr_pages < PAGES_PER_SECTION && early_section(ms)) | |
876 | return pfn_to_page(pfn); | |
877 | ||
878 | memmap = populate_section_memmap(pfn, nr_pages, nid, altmap); | |
879 | if (!memmap) { | |
880 | section_deactivate(pfn, nr_pages, altmap); | |
881 | return ERR_PTR(-ENOMEM); | |
882 | } | |
883 | ||
884 | return memmap; | |
885 | } | |
886 | ||
887 | /** | |
888 | * sparse_add_section - add a memory section, or populate an existing one | |
889 | * @nid: The node to add section on | |
890 | * @start_pfn: start pfn of the memory range | |
891 | * @nr_pages: number of pfns to add in the section | |
892 | * @altmap: device page map | |
893 | * | |
894 | * This is only intended for hotplug. | |
895 | * | |
896 | * Note that only VMEMMAP supports sub-section aligned hotplug, | |
897 | * the proper alignment and size are gated by check_pfn_span(). | |
898 | * | |
899 | * | |
900 | * Return: | |
901 | * * 0 - On success. | |
902 | * * -EEXIST - Section has been present. | |
903 | * * -ENOMEM - Out of memory. | |
904 | */ | |
905 | int __meminit sparse_add_section(int nid, unsigned long start_pfn, | |
906 | unsigned long nr_pages, struct vmem_altmap *altmap) | |
907 | { | |
908 | unsigned long section_nr = pfn_to_section_nr(start_pfn); | |
909 | struct mem_section *ms; | |
910 | struct page *memmap; | |
911 | int ret; | |
912 | ||
913 | ret = sparse_index_init(section_nr, nid); | |
914 | if (ret < 0) | |
915 | return ret; | |
916 | ||
917 | memmap = section_activate(nid, start_pfn, nr_pages, altmap); | |
918 | if (IS_ERR(memmap)) | |
919 | return PTR_ERR(memmap); | |
920 | ||
921 | /* | |
922 | * Poison uninitialized struct pages in order to catch invalid flags | |
923 | * combinations. | |
924 | */ | |
925 | page_init_poison(memmap, sizeof(struct page) * nr_pages); | |
926 | ||
927 | ms = __nr_to_section(section_nr); | |
928 | set_section_nid(section_nr, nid); | |
929 | section_mark_present(ms); | |
930 | ||
931 | /* Align memmap to section boundary in the subsection case */ | |
932 | if (section_nr_to_pfn(section_nr) != start_pfn) | |
933 | memmap = pfn_to_page(section_nr_to_pfn(section_nr)); | |
934 | sparse_init_one_section(ms, section_nr, memmap, ms->usage, 0); | |
935 | ||
936 | return 0; | |
937 | } | |
938 | ||
939 | #ifdef CONFIG_MEMORY_FAILURE | |
940 | static void clear_hwpoisoned_pages(struct page *memmap, int nr_pages) | |
941 | { | |
942 | int i; | |
943 | ||
944 | /* | |
945 | * A further optimization is to have per section refcounted | |
946 | * num_poisoned_pages. But that would need more space per memmap, so | |
947 | * for now just do a quick global check to speed up this routine in the | |
948 | * absence of bad pages. | |
949 | */ | |
950 | if (atomic_long_read(&num_poisoned_pages) == 0) | |
951 | return; | |
952 | ||
953 | for (i = 0; i < nr_pages; i++) { | |
954 | if (PageHWPoison(&memmap[i])) { | |
955 | num_poisoned_pages_dec(); | |
956 | ClearPageHWPoison(&memmap[i]); | |
957 | } | |
958 | } | |
959 | } | |
960 | #else | |
961 | static inline void clear_hwpoisoned_pages(struct page *memmap, int nr_pages) | |
962 | { | |
963 | } | |
964 | #endif | |
965 | ||
966 | void sparse_remove_section(struct mem_section *ms, unsigned long pfn, | |
967 | unsigned long nr_pages, unsigned long map_offset, | |
968 | struct vmem_altmap *altmap) | |
969 | { | |
970 | clear_hwpoisoned_pages(pfn_to_page(pfn) + map_offset, | |
971 | nr_pages - map_offset); | |
972 | section_deactivate(pfn, nr_pages, altmap); | |
973 | } | |
974 | #endif /* CONFIG_MEMORY_HOTPLUG */ |