Loading...
1// SPDX-License-Identifier: GPL-2.0
2/*
3 * Virtual Memory Map support
4 *
5 * (C) 2007 sgi. Christoph Lameter.
6 *
7 * Virtual memory maps allow VM primitives pfn_to_page, page_to_pfn,
8 * virt_to_page, page_address() to be implemented as a base offset
9 * calculation without memory access.
10 *
11 * However, virtual mappings need a page table and TLBs. Many Linux
12 * architectures already map their physical space using 1-1 mappings
13 * via TLBs. For those arches the virtual memory map is essentially
14 * for free if we use the same page size as the 1-1 mappings. In that
15 * case the overhead consists of a few additional pages that are
16 * allocated to create a view of memory for vmemmap.
17 *
18 * The architecture is expected to provide a vmemmap_populate() function
19 * to instantiate the mapping.
20 */
21#include <linux/mm.h>
22#include <linux/mmzone.h>
23#include <linux/memblock.h>
24#include <linux/memremap.h>
25#include <linux/highmem.h>
26#include <linux/slab.h>
27#include <linux/spinlock.h>
28#include <linux/vmalloc.h>
29#include <linux/sched.h>
30#include <linux/pgtable.h>
31#include <linux/bootmem_info.h>
32
33#include <asm/dma.h>
34#include <asm/pgalloc.h>
35#include <asm/tlbflush.h>
36
37/**
38 * struct vmemmap_remap_walk - walk vmemmap page table
39 *
40 * @remap_pte: called for each lowest-level entry (PTE).
41 * @nr_walked: the number of walked pte.
42 * @reuse_page: the page which is reused for the tail vmemmap pages.
43 * @reuse_addr: the virtual address of the @reuse_page page.
44 * @vmemmap_pages: the list head of the vmemmap pages that can be freed
45 * or is mapped from.
46 */
47struct vmemmap_remap_walk {
48 void (*remap_pte)(pte_t *pte, unsigned long addr,
49 struct vmemmap_remap_walk *walk);
50 unsigned long nr_walked;
51 struct page *reuse_page;
52 unsigned long reuse_addr;
53 struct list_head *vmemmap_pages;
54};
55
56static int split_vmemmap_huge_pmd(pmd_t *pmd, unsigned long start,
57 struct vmemmap_remap_walk *walk)
58{
59 pmd_t __pmd;
60 int i;
61 unsigned long addr = start;
62 struct page *page = pmd_page(*pmd);
63 pte_t *pgtable = pte_alloc_one_kernel(&init_mm);
64
65 if (!pgtable)
66 return -ENOMEM;
67
68 pmd_populate_kernel(&init_mm, &__pmd, pgtable);
69
70 for (i = 0; i < PMD_SIZE / PAGE_SIZE; i++, addr += PAGE_SIZE) {
71 pte_t entry, *pte;
72 pgprot_t pgprot = PAGE_KERNEL;
73
74 entry = mk_pte(page + i, pgprot);
75 pte = pte_offset_kernel(&__pmd, addr);
76 set_pte_at(&init_mm, addr, pte, entry);
77 }
78
79 /* Make pte visible before pmd. See comment in __pte_alloc(). */
80 smp_wmb();
81 pmd_populate_kernel(&init_mm, pmd, pgtable);
82
83 flush_tlb_kernel_range(start, start + PMD_SIZE);
84
85 return 0;
86}
87
88static void vmemmap_pte_range(pmd_t *pmd, unsigned long addr,
89 unsigned long end,
90 struct vmemmap_remap_walk *walk)
91{
92 pte_t *pte = pte_offset_kernel(pmd, addr);
93
94 /*
95 * The reuse_page is found 'first' in table walk before we start
96 * remapping (which is calling @walk->remap_pte).
97 */
98 if (!walk->reuse_page) {
99 walk->reuse_page = pte_page(*pte);
100 /*
101 * Because the reuse address is part of the range that we are
102 * walking, skip the reuse address range.
103 */
104 addr += PAGE_SIZE;
105 pte++;
106 walk->nr_walked++;
107 }
108
109 for (; addr != end; addr += PAGE_SIZE, pte++) {
110 walk->remap_pte(pte, addr, walk);
111 walk->nr_walked++;
112 }
113}
114
115static int vmemmap_pmd_range(pud_t *pud, unsigned long addr,
116 unsigned long end,
117 struct vmemmap_remap_walk *walk)
118{
119 pmd_t *pmd;
120 unsigned long next;
121
122 pmd = pmd_offset(pud, addr);
123 do {
124 if (pmd_leaf(*pmd)) {
125 int ret;
126
127 ret = split_vmemmap_huge_pmd(pmd, addr & PMD_MASK, walk);
128 if (ret)
129 return ret;
130 }
131 next = pmd_addr_end(addr, end);
132 vmemmap_pte_range(pmd, addr, next, walk);
133 } while (pmd++, addr = next, addr != end);
134
135 return 0;
136}
137
138static int vmemmap_pud_range(p4d_t *p4d, unsigned long addr,
139 unsigned long end,
140 struct vmemmap_remap_walk *walk)
141{
142 pud_t *pud;
143 unsigned long next;
144
145 pud = pud_offset(p4d, addr);
146 do {
147 int ret;
148
149 next = pud_addr_end(addr, end);
150 ret = vmemmap_pmd_range(pud, addr, next, walk);
151 if (ret)
152 return ret;
153 } while (pud++, addr = next, addr != end);
154
155 return 0;
156}
157
158static int vmemmap_p4d_range(pgd_t *pgd, unsigned long addr,
159 unsigned long end,
160 struct vmemmap_remap_walk *walk)
161{
162 p4d_t *p4d;
163 unsigned long next;
164
165 p4d = p4d_offset(pgd, addr);
166 do {
167 int ret;
168
169 next = p4d_addr_end(addr, end);
170 ret = vmemmap_pud_range(p4d, addr, next, walk);
171 if (ret)
172 return ret;
173 } while (p4d++, addr = next, addr != end);
174
175 return 0;
176}
177
178static int vmemmap_remap_range(unsigned long start, unsigned long end,
179 struct vmemmap_remap_walk *walk)
180{
181 unsigned long addr = start;
182 unsigned long next;
183 pgd_t *pgd;
184
185 VM_BUG_ON(!IS_ALIGNED(start, PAGE_SIZE));
186 VM_BUG_ON(!IS_ALIGNED(end, PAGE_SIZE));
187
188 pgd = pgd_offset_k(addr);
189 do {
190 int ret;
191
192 next = pgd_addr_end(addr, end);
193 ret = vmemmap_p4d_range(pgd, addr, next, walk);
194 if (ret)
195 return ret;
196 } while (pgd++, addr = next, addr != end);
197
198 /*
199 * We only change the mapping of the vmemmap virtual address range
200 * [@start + PAGE_SIZE, end), so we only need to flush the TLB which
201 * belongs to the range.
202 */
203 flush_tlb_kernel_range(start + PAGE_SIZE, end);
204
205 return 0;
206}
207
208/*
209 * Free a vmemmap page. A vmemmap page can be allocated from the memblock
210 * allocator or buddy allocator. If the PG_reserved flag is set, it means
211 * that it allocated from the memblock allocator, just free it via the
212 * free_bootmem_page(). Otherwise, use __free_page().
213 */
214static inline void free_vmemmap_page(struct page *page)
215{
216 if (PageReserved(page))
217 free_bootmem_page(page);
218 else
219 __free_page(page);
220}
221
222/* Free a list of the vmemmap pages */
223static void free_vmemmap_page_list(struct list_head *list)
224{
225 struct page *page, *next;
226
227 list_for_each_entry_safe(page, next, list, lru) {
228 list_del(&page->lru);
229 free_vmemmap_page(page);
230 }
231}
232
233static void vmemmap_remap_pte(pte_t *pte, unsigned long addr,
234 struct vmemmap_remap_walk *walk)
235{
236 /*
237 * Remap the tail pages as read-only to catch illegal write operation
238 * to the tail pages.
239 */
240 pgprot_t pgprot = PAGE_KERNEL_RO;
241 pte_t entry = mk_pte(walk->reuse_page, pgprot);
242 struct page *page = pte_page(*pte);
243
244 list_add_tail(&page->lru, walk->vmemmap_pages);
245 set_pte_at(&init_mm, addr, pte, entry);
246}
247
248static void vmemmap_restore_pte(pte_t *pte, unsigned long addr,
249 struct vmemmap_remap_walk *walk)
250{
251 pgprot_t pgprot = PAGE_KERNEL;
252 struct page *page;
253 void *to;
254
255 BUG_ON(pte_page(*pte) != walk->reuse_page);
256
257 page = list_first_entry(walk->vmemmap_pages, struct page, lru);
258 list_del(&page->lru);
259 to = page_to_virt(page);
260 copy_page(to, (void *)walk->reuse_addr);
261
262 set_pte_at(&init_mm, addr, pte, mk_pte(page, pgprot));
263}
264
265/**
266 * vmemmap_remap_free - remap the vmemmap virtual address range [@start, @end)
267 * to the page which @reuse is mapped to, then free vmemmap
268 * which the range are mapped to.
269 * @start: start address of the vmemmap virtual address range that we want
270 * to remap.
271 * @end: end address of the vmemmap virtual address range that we want to
272 * remap.
273 * @reuse: reuse address.
274 *
275 * Return: %0 on success, negative error code otherwise.
276 */
277int vmemmap_remap_free(unsigned long start, unsigned long end,
278 unsigned long reuse)
279{
280 int ret;
281 LIST_HEAD(vmemmap_pages);
282 struct vmemmap_remap_walk walk = {
283 .remap_pte = vmemmap_remap_pte,
284 .reuse_addr = reuse,
285 .vmemmap_pages = &vmemmap_pages,
286 };
287
288 /*
289 * In order to make remapping routine most efficient for the huge pages,
290 * the routine of vmemmap page table walking has the following rules
291 * (see more details from the vmemmap_pte_range()):
292 *
293 * - The range [@start, @end) and the range [@reuse, @reuse + PAGE_SIZE)
294 * should be continuous.
295 * - The @reuse address is part of the range [@reuse, @end) that we are
296 * walking which is passed to vmemmap_remap_range().
297 * - The @reuse address is the first in the complete range.
298 *
299 * So we need to make sure that @start and @reuse meet the above rules.
300 */
301 BUG_ON(start - reuse != PAGE_SIZE);
302
303 mmap_write_lock(&init_mm);
304 ret = vmemmap_remap_range(reuse, end, &walk);
305 mmap_write_downgrade(&init_mm);
306
307 if (ret && walk.nr_walked) {
308 end = reuse + walk.nr_walked * PAGE_SIZE;
309 /*
310 * vmemmap_pages contains pages from the previous
311 * vmemmap_remap_range call which failed. These
312 * are pages which were removed from the vmemmap.
313 * They will be restored in the following call.
314 */
315 walk = (struct vmemmap_remap_walk) {
316 .remap_pte = vmemmap_restore_pte,
317 .reuse_addr = reuse,
318 .vmemmap_pages = &vmemmap_pages,
319 };
320
321 vmemmap_remap_range(reuse, end, &walk);
322 }
323 mmap_read_unlock(&init_mm);
324
325 free_vmemmap_page_list(&vmemmap_pages);
326
327 return ret;
328}
329
330static int alloc_vmemmap_page_list(unsigned long start, unsigned long end,
331 gfp_t gfp_mask, struct list_head *list)
332{
333 unsigned long nr_pages = (end - start) >> PAGE_SHIFT;
334 int nid = page_to_nid((struct page *)start);
335 struct page *page, *next;
336
337 while (nr_pages--) {
338 page = alloc_pages_node(nid, gfp_mask, 0);
339 if (!page)
340 goto out;
341 list_add_tail(&page->lru, list);
342 }
343
344 return 0;
345out:
346 list_for_each_entry_safe(page, next, list, lru)
347 __free_pages(page, 0);
348 return -ENOMEM;
349}
350
351/**
352 * vmemmap_remap_alloc - remap the vmemmap virtual address range [@start, end)
353 * to the page which is from the @vmemmap_pages
354 * respectively.
355 * @start: start address of the vmemmap virtual address range that we want
356 * to remap.
357 * @end: end address of the vmemmap virtual address range that we want to
358 * remap.
359 * @reuse: reuse address.
360 * @gfp_mask: GFP flag for allocating vmemmap pages.
361 *
362 * Return: %0 on success, negative error code otherwise.
363 */
364int vmemmap_remap_alloc(unsigned long start, unsigned long end,
365 unsigned long reuse, gfp_t gfp_mask)
366{
367 LIST_HEAD(vmemmap_pages);
368 struct vmemmap_remap_walk walk = {
369 .remap_pte = vmemmap_restore_pte,
370 .reuse_addr = reuse,
371 .vmemmap_pages = &vmemmap_pages,
372 };
373
374 /* See the comment in the vmemmap_remap_free(). */
375 BUG_ON(start - reuse != PAGE_SIZE);
376
377 if (alloc_vmemmap_page_list(start, end, gfp_mask, &vmemmap_pages))
378 return -ENOMEM;
379
380 mmap_read_lock(&init_mm);
381 vmemmap_remap_range(reuse, end, &walk);
382 mmap_read_unlock(&init_mm);
383
384 return 0;
385}
386
387/*
388 * Allocate a block of memory to be used to back the virtual memory map
389 * or to back the page tables that are used to create the mapping.
390 * Uses the main allocators if they are available, else bootmem.
391 */
392
393static void * __ref __earlyonly_bootmem_alloc(int node,
394 unsigned long size,
395 unsigned long align,
396 unsigned long goal)
397{
398 return memblock_alloc_try_nid_raw(size, align, goal,
399 MEMBLOCK_ALLOC_ACCESSIBLE, node);
400}
401
402void * __meminit vmemmap_alloc_block(unsigned long size, int node)
403{
404 /* If the main allocator is up use that, fallback to bootmem. */
405 if (slab_is_available()) {
406 gfp_t gfp_mask = GFP_KERNEL|__GFP_RETRY_MAYFAIL|__GFP_NOWARN;
407 int order = get_order(size);
408 static bool warned;
409 struct page *page;
410
411 page = alloc_pages_node(node, gfp_mask, order);
412 if (page)
413 return page_address(page);
414
415 if (!warned) {
416 warn_alloc(gfp_mask & ~__GFP_NOWARN, NULL,
417 "vmemmap alloc failure: order:%u", order);
418 warned = true;
419 }
420 return NULL;
421 } else
422 return __earlyonly_bootmem_alloc(node, size, size,
423 __pa(MAX_DMA_ADDRESS));
424}
425
426static void * __meminit altmap_alloc_block_buf(unsigned long size,
427 struct vmem_altmap *altmap);
428
429/* need to make sure size is all the same during early stage */
430void * __meminit vmemmap_alloc_block_buf(unsigned long size, int node,
431 struct vmem_altmap *altmap)
432{
433 void *ptr;
434
435 if (altmap)
436 return altmap_alloc_block_buf(size, altmap);
437
438 ptr = sparse_buffer_alloc(size);
439 if (!ptr)
440 ptr = vmemmap_alloc_block(size, node);
441 return ptr;
442}
443
444static unsigned long __meminit vmem_altmap_next_pfn(struct vmem_altmap *altmap)
445{
446 return altmap->base_pfn + altmap->reserve + altmap->alloc
447 + altmap->align;
448}
449
450static unsigned long __meminit vmem_altmap_nr_free(struct vmem_altmap *altmap)
451{
452 unsigned long allocated = altmap->alloc + altmap->align;
453
454 if (altmap->free > allocated)
455 return altmap->free - allocated;
456 return 0;
457}
458
459static void * __meminit altmap_alloc_block_buf(unsigned long size,
460 struct vmem_altmap *altmap)
461{
462 unsigned long pfn, nr_pfns, nr_align;
463
464 if (size & ~PAGE_MASK) {
465 pr_warn_once("%s: allocations must be multiple of PAGE_SIZE (%ld)\n",
466 __func__, size);
467 return NULL;
468 }
469
470 pfn = vmem_altmap_next_pfn(altmap);
471 nr_pfns = size >> PAGE_SHIFT;
472 nr_align = 1UL << find_first_bit(&nr_pfns, BITS_PER_LONG);
473 nr_align = ALIGN(pfn, nr_align) - pfn;
474 if (nr_pfns + nr_align > vmem_altmap_nr_free(altmap))
475 return NULL;
476
477 altmap->alloc += nr_pfns;
478 altmap->align += nr_align;
479 pfn += nr_align;
480
481 pr_debug("%s: pfn: %#lx alloc: %ld align: %ld nr: %#lx\n",
482 __func__, pfn, altmap->alloc, altmap->align, nr_pfns);
483 return __va(__pfn_to_phys(pfn));
484}
485
486void __meminit vmemmap_verify(pte_t *pte, int node,
487 unsigned long start, unsigned long end)
488{
489 unsigned long pfn = pte_pfn(*pte);
490 int actual_node = early_pfn_to_nid(pfn);
491
492 if (node_distance(actual_node, node) > LOCAL_DISTANCE)
493 pr_warn("[%lx-%lx] potential offnode page_structs\n",
494 start, end - 1);
495}
496
497pte_t * __meminit vmemmap_pte_populate(pmd_t *pmd, unsigned long addr, int node,
498 struct vmem_altmap *altmap)
499{
500 pte_t *pte = pte_offset_kernel(pmd, addr);
501 if (pte_none(*pte)) {
502 pte_t entry;
503 void *p;
504
505 p = vmemmap_alloc_block_buf(PAGE_SIZE, node, altmap);
506 if (!p)
507 return NULL;
508 entry = pfn_pte(__pa(p) >> PAGE_SHIFT, PAGE_KERNEL);
509 set_pte_at(&init_mm, addr, pte, entry);
510 }
511 return pte;
512}
513
514static void * __meminit vmemmap_alloc_block_zero(unsigned long size, int node)
515{
516 void *p = vmemmap_alloc_block(size, node);
517
518 if (!p)
519 return NULL;
520 memset(p, 0, size);
521
522 return p;
523}
524
525pmd_t * __meminit vmemmap_pmd_populate(pud_t *pud, unsigned long addr, int node)
526{
527 pmd_t *pmd = pmd_offset(pud, addr);
528 if (pmd_none(*pmd)) {
529 void *p = vmemmap_alloc_block_zero(PAGE_SIZE, node);
530 if (!p)
531 return NULL;
532 pmd_populate_kernel(&init_mm, pmd, p);
533 }
534 return pmd;
535}
536
537pud_t * __meminit vmemmap_pud_populate(p4d_t *p4d, unsigned long addr, int node)
538{
539 pud_t *pud = pud_offset(p4d, addr);
540 if (pud_none(*pud)) {
541 void *p = vmemmap_alloc_block_zero(PAGE_SIZE, node);
542 if (!p)
543 return NULL;
544 pud_populate(&init_mm, pud, p);
545 }
546 return pud;
547}
548
549p4d_t * __meminit vmemmap_p4d_populate(pgd_t *pgd, unsigned long addr, int node)
550{
551 p4d_t *p4d = p4d_offset(pgd, addr);
552 if (p4d_none(*p4d)) {
553 void *p = vmemmap_alloc_block_zero(PAGE_SIZE, node);
554 if (!p)
555 return NULL;
556 p4d_populate(&init_mm, p4d, p);
557 }
558 return p4d;
559}
560
561pgd_t * __meminit vmemmap_pgd_populate(unsigned long addr, int node)
562{
563 pgd_t *pgd = pgd_offset_k(addr);
564 if (pgd_none(*pgd)) {
565 void *p = vmemmap_alloc_block_zero(PAGE_SIZE, node);
566 if (!p)
567 return NULL;
568 pgd_populate(&init_mm, pgd, p);
569 }
570 return pgd;
571}
572
573int __meminit vmemmap_populate_basepages(unsigned long start, unsigned long end,
574 int node, struct vmem_altmap *altmap)
575{
576 unsigned long addr = start;
577 pgd_t *pgd;
578 p4d_t *p4d;
579 pud_t *pud;
580 pmd_t *pmd;
581 pte_t *pte;
582
583 for (; addr < end; addr += PAGE_SIZE) {
584 pgd = vmemmap_pgd_populate(addr, node);
585 if (!pgd)
586 return -ENOMEM;
587 p4d = vmemmap_p4d_populate(pgd, addr, node);
588 if (!p4d)
589 return -ENOMEM;
590 pud = vmemmap_pud_populate(p4d, addr, node);
591 if (!pud)
592 return -ENOMEM;
593 pmd = vmemmap_pmd_populate(pud, addr, node);
594 if (!pmd)
595 return -ENOMEM;
596 pte = vmemmap_pte_populate(pmd, addr, node, altmap);
597 if (!pte)
598 return -ENOMEM;
599 vmemmap_verify(pte, node, addr, addr + PAGE_SIZE);
600 }
601
602 return 0;
603}
604
605struct page * __meminit __populate_section_memmap(unsigned long pfn,
606 unsigned long nr_pages, int nid, struct vmem_altmap *altmap)
607{
608 unsigned long start = (unsigned long) pfn_to_page(pfn);
609 unsigned long end = start + nr_pages * sizeof(struct page);
610
611 if (WARN_ON_ONCE(!IS_ALIGNED(pfn, PAGES_PER_SUBSECTION) ||
612 !IS_ALIGNED(nr_pages, PAGES_PER_SUBSECTION)))
613 return NULL;
614
615 if (vmemmap_populate(start, end, nid, altmap))
616 return NULL;
617
618 return pfn_to_page(pfn);
619}
1// SPDX-License-Identifier: GPL-2.0
2/*
3 * Virtual Memory Map support
4 *
5 * (C) 2007 sgi. Christoph Lameter.
6 *
7 * Virtual memory maps allow VM primitives pfn_to_page, page_to_pfn,
8 * virt_to_page, page_address() to be implemented as a base offset
9 * calculation without memory access.
10 *
11 * However, virtual mappings need a page table and TLBs. Many Linux
12 * architectures already map their physical space using 1-1 mappings
13 * via TLBs. For those arches the virtual memory map is essentially
14 * for free if we use the same page size as the 1-1 mappings. In that
15 * case the overhead consists of a few additional pages that are
16 * allocated to create a view of memory for vmemmap.
17 *
18 * The architecture is expected to provide a vmemmap_populate() function
19 * to instantiate the mapping.
20 */
21#include <linux/mm.h>
22#include <linux/mmzone.h>
23#include <linux/memblock.h>
24#include <linux/memremap.h>
25#include <linux/highmem.h>
26#include <linux/slab.h>
27#include <linux/spinlock.h>
28#include <linux/vmalloc.h>
29#include <linux/sched.h>
30
31#include <asm/dma.h>
32#include <asm/pgalloc.h>
33
34/*
35 * Allocate a block of memory to be used to back the virtual memory map
36 * or to back the page tables that are used to create the mapping.
37 * Uses the main allocators if they are available, else bootmem.
38 */
39
40static void * __ref __earlyonly_bootmem_alloc(int node,
41 unsigned long size,
42 unsigned long align,
43 unsigned long goal)
44{
45 return memblock_alloc_try_nid_raw(size, align, goal,
46 MEMBLOCK_ALLOC_ACCESSIBLE, node);
47}
48
49void * __meminit vmemmap_alloc_block(unsigned long size, int node)
50{
51 /* If the main allocator is up use that, fallback to bootmem. */
52 if (slab_is_available()) {
53 gfp_t gfp_mask = GFP_KERNEL|__GFP_RETRY_MAYFAIL|__GFP_NOWARN;
54 int order = get_order(size);
55 static bool warned;
56 struct page *page;
57
58 page = alloc_pages_node(node, gfp_mask, order);
59 if (page)
60 return page_address(page);
61
62 if (!warned) {
63 warn_alloc(gfp_mask & ~__GFP_NOWARN, NULL,
64 "vmemmap alloc failure: order:%u", order);
65 warned = true;
66 }
67 return NULL;
68 } else
69 return __earlyonly_bootmem_alloc(node, size, size,
70 __pa(MAX_DMA_ADDRESS));
71}
72
73static void * __meminit altmap_alloc_block_buf(unsigned long size,
74 struct vmem_altmap *altmap);
75
76/* need to make sure size is all the same during early stage */
77void * __meminit vmemmap_alloc_block_buf(unsigned long size, int node,
78 struct vmem_altmap *altmap)
79{
80 void *ptr;
81
82 if (altmap)
83 return altmap_alloc_block_buf(size, altmap);
84
85 ptr = sparse_buffer_alloc(size);
86 if (!ptr)
87 ptr = vmemmap_alloc_block(size, node);
88 return ptr;
89}
90
91static unsigned long __meminit vmem_altmap_next_pfn(struct vmem_altmap *altmap)
92{
93 return altmap->base_pfn + altmap->reserve + altmap->alloc
94 + altmap->align;
95}
96
97static unsigned long __meminit vmem_altmap_nr_free(struct vmem_altmap *altmap)
98{
99 unsigned long allocated = altmap->alloc + altmap->align;
100
101 if (altmap->free > allocated)
102 return altmap->free - allocated;
103 return 0;
104}
105
106static void * __meminit altmap_alloc_block_buf(unsigned long size,
107 struct vmem_altmap *altmap)
108{
109 unsigned long pfn, nr_pfns, nr_align;
110
111 if (size & ~PAGE_MASK) {
112 pr_warn_once("%s: allocations must be multiple of PAGE_SIZE (%ld)\n",
113 __func__, size);
114 return NULL;
115 }
116
117 pfn = vmem_altmap_next_pfn(altmap);
118 nr_pfns = size >> PAGE_SHIFT;
119 nr_align = 1UL << find_first_bit(&nr_pfns, BITS_PER_LONG);
120 nr_align = ALIGN(pfn, nr_align) - pfn;
121 if (nr_pfns + nr_align > vmem_altmap_nr_free(altmap))
122 return NULL;
123
124 altmap->alloc += nr_pfns;
125 altmap->align += nr_align;
126 pfn += nr_align;
127
128 pr_debug("%s: pfn: %#lx alloc: %ld align: %ld nr: %#lx\n",
129 __func__, pfn, altmap->alloc, altmap->align, nr_pfns);
130 return __va(__pfn_to_phys(pfn));
131}
132
133void __meminit vmemmap_verify(pte_t *pte, int node,
134 unsigned long start, unsigned long end)
135{
136 unsigned long pfn = pte_pfn(ptep_get(pte));
137 int actual_node = early_pfn_to_nid(pfn);
138
139 if (node_distance(actual_node, node) > LOCAL_DISTANCE)
140 pr_warn_once("[%lx-%lx] potential offnode page_structs\n",
141 start, end - 1);
142}
143
144pte_t * __meminit vmemmap_pte_populate(pmd_t *pmd, unsigned long addr, int node,
145 struct vmem_altmap *altmap,
146 struct page *reuse)
147{
148 pte_t *pte = pte_offset_kernel(pmd, addr);
149 if (pte_none(ptep_get(pte))) {
150 pte_t entry;
151 void *p;
152
153 if (!reuse) {
154 p = vmemmap_alloc_block_buf(PAGE_SIZE, node, altmap);
155 if (!p)
156 return NULL;
157 } else {
158 /*
159 * When a PTE/PMD entry is freed from the init_mm
160 * there's a free_pages() call to this page allocated
161 * above. Thus this get_page() is paired with the
162 * put_page_testzero() on the freeing path.
163 * This can only called by certain ZONE_DEVICE path,
164 * and through vmemmap_populate_compound_pages() when
165 * slab is available.
166 */
167 get_page(reuse);
168 p = page_to_virt(reuse);
169 }
170 entry = pfn_pte(__pa(p) >> PAGE_SHIFT, PAGE_KERNEL);
171 set_pte_at(&init_mm, addr, pte, entry);
172 }
173 return pte;
174}
175
176static void * __meminit vmemmap_alloc_block_zero(unsigned long size, int node)
177{
178 void *p = vmemmap_alloc_block(size, node);
179
180 if (!p)
181 return NULL;
182 memset(p, 0, size);
183
184 return p;
185}
186
187pmd_t * __meminit vmemmap_pmd_populate(pud_t *pud, unsigned long addr, int node)
188{
189 pmd_t *pmd = pmd_offset(pud, addr);
190 if (pmd_none(*pmd)) {
191 void *p = vmemmap_alloc_block_zero(PAGE_SIZE, node);
192 if (!p)
193 return NULL;
194 kernel_pte_init(p);
195 pmd_populate_kernel(&init_mm, pmd, p);
196 }
197 return pmd;
198}
199
200pud_t * __meminit vmemmap_pud_populate(p4d_t *p4d, unsigned long addr, int node)
201{
202 pud_t *pud = pud_offset(p4d, addr);
203 if (pud_none(*pud)) {
204 void *p = vmemmap_alloc_block_zero(PAGE_SIZE, node);
205 if (!p)
206 return NULL;
207 pmd_init(p);
208 pud_populate(&init_mm, pud, p);
209 }
210 return pud;
211}
212
213p4d_t * __meminit vmemmap_p4d_populate(pgd_t *pgd, unsigned long addr, int node)
214{
215 p4d_t *p4d = p4d_offset(pgd, addr);
216 if (p4d_none(*p4d)) {
217 void *p = vmemmap_alloc_block_zero(PAGE_SIZE, node);
218 if (!p)
219 return NULL;
220 pud_init(p);
221 p4d_populate(&init_mm, p4d, p);
222 }
223 return p4d;
224}
225
226pgd_t * __meminit vmemmap_pgd_populate(unsigned long addr, int node)
227{
228 pgd_t *pgd = pgd_offset_k(addr);
229 if (pgd_none(*pgd)) {
230 void *p = vmemmap_alloc_block_zero(PAGE_SIZE, node);
231 if (!p)
232 return NULL;
233 pgd_populate(&init_mm, pgd, p);
234 }
235 return pgd;
236}
237
238static pte_t * __meminit vmemmap_populate_address(unsigned long addr, int node,
239 struct vmem_altmap *altmap,
240 struct page *reuse)
241{
242 pgd_t *pgd;
243 p4d_t *p4d;
244 pud_t *pud;
245 pmd_t *pmd;
246 pte_t *pte;
247
248 pgd = vmemmap_pgd_populate(addr, node);
249 if (!pgd)
250 return NULL;
251 p4d = vmemmap_p4d_populate(pgd, addr, node);
252 if (!p4d)
253 return NULL;
254 pud = vmemmap_pud_populate(p4d, addr, node);
255 if (!pud)
256 return NULL;
257 pmd = vmemmap_pmd_populate(pud, addr, node);
258 if (!pmd)
259 return NULL;
260 pte = vmemmap_pte_populate(pmd, addr, node, altmap, reuse);
261 if (!pte)
262 return NULL;
263 vmemmap_verify(pte, node, addr, addr + PAGE_SIZE);
264
265 return pte;
266}
267
268static int __meminit vmemmap_populate_range(unsigned long start,
269 unsigned long end, int node,
270 struct vmem_altmap *altmap,
271 struct page *reuse)
272{
273 unsigned long addr = start;
274 pte_t *pte;
275
276 for (; addr < end; addr += PAGE_SIZE) {
277 pte = vmemmap_populate_address(addr, node, altmap, reuse);
278 if (!pte)
279 return -ENOMEM;
280 }
281
282 return 0;
283}
284
285int __meminit vmemmap_populate_basepages(unsigned long start, unsigned long end,
286 int node, struct vmem_altmap *altmap)
287{
288 return vmemmap_populate_range(start, end, node, altmap, NULL);
289}
290
291void __weak __meminit vmemmap_set_pmd(pmd_t *pmd, void *p, int node,
292 unsigned long addr, unsigned long next)
293{
294}
295
296int __weak __meminit vmemmap_check_pmd(pmd_t *pmd, int node,
297 unsigned long addr, unsigned long next)
298{
299 return 0;
300}
301
302int __meminit vmemmap_populate_hugepages(unsigned long start, unsigned long end,
303 int node, struct vmem_altmap *altmap)
304{
305 unsigned long addr;
306 unsigned long next;
307 pgd_t *pgd;
308 p4d_t *p4d;
309 pud_t *pud;
310 pmd_t *pmd;
311
312 for (addr = start; addr < end; addr = next) {
313 next = pmd_addr_end(addr, end);
314
315 pgd = vmemmap_pgd_populate(addr, node);
316 if (!pgd)
317 return -ENOMEM;
318
319 p4d = vmemmap_p4d_populate(pgd, addr, node);
320 if (!p4d)
321 return -ENOMEM;
322
323 pud = vmemmap_pud_populate(p4d, addr, node);
324 if (!pud)
325 return -ENOMEM;
326
327 pmd = pmd_offset(pud, addr);
328 if (pmd_none(READ_ONCE(*pmd))) {
329 void *p;
330
331 p = vmemmap_alloc_block_buf(PMD_SIZE, node, altmap);
332 if (p) {
333 vmemmap_set_pmd(pmd, p, node, addr, next);
334 continue;
335 } else if (altmap) {
336 /*
337 * No fallback: In any case we care about, the
338 * altmap should be reasonably sized and aligned
339 * such that vmemmap_alloc_block_buf() will always
340 * succeed. For consistency with the PTE case,
341 * return an error here as failure could indicate
342 * a configuration issue with the size of the altmap.
343 */
344 return -ENOMEM;
345 }
346 } else if (vmemmap_check_pmd(pmd, node, addr, next))
347 continue;
348 if (vmemmap_populate_basepages(addr, next, node, altmap))
349 return -ENOMEM;
350 }
351 return 0;
352}
353
354#ifndef vmemmap_populate_compound_pages
355/*
356 * For compound pages bigger than section size (e.g. x86 1G compound
357 * pages with 2M subsection size) fill the rest of sections as tail
358 * pages.
359 *
360 * Note that memremap_pages() resets @nr_range value and will increment
361 * it after each range successful onlining. Thus the value or @nr_range
362 * at section memmap populate corresponds to the in-progress range
363 * being onlined here.
364 */
365static bool __meminit reuse_compound_section(unsigned long start_pfn,
366 struct dev_pagemap *pgmap)
367{
368 unsigned long nr_pages = pgmap_vmemmap_nr(pgmap);
369 unsigned long offset = start_pfn -
370 PHYS_PFN(pgmap->ranges[pgmap->nr_range].start);
371
372 return !IS_ALIGNED(offset, nr_pages) && nr_pages > PAGES_PER_SUBSECTION;
373}
374
375static pte_t * __meminit compound_section_tail_page(unsigned long addr)
376{
377 pte_t *pte;
378
379 addr -= PAGE_SIZE;
380
381 /*
382 * Assuming sections are populated sequentially, the previous section's
383 * page data can be reused.
384 */
385 pte = pte_offset_kernel(pmd_off_k(addr), addr);
386 if (!pte)
387 return NULL;
388
389 return pte;
390}
391
392static int __meminit vmemmap_populate_compound_pages(unsigned long start_pfn,
393 unsigned long start,
394 unsigned long end, int node,
395 struct dev_pagemap *pgmap)
396{
397 unsigned long size, addr;
398 pte_t *pte;
399 int rc;
400
401 if (reuse_compound_section(start_pfn, pgmap)) {
402 pte = compound_section_tail_page(start);
403 if (!pte)
404 return -ENOMEM;
405
406 /*
407 * Reuse the page that was populated in the prior iteration
408 * with just tail struct pages.
409 */
410 return vmemmap_populate_range(start, end, node, NULL,
411 pte_page(ptep_get(pte)));
412 }
413
414 size = min(end - start, pgmap_vmemmap_nr(pgmap) * sizeof(struct page));
415 for (addr = start; addr < end; addr += size) {
416 unsigned long next, last = addr + size;
417
418 /* Populate the head page vmemmap page */
419 pte = vmemmap_populate_address(addr, node, NULL, NULL);
420 if (!pte)
421 return -ENOMEM;
422
423 /* Populate the tail pages vmemmap page */
424 next = addr + PAGE_SIZE;
425 pte = vmemmap_populate_address(next, node, NULL, NULL);
426 if (!pte)
427 return -ENOMEM;
428
429 /*
430 * Reuse the previous page for the rest of tail pages
431 * See layout diagram in Documentation/mm/vmemmap_dedup.rst
432 */
433 next += PAGE_SIZE;
434 rc = vmemmap_populate_range(next, last, node, NULL,
435 pte_page(ptep_get(pte)));
436 if (rc)
437 return -ENOMEM;
438 }
439
440 return 0;
441}
442
443#endif
444
445struct page * __meminit __populate_section_memmap(unsigned long pfn,
446 unsigned long nr_pages, int nid, struct vmem_altmap *altmap,
447 struct dev_pagemap *pgmap)
448{
449 unsigned long start = (unsigned long) pfn_to_page(pfn);
450 unsigned long end = start + nr_pages * sizeof(struct page);
451 int r;
452
453 if (WARN_ON_ONCE(!IS_ALIGNED(pfn, PAGES_PER_SUBSECTION) ||
454 !IS_ALIGNED(nr_pages, PAGES_PER_SUBSECTION)))
455 return NULL;
456
457 if (vmemmap_can_optimize(altmap, pgmap))
458 r = vmemmap_populate_compound_pages(pfn, start, end, nid, pgmap);
459 else
460 r = vmemmap_populate(start, end, nid, altmap);
461
462 if (r < 0)
463 return NULL;
464
465 if (system_state == SYSTEM_BOOTING)
466 memmap_boot_pages_add(DIV_ROUND_UP(end - start, PAGE_SIZE));
467 else
468 memmap_pages_add(DIV_ROUND_UP(end - start, PAGE_SIZE));
469
470 return pfn_to_page(pfn);
471}