Line data Source code
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 9 : static inline void set_section_nid(unsigned long section_nr, int nid)
57 : {
58 9 : }
59 : #endif
60 :
61 : #ifdef CONFIG_SPARSEMEM_EXTREME
62 1 : static noinline struct mem_section __ref *sparse_index_alloc(int nid)
63 : {
64 1 : struct mem_section *section = NULL;
65 1 : unsigned long array_size = SECTIONS_PER_ROOT *
66 : sizeof(struct mem_section);
67 :
68 1 : if (slab_is_available()) {
69 0 : section = kzalloc_node(array_size, GFP_KERNEL, nid);
70 : } else {
71 1 : section = memblock_alloc_node(array_size, SMP_CACHE_BYTES,
72 : nid);
73 1 : if (!section)
74 0 : panic("%s: Failed to allocate %lu bytes nid=%d\n",
75 : __func__, array_size, nid);
76 : }
77 :
78 1 : return section;
79 : }
80 :
81 9 : static int __meminit sparse_index_init(unsigned long section_nr, int nid)
82 : {
83 9 : unsigned long root = SECTION_NR_TO_ROOT(section_nr);
84 9 : 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 9 : if (mem_section[root])
94 : return 0;
95 :
96 1 : section = sparse_index_alloc(nid);
97 1 : if (!section)
98 : return -ENOMEM;
99 :
100 1 : mem_section[root] = section;
101 :
102 1 : 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 8 : unsigned long __section_nr(struct mem_section *ms)
113 : {
114 8 : unsigned long root_nr;
115 8 : struct mem_section *root = NULL;
116 :
117 8 : for (root_nr = 0; root_nr < NR_SECTION_ROOTS; root_nr++) {
118 8 : root = __nr_to_section(root_nr * SECTIONS_PER_ROOT);
119 8 : if (!root)
120 0 : continue;
121 :
122 8 : if ((ms >= root) && (ms < (root + SECTIONS_PER_ROOT)))
123 : break;
124 : }
125 :
126 8 : VM_BUG_ON(!root);
127 :
128 8 : 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 8 : static inline unsigned long sparse_encode_early_nid(int nid)
144 : {
145 8 : return (nid << SECTION_NID_SHIFT);
146 : }
147 :
148 8 : static inline int sparse_early_nid(struct mem_section *section)
149 : {
150 8 : return (section->section_mem_map >> SECTION_NID_SHIFT);
151 : }
152 :
153 : /* Validate the physical addressing limitations of the model */
154 2 : void __meminit mminit_validate_memmodel_limits(unsigned long *start_pfn,
155 : unsigned long *end_pfn)
156 : {
157 2 : 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 2 : if (*start_pfn > max_sparsemem_pfn) {
164 0 : 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 0 : WARN_ON_ONCE(1);
168 0 : *start_pfn = max_sparsemem_pfn;
169 0 : *end_pfn = max_sparsemem_pfn;
170 2 : } else if (*end_pfn > max_sparsemem_pfn) {
171 0 : 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 0 : WARN_ON_ONCE(1);
175 0 : *end_pfn = max_sparsemem_pfn;
176 : }
177 2 : }
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 8 : static void section_mark_present(struct mem_section *ms)
190 : {
191 8 : unsigned long section_nr = __section_nr(ms);
192 :
193 8 : if (section_nr > __highest_present_section_nr)
194 7 : __highest_present_section_nr = section_nr;
195 :
196 8 : ms->section_mem_map |= SECTION_MARKED_PRESENT;
197 8 : }
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 1 : static inline unsigned long first_present_section_nr(void)
206 : {
207 1 : 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 2 : void __init subsection_map_init(unsigned long pfn, unsigned long nr_pages)
247 : {
248 2 : }
249 : #endif
250 :
251 : /* Record a memory area against a node. */
252 2 : static void __init memory_present(int nid, unsigned long start, unsigned long end)
253 : {
254 2 : unsigned long pfn;
255 :
256 : #ifdef CONFIG_SPARSEMEM_EXTREME
257 2 : if (unlikely(!mem_section)) {
258 1 : unsigned long size, align;
259 :
260 1 : size = sizeof(struct mem_section*) * NR_SECTION_ROOTS;
261 1 : align = 1 << (INTERNODE_CACHE_SHIFT);
262 1 : mem_section = memblock_alloc(size, align);
263 1 : if (!mem_section)
264 0 : panic("%s: Failed to allocate %lu bytes align=0x%lx\n",
265 : __func__, size, align);
266 : }
267 : #endif
268 :
269 2 : start &= PAGE_SECTION_MASK;
270 2 : mminit_validate_memmodel_limits(&start, &end);
271 11 : for (pfn = start; pfn < end; pfn += PAGES_PER_SECTION) {
272 9 : unsigned long section = pfn_to_section_nr(pfn);
273 9 : struct mem_section *ms;
274 :
275 9 : sparse_index_init(section, nid);
276 9 : set_section_nid(section, nid);
277 :
278 9 : ms = __nr_to_section(section);
279 9 : if (!ms->section_mem_map) {
280 8 : ms->section_mem_map = sparse_encode_early_nid(nid) |
281 : SECTION_IS_ONLINE;
282 8 : section_mark_present(ms);
283 : }
284 : }
285 2 : }
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 1 : static void __init memblocks_present(void)
293 : {
294 1 : unsigned long start, end;
295 1 : int i, nid;
296 :
297 3 : for_each_mem_pfn_range(i, MAX_NUMNODES, &start, &end, &nid)
298 2 : memory_present(nid, start, end);
299 1 : }
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 8 : static unsigned long sparse_encode_mem_map(struct page *mem_map, unsigned long pnum)
307 : {
308 8 : unsigned long coded_mem_map =
309 8 : (unsigned long)(mem_map - (section_nr_to_pfn(pnum)));
310 8 : BUILD_BUG_ON(SECTION_MAP_LAST_BIT > (1UL<<PFN_SECTION_SHIFT));
311 8 : BUG_ON(coded_mem_map & ~SECTION_MAP_MASK);
312 8 : 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 8 : 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 8 : ms->section_mem_map &= ~SECTION_MAP_MASK;
332 8 : ms->section_mem_map |= sparse_encode_mem_map(mem_map, pnum)
333 8 : | SECTION_HAS_MEM_MAP | flags;
334 8 : ms->usage = usage;
335 8 : }
336 :
337 9 : static unsigned long usemap_size(void)
338 : {
339 9 : return BITS_TO_LONGS(SECTION_BLOCKFLAGS_BITS) * sizeof(unsigned long);
340 : }
341 :
342 9 : size_t mem_section_usage_size(void)
343 : {
344 9 : 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 1 : sparse_early_usemaps_alloc_pgdat_section(struct pglist_data *pgdat,
422 : unsigned long size)
423 : {
424 1 : return memblock_alloc_node(size, SMP_CACHE_BYTES, pgdat->node_id);
425 : }
426 :
427 8 : static void __init check_usemap_section_nr(int nid,
428 : struct mem_section_usage *usage)
429 : {
430 8 : }
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 10 : static unsigned long __init section_map_size(void)
441 : {
442 10 : return PAGE_ALIGN(sizeof(struct page) * PAGES_PER_SECTION);
443 : }
444 :
445 8 : struct page __init *__populate_section_memmap(unsigned long pfn,
446 : unsigned long nr_pages, int nid, struct vmem_altmap *altmap)
447 : {
448 8 : unsigned long size = section_map_size();
449 8 : struct page *map = sparse_buffer_alloc(size);
450 8 : phys_addr_t addr = __pa(MAX_DMA_ADDRESS);
451 :
452 8 : if (map)
453 : return map;
454 :
455 0 : map = memblock_alloc_try_nid_raw(size, size, addr,
456 : MEMBLOCK_ALLOC_ACCESSIBLE, nid);
457 0 : if (!map)
458 0 : 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 0 : static inline void __meminit sparse_buffer_free(unsigned long size)
469 : {
470 0 : WARN_ON(!sparsemap_buf || size == 0);
471 0 : memblock_free_early(__pa(sparsemap_buf), size);
472 0 : }
473 :
474 1 : static void __init sparse_buffer_init(unsigned long size, int nid)
475 : {
476 1 : phys_addr_t addr = __pa(MAX_DMA_ADDRESS);
477 1 : 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 1 : sparsemap_buf = memblock_alloc_exact_nid_raw(size, section_map_size(),
484 : addr, MEMBLOCK_ALLOC_ACCESSIBLE, nid);
485 1 : sparsemap_buf_end = sparsemap_buf + size;
486 1 : }
487 :
488 1 : static void __init sparse_buffer_fini(void)
489 : {
490 1 : unsigned long size = sparsemap_buf_end - sparsemap_buf;
491 :
492 1 : if (sparsemap_buf && size > 0)
493 0 : sparse_buffer_free(size);
494 1 : sparsemap_buf = NULL;
495 1 : }
496 :
497 8 : void * __meminit sparse_buffer_alloc(unsigned long size)
498 : {
499 8 : void *ptr = NULL;
500 :
501 8 : if (sparsemap_buf) {
502 8 : ptr = (void *) roundup((unsigned long)sparsemap_buf, size);
503 8 : if (ptr + size > sparsemap_buf_end)
504 : ptr = NULL;
505 : else {
506 : /* Free redundant aligned space */
507 8 : if ((unsigned long)(ptr - sparsemap_buf) > 0)
508 0 : sparse_buffer_free((unsigned long)(ptr - sparsemap_buf));
509 8 : sparsemap_buf = ptr + size;
510 : }
511 : }
512 8 : return ptr;
513 : }
514 :
515 1 : void __weak __meminit vmemmap_populate_print_last(void)
516 : {
517 1 : }
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 1 : static void __init sparse_init_nid(int nid, unsigned long pnum_begin,
524 : unsigned long pnum_end,
525 : unsigned long map_count)
526 : {
527 1 : struct mem_section_usage *usage;
528 1 : unsigned long pnum;
529 1 : struct page *map;
530 :
531 1 : usage = sparse_early_usemaps_alloc_pgdat_section(NODE_DATA(nid),
532 : mem_section_usage_size() * map_count);
533 1 : if (!usage) {
534 0 : pr_err("%s: node[%d] usemap allocation failed", __func__, nid);
535 0 : goto failed;
536 : }
537 1 : sparse_buffer_init(map_count * section_map_size(), nid);
538 9 : for_each_present_section_nr(pnum_begin, pnum) {
539 8 : unsigned long pfn = section_nr_to_pfn(pnum);
540 :
541 8 : if (pnum >= pnum_end)
542 : break;
543 :
544 8 : map = __populate_section_memmap(pfn, PAGES_PER_SECTION,
545 : nid, NULL);
546 8 : if (!map) {
547 0 : pr_err("%s: node[%d] memory map backing failed. Some memory will not be available.",
548 : __func__, nid);
549 0 : pnum_begin = pnum;
550 0 : goto failed;
551 : }
552 8 : check_usemap_section_nr(nid, usage);
553 8 : sparse_init_one_section(__nr_to_section(pnum), pnum, map, usage,
554 : SECTION_IS_EARLY);
555 8 : usage = (void *) usage + mem_section_usage_size();
556 : }
557 1 : sparse_buffer_fini();
558 1 : return;
559 0 : failed:
560 : /* We failed to allocate, mark all the following pnums as not present */
561 0 : for_each_present_section_nr(pnum_begin, pnum) {
562 0 : struct mem_section *ms;
563 :
564 0 : if (pnum >= pnum_end)
565 : break;
566 0 : ms = __nr_to_section(pnum);
567 0 : ms->section_mem_map = 0;
568 : }
569 : }
570 :
571 : /*
572 : * Allocate the accumulated non-linear sections, allocate a mem_map
573 : * for each and record the physical to section mapping.
574 : */
575 1 : void __init sparse_init(void)
576 : {
577 1 : unsigned long pnum_end, pnum_begin, map_count = 1;
578 1 : int nid_begin;
579 :
580 1 : memblocks_present();
581 :
582 1 : pnum_begin = first_present_section_nr();
583 1 : nid_begin = sparse_early_nid(__nr_to_section(pnum_begin));
584 :
585 : /* Setup pageblock_order for HUGETLB_PAGE_SIZE_VARIABLE */
586 1 : set_pageblock_order();
587 :
588 8 : for_each_present_section_nr(pnum_begin + 1, pnum_end) {
589 7 : int nid = sparse_early_nid(__nr_to_section(pnum_end));
590 :
591 7 : if (nid == nid_begin) {
592 7 : map_count++;
593 7 : continue;
594 : }
595 : /* Init node with sections in range [pnum_begin, pnum_end) */
596 0 : sparse_init_nid(nid_begin, pnum_begin, pnum_end, map_count);
597 0 : nid_begin = nid;
598 0 : pnum_begin = pnum_end;
599 0 : map_count = 1;
600 : }
601 : /* cover the last node */
602 1 : sparse_init_nid(nid_begin, pnum_begin, pnum_end, map_count);
603 1 : vmemmap_populate_print_last();
604 1 : }
605 :
606 : #ifdef CONFIG_MEMORY_HOTPLUG
607 :
608 : /* Mark all memory sections within the pfn range as online */
609 : void online_mem_sections(unsigned long start_pfn, unsigned long end_pfn)
610 : {
611 : unsigned long pfn;
612 :
613 : for (pfn = start_pfn; pfn < end_pfn; pfn += PAGES_PER_SECTION) {
614 : unsigned long section_nr = pfn_to_section_nr(pfn);
615 : struct mem_section *ms;
616 :
617 : /* onlining code should never touch invalid ranges */
618 : if (WARN_ON(!valid_section_nr(section_nr)))
619 : continue;
620 :
621 : ms = __nr_to_section(section_nr);
622 : ms->section_mem_map |= SECTION_IS_ONLINE;
623 : }
624 : }
625 :
626 : #ifdef CONFIG_MEMORY_HOTREMOVE
627 : /* Mark all memory sections within the pfn range as offline */
628 : void offline_mem_sections(unsigned long start_pfn, unsigned long end_pfn)
629 : {
630 : unsigned long pfn;
631 :
632 : for (pfn = start_pfn; pfn < end_pfn; pfn += PAGES_PER_SECTION) {
633 : unsigned long section_nr = pfn_to_section_nr(pfn);
634 : struct mem_section *ms;
635 :
636 : /*
637 : * TODO this needs some double checking. Offlining code makes
638 : * sure to check pfn_valid but those checks might be just bogus
639 : */
640 : if (WARN_ON(!valid_section_nr(section_nr)))
641 : continue;
642 :
643 : ms = __nr_to_section(section_nr);
644 : ms->section_mem_map &= ~SECTION_IS_ONLINE;
645 : }
646 : }
647 : #endif
648 :
649 : #ifdef CONFIG_SPARSEMEM_VMEMMAP
650 : static struct page * __meminit populate_section_memmap(unsigned long pfn,
651 : unsigned long nr_pages, int nid, struct vmem_altmap *altmap)
652 : {
653 : return __populate_section_memmap(pfn, nr_pages, nid, altmap);
654 : }
655 :
656 : static void depopulate_section_memmap(unsigned long pfn, unsigned long nr_pages,
657 : struct vmem_altmap *altmap)
658 : {
659 : unsigned long start = (unsigned long) pfn_to_page(pfn);
660 : unsigned long end = start + nr_pages * sizeof(struct page);
661 :
662 : vmemmap_free(start, end, altmap);
663 : }
664 : static void free_map_bootmem(struct page *memmap)
665 : {
666 : unsigned long start = (unsigned long)memmap;
667 : unsigned long end = (unsigned long)(memmap + PAGES_PER_SECTION);
668 :
669 : vmemmap_free(start, end, NULL);
670 : }
671 :
672 : static int clear_subsection_map(unsigned long pfn, unsigned long nr_pages)
673 : {
674 : DECLARE_BITMAP(map, SUBSECTIONS_PER_SECTION) = { 0 };
675 : DECLARE_BITMAP(tmp, SUBSECTIONS_PER_SECTION) = { 0 };
676 : struct mem_section *ms = __pfn_to_section(pfn);
677 : unsigned long *subsection_map = ms->usage
678 : ? &ms->usage->subsection_map[0] : NULL;
679 :
680 : subsection_mask_set(map, pfn, nr_pages);
681 : if (subsection_map)
682 : bitmap_and(tmp, map, subsection_map, SUBSECTIONS_PER_SECTION);
683 :
684 : if (WARN(!subsection_map || !bitmap_equal(tmp, map, SUBSECTIONS_PER_SECTION),
685 : "section already deactivated (%#lx + %ld)\n",
686 : pfn, nr_pages))
687 : return -EINVAL;
688 :
689 : bitmap_xor(subsection_map, map, subsection_map, SUBSECTIONS_PER_SECTION);
690 : return 0;
691 : }
692 :
693 : static bool is_subsection_map_empty(struct mem_section *ms)
694 : {
695 : return bitmap_empty(&ms->usage->subsection_map[0],
696 : SUBSECTIONS_PER_SECTION);
697 : }
698 :
699 : static int fill_subsection_map(unsigned long pfn, unsigned long nr_pages)
700 : {
701 : struct mem_section *ms = __pfn_to_section(pfn);
702 : DECLARE_BITMAP(map, SUBSECTIONS_PER_SECTION) = { 0 };
703 : unsigned long *subsection_map;
704 : int rc = 0;
705 :
706 : subsection_mask_set(map, pfn, nr_pages);
707 :
708 : subsection_map = &ms->usage->subsection_map[0];
709 :
710 : if (bitmap_empty(map, SUBSECTIONS_PER_SECTION))
711 : rc = -EINVAL;
712 : else if (bitmap_intersects(map, subsection_map, SUBSECTIONS_PER_SECTION))
713 : rc = -EEXIST;
714 : else
715 : bitmap_or(subsection_map, map, subsection_map,
716 : SUBSECTIONS_PER_SECTION);
717 :
718 : return rc;
719 : }
720 : #else
721 : struct page * __meminit populate_section_memmap(unsigned long pfn,
722 : unsigned long nr_pages, int nid, struct vmem_altmap *altmap)
723 : {
724 : return kvmalloc_node(array_size(sizeof(struct page),
725 : PAGES_PER_SECTION), GFP_KERNEL, nid);
726 : }
727 :
728 : static void depopulate_section_memmap(unsigned long pfn, unsigned long nr_pages,
729 : struct vmem_altmap *altmap)
730 : {
731 : kvfree(pfn_to_page(pfn));
732 : }
733 :
734 : static void free_map_bootmem(struct page *memmap)
735 : {
736 : unsigned long maps_section_nr, removing_section_nr, i;
737 : unsigned long magic, nr_pages;
738 : struct page *page = virt_to_page(memmap);
739 :
740 : nr_pages = PAGE_ALIGN(PAGES_PER_SECTION * sizeof(struct page))
741 : >> PAGE_SHIFT;
742 :
743 : for (i = 0; i < nr_pages; i++, page++) {
744 : magic = (unsigned long) page->freelist;
745 :
746 : BUG_ON(magic == NODE_INFO);
747 :
748 : maps_section_nr = pfn_to_section_nr(page_to_pfn(page));
749 : removing_section_nr = page_private(page);
750 :
751 : /*
752 : * When this function is called, the removing section is
753 : * logical offlined state. This means all pages are isolated
754 : * from page allocator. If removing section's memmap is placed
755 : * on the same section, it must not be freed.
756 : * If it is freed, page allocator may allocate it which will
757 : * be removed physically soon.
758 : */
759 : if (maps_section_nr != removing_section_nr)
760 : put_page_bootmem(page);
761 : }
762 : }
763 :
764 : static int clear_subsection_map(unsigned long pfn, unsigned long nr_pages)
765 : {
766 : return 0;
767 : }
768 :
769 : static bool is_subsection_map_empty(struct mem_section *ms)
770 : {
771 : return true;
772 : }
773 :
774 : static int fill_subsection_map(unsigned long pfn, unsigned long nr_pages)
775 : {
776 : return 0;
777 : }
778 : #endif /* CONFIG_SPARSEMEM_VMEMMAP */
779 :
780 : /*
781 : * To deactivate a memory region, there are 3 cases to handle across
782 : * two configurations (SPARSEMEM_VMEMMAP={y,n}):
783 : *
784 : * 1. deactivation of a partial hot-added section (only possible in
785 : * the SPARSEMEM_VMEMMAP=y case).
786 : * a) section was present at memory init.
787 : * b) section was hot-added post memory init.
788 : * 2. deactivation of a complete hot-added section.
789 : * 3. deactivation of a complete section from memory init.
790 : *
791 : * For 1, when subsection_map does not empty we will not be freeing the
792 : * usage map, but still need to free the vmemmap range.
793 : *
794 : * For 2 and 3, the SPARSEMEM_VMEMMAP={y,n} cases are unified
795 : */
796 : static void section_deactivate(unsigned long pfn, unsigned long nr_pages,
797 : struct vmem_altmap *altmap)
798 : {
799 : struct mem_section *ms = __pfn_to_section(pfn);
800 : bool section_is_early = early_section(ms);
801 : struct page *memmap = NULL;
802 : bool empty;
803 :
804 : if (clear_subsection_map(pfn, nr_pages))
805 : return;
806 :
807 : empty = is_subsection_map_empty(ms);
808 : if (empty) {
809 : unsigned long section_nr = pfn_to_section_nr(pfn);
810 :
811 : /*
812 : * When removing an early section, the usage map is kept (as the
813 : * usage maps of other sections fall into the same page). It
814 : * will be re-used when re-adding the section - which is then no
815 : * longer an early section. If the usage map is PageReserved, it
816 : * was allocated during boot.
817 : */
818 : if (!PageReserved(virt_to_page(ms->usage))) {
819 : kfree(ms->usage);
820 : ms->usage = NULL;
821 : }
822 : memmap = sparse_decode_mem_map(ms->section_mem_map, section_nr);
823 : /*
824 : * Mark the section invalid so that valid_section()
825 : * return false. This prevents code from dereferencing
826 : * ms->usage array.
827 : */
828 : ms->section_mem_map &= ~SECTION_HAS_MEM_MAP;
829 : }
830 :
831 : /*
832 : * The memmap of early sections is always fully populated. See
833 : * section_activate() and pfn_valid() .
834 : */
835 : if (!section_is_early)
836 : depopulate_section_memmap(pfn, nr_pages, altmap);
837 : else if (memmap)
838 : free_map_bootmem(memmap);
839 :
840 : if (empty)
841 : ms->section_mem_map = (unsigned long)NULL;
842 : }
843 :
844 : static struct page * __meminit section_activate(int nid, unsigned long pfn,
845 : unsigned long nr_pages, struct vmem_altmap *altmap)
846 : {
847 : struct mem_section *ms = __pfn_to_section(pfn);
848 : struct mem_section_usage *usage = NULL;
849 : struct page *memmap;
850 : int rc = 0;
851 :
852 : if (!ms->usage) {
853 : usage = kzalloc(mem_section_usage_size(), GFP_KERNEL);
854 : if (!usage)
855 : return ERR_PTR(-ENOMEM);
856 : ms->usage = usage;
857 : }
858 :
859 : rc = fill_subsection_map(pfn, nr_pages);
860 : if (rc) {
861 : if (usage)
862 : ms->usage = NULL;
863 : kfree(usage);
864 : return ERR_PTR(rc);
865 : }
866 :
867 : /*
868 : * The early init code does not consider partially populated
869 : * initial sections, it simply assumes that memory will never be
870 : * referenced. If we hot-add memory into such a section then we
871 : * do not need to populate the memmap and can simply reuse what
872 : * is already there.
873 : */
874 : if (nr_pages < PAGES_PER_SECTION && early_section(ms))
875 : return pfn_to_page(pfn);
876 :
877 : memmap = populate_section_memmap(pfn, nr_pages, nid, altmap);
878 : if (!memmap) {
879 : section_deactivate(pfn, nr_pages, altmap);
880 : return ERR_PTR(-ENOMEM);
881 : }
882 :
883 : return memmap;
884 : }
885 :
886 : /**
887 : * sparse_add_section - add a memory section, or populate an existing one
888 : * @nid: The node to add section on
889 : * @start_pfn: start pfn of the memory range
890 : * @nr_pages: number of pfns to add in the section
891 : * @altmap: device page map
892 : *
893 : * This is only intended for hotplug.
894 : *
895 : * Note that only VMEMMAP supports sub-section aligned hotplug,
896 : * the proper alignment and size are gated by check_pfn_span().
897 : *
898 : *
899 : * Return:
900 : * * 0 - On success.
901 : * * -EEXIST - Section has been present.
902 : * * -ENOMEM - Out of memory.
903 : */
904 : int __meminit sparse_add_section(int nid, unsigned long start_pfn,
905 : unsigned long nr_pages, struct vmem_altmap *altmap)
906 : {
907 : unsigned long section_nr = pfn_to_section_nr(start_pfn);
908 : struct mem_section *ms;
909 : struct page *memmap;
910 : int ret;
911 :
912 : ret = sparse_index_init(section_nr, nid);
913 : if (ret < 0)
914 : return ret;
915 :
916 : memmap = section_activate(nid, start_pfn, nr_pages, altmap);
917 : if (IS_ERR(memmap))
918 : return PTR_ERR(memmap);
919 :
920 : /*
921 : * Poison uninitialized struct pages in order to catch invalid flags
922 : * combinations.
923 : */
924 : page_init_poison(memmap, sizeof(struct page) * nr_pages);
925 :
926 : ms = __nr_to_section(section_nr);
927 : set_section_nid(section_nr, nid);
928 : section_mark_present(ms);
929 :
930 : /* Align memmap to section boundary in the subsection case */
931 : if (section_nr_to_pfn(section_nr) != start_pfn)
932 : memmap = pfn_to_page(section_nr_to_pfn(section_nr));
933 : sparse_init_one_section(ms, section_nr, memmap, ms->usage, 0);
934 :
935 : return 0;
936 : }
937 :
938 : #ifdef CONFIG_MEMORY_FAILURE
939 : static void clear_hwpoisoned_pages(struct page *memmap, int nr_pages)
940 : {
941 : int i;
942 :
943 : /*
944 : * A further optimization is to have per section refcounted
945 : * num_poisoned_pages. But that would need more space per memmap, so
946 : * for now just do a quick global check to speed up this routine in the
947 : * absence of bad pages.
948 : */
949 : if (atomic_long_read(&num_poisoned_pages) == 0)
950 : return;
951 :
952 : for (i = 0; i < nr_pages; i++) {
953 : if (PageHWPoison(&memmap[i])) {
954 : num_poisoned_pages_dec();
955 : ClearPageHWPoison(&memmap[i]);
956 : }
957 : }
958 : }
959 : #else
960 : static inline void clear_hwpoisoned_pages(struct page *memmap, int nr_pages)
961 : {
962 : }
963 : #endif
964 :
965 : void sparse_remove_section(struct mem_section *ms, unsigned long pfn,
966 : unsigned long nr_pages, unsigned long map_offset,
967 : struct vmem_altmap *altmap)
968 : {
969 : clear_hwpoisoned_pages(pfn_to_page(pfn) + map_offset,
970 : nr_pages - map_offset);
971 : section_deactivate(pfn, nr_pages, altmap);
972 : }
973 : #endif /* CONFIG_MEMORY_HOTPLUG */
|
