sparse-vmemmap.c

  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/pgalloc.h>
 31
 32#include <asm/dma.h>
 33#include <asm/tlbflush.h>
 34
 35#include "hugetlb_vmemmap.h"
 36
 37/*
 38 * Flags for vmemmap_populate_range and friends.
 39 */
 40/* Get a ref on the head page struct page, for ZONE_DEVICE compound pages */
 41#define VMEMMAP_POPULATE_PAGEREF	0x0001
 42
 43#include "internal.h"
 44
 45/*
 46 * Allocate a block of memory to be used to back the virtual memory map
 47 * or to back the page tables that are used to create the mapping.
 48 * Uses the main allocators if they are available, else bootmem.
 49 */
 50
 51static void * __ref __earlyonly_bootmem_alloc(int node,
 52				unsigned long size,
 53				unsigned long align,
 54				unsigned long goal)
 55{
 56	return memmap_alloc(size, align, goal, node, false);
 57}
 58
 59void * __meminit vmemmap_alloc_block(unsigned long size, int node)
 60{
 61	/* If the main allocator is up use that, fallback to bootmem. */
 62	if (slab_is_available()) {
 63		gfp_t gfp_mask = GFP_KERNEL|__GFP_RETRY_MAYFAIL|__GFP_NOWARN;
 64		int order = get_order(size);
 65		static bool warned;
 66		struct page *page;
 67
 68		page = alloc_pages_node(node, gfp_mask, order);
 69		if (page)
 70			return page_address(page);
 71
 72		if (!warned) {
 73			warn_alloc(gfp_mask & ~__GFP_NOWARN, NULL,
 74				   "vmemmap alloc failure: order:%u", order);
 75			warned = true;
 76		}
 77		return NULL;
 78	} else
 79		return __earlyonly_bootmem_alloc(node, size, size,
 80				__pa(MAX_DMA_ADDRESS));
 81}
 82
 83static void * __meminit altmap_alloc_block_buf(unsigned long size,
 84					       struct vmem_altmap *altmap);
 85
 86/* need to make sure size is all the same during early stage */
 87void * __meminit vmemmap_alloc_block_buf(unsigned long size, int node,
 88					 struct vmem_altmap *altmap)
 89{
 90	void *ptr;
 91
 92	if (altmap)
 93		return altmap_alloc_block_buf(size, altmap);
 94
 95	ptr = sparse_buffer_alloc(size);
 96	if (!ptr)
 97		ptr = vmemmap_alloc_block(size, node);
 98	return ptr;
 99}
100
101static unsigned long __meminit vmem_altmap_next_pfn(struct vmem_altmap *altmap)
102{
103	return altmap->base_pfn + altmap->reserve + altmap->alloc
104		+ altmap->align;
105}
106
107static unsigned long __meminit vmem_altmap_nr_free(struct vmem_altmap *altmap)
108{
109	unsigned long allocated = altmap->alloc + altmap->align;
110
111	if (altmap->free > allocated)
112		return altmap->free - allocated;
113	return 0;
114}
115
116static void * __meminit altmap_alloc_block_buf(unsigned long size,
117					       struct vmem_altmap *altmap)
118{
119	unsigned long pfn, nr_pfns, nr_align;
120
121	if (size & ~PAGE_MASK) {
122		pr_warn_once("%s: allocations must be multiple of PAGE_SIZE (%ld)\n",
123				__func__, size);
124		return NULL;
125	}
126
127	pfn = vmem_altmap_next_pfn(altmap);
128	nr_pfns = size >> PAGE_SHIFT;
129	nr_align = 1UL << find_first_bit(&nr_pfns, BITS_PER_LONG);
130	nr_align = ALIGN(pfn, nr_align) - pfn;
131	if (nr_pfns + nr_align > vmem_altmap_nr_free(altmap))
132		return NULL;
133
134	altmap->alloc += nr_pfns;
135	altmap->align += nr_align;
136	pfn += nr_align;
137
138	pr_debug("%s: pfn: %#lx alloc: %ld align: %ld nr: %#lx\n",
139			__func__, pfn, altmap->alloc, altmap->align, nr_pfns);
140	return __va(__pfn_to_phys(pfn));
141}
142
143void __meminit vmemmap_verify(pte_t *pte, int node,
144				unsigned long start, unsigned long end)
145{
146	unsigned long pfn = pte_pfn(ptep_get(pte));
147	int actual_node = early_pfn_to_nid(pfn);
148
149	if (node_distance(actual_node, node) > LOCAL_DISTANCE)
150		pr_warn_once("[%lx-%lx] potential offnode page_structs\n",
151			start, end - 1);
152}
153
154pte_t * __meminit vmemmap_pte_populate(pmd_t *pmd, unsigned long addr, int node,
155				       struct vmem_altmap *altmap,
156				       unsigned long ptpfn, unsigned long flags)
157{
158	pte_t *pte = pte_offset_kernel(pmd, addr);
159	if (pte_none(ptep_get(pte))) {
160		pte_t entry;
161		void *p;
162
163		if (ptpfn == (unsigned long)-1) {
164			p = vmemmap_alloc_block_buf(PAGE_SIZE, node, altmap);
165			if (!p)
166				return NULL;
167			ptpfn = PHYS_PFN(__pa(p));
168		} else {
169			/*
170			 * When a PTE/PMD entry is freed from the init_mm
171			 * there's a free_pages() call to this page allocated
172			 * above. Thus this get_page() is paired with the
173			 * put_page_testzero() on the freeing path.
174			 * This can only called by certain ZONE_DEVICE path,
175			 * and through vmemmap_populate_compound_pages() when
176			 * slab is available.
177			 */
178			if (flags & VMEMMAP_POPULATE_PAGEREF)
179				get_page(pfn_to_page(ptpfn));
180		}
181		entry = pfn_pte(ptpfn, PAGE_KERNEL);
182		set_pte_at(&init_mm, addr, pte, entry);
183	}
184	return pte;
185}
186
187static void * __meminit vmemmap_alloc_block_zero(unsigned long size, int node)
188{
189	void *p = vmemmap_alloc_block(size, node);
190
191	if (!p)
192		return NULL;
193	memset(p, 0, size);
194
195	return p;
196}
197
198pmd_t * __meminit vmemmap_pmd_populate(pud_t *pud, unsigned long addr, int node)
199{
200	pmd_t *pmd = pmd_offset(pud, addr);
201	if (pmd_none(*pmd)) {
202		void *p = vmemmap_alloc_block_zero(PAGE_SIZE, node);
203		if (!p)
204			return NULL;
205		kernel_pte_init(p);
206		pmd_populate_kernel(&init_mm, pmd, p);
207	}
208	return pmd;
209}
210
211pud_t * __meminit vmemmap_pud_populate(p4d_t *p4d, unsigned long addr, int node)
212{
213	pud_t *pud = pud_offset(p4d, addr);
214	if (pud_none(*pud)) {
215		void *p = vmemmap_alloc_block_zero(PAGE_SIZE, node);
216		if (!p)
217			return NULL;
218		pmd_init(p);
219		pud_populate(&init_mm, pud, p);
220	}
221	return pud;
222}
223
224p4d_t * __meminit vmemmap_p4d_populate(pgd_t *pgd, unsigned long addr, int node)
225{
226	p4d_t *p4d = p4d_offset(pgd, addr);
227	if (p4d_none(*p4d)) {
228		void *p = vmemmap_alloc_block_zero(PAGE_SIZE, node);
229		if (!p)
230			return NULL;
231		pud_init(p);
232		p4d_populate_kernel(addr, p4d, p);
233	}
234	return p4d;
235}
236
237pgd_t * __meminit vmemmap_pgd_populate(unsigned long addr, int node)
238{
239	pgd_t *pgd = pgd_offset_k(addr);
240	if (pgd_none(*pgd)) {
241		void *p = vmemmap_alloc_block_zero(PAGE_SIZE, node);
242		if (!p)
243			return NULL;
244		pgd_populate_kernel(addr, pgd, p);
245	}
246	return pgd;
247}
248
249static pte_t * __meminit vmemmap_populate_address(unsigned long addr, int node,
250					      struct vmem_altmap *altmap,
251					      unsigned long ptpfn,
252					      unsigned long flags)
253{
254	pgd_t *pgd;
255	p4d_t *p4d;
256	pud_t *pud;
257	pmd_t *pmd;
258	pte_t *pte;
259
260	pgd = vmemmap_pgd_populate(addr, node);
261	if (!pgd)
262		return NULL;
263	p4d = vmemmap_p4d_populate(pgd, addr, node);
264	if (!p4d)
265		return NULL;
266	pud = vmemmap_pud_populate(p4d, addr, node);
267	if (!pud)
268		return NULL;
269	pmd = vmemmap_pmd_populate(pud, addr, node);
270	if (!pmd)
271		return NULL;
272	pte = vmemmap_pte_populate(pmd, addr, node, altmap, ptpfn, flags);
273	if (!pte)
274		return NULL;
275	vmemmap_verify(pte, node, addr, addr + PAGE_SIZE);
276
277	return pte;
278}
279
280static int __meminit vmemmap_populate_range(unsigned long start,
281					    unsigned long end, int node,
282					    struct vmem_altmap *altmap,
283					    unsigned long ptpfn,
284					    unsigned long flags)
285{
286	unsigned long addr = start;
287	pte_t *pte;
288
289	for (; addr < end; addr += PAGE_SIZE) {
290		pte = vmemmap_populate_address(addr, node, altmap,
291					       ptpfn, flags);
292		if (!pte)
293			return -ENOMEM;
294	}
295
296	return 0;
297}
298
299int __meminit vmemmap_populate_basepages(unsigned long start, unsigned long end,
300					 int node, struct vmem_altmap *altmap)
301{
302	return vmemmap_populate_range(start, end, node, altmap, -1, 0);
303}
304
305/*
306 * Undo populate_hvo, and replace it with a normal base page mapping.
307 * Used in memory init in case a HVO mapping needs to be undone.
308 *
309 * This can happen when it is discovered that a memblock allocated
310 * hugetlb page spans multiple zones, which can only be verified
311 * after zones have been initialized.
312 *
313 * We know that:
314 * 1) The first @headsize / PAGE_SIZE vmemmap pages were individually
315 *    allocated through memblock, and mapped.
316 *
317 * 2) The rest of the vmemmap pages are mirrors of the last head page.
318 */
319int __meminit vmemmap_undo_hvo(unsigned long addr, unsigned long end,
320				      int node, unsigned long headsize)
321{
322	unsigned long maddr, pfn;
323	pte_t *pte;
324	int headpages;
325
326	/*
327	 * Should only be called early in boot, so nothing will
328	 * be accessing these page structures.
329	 */
330	WARN_ON(!early_boot_irqs_disabled);
331
332	headpages = headsize >> PAGE_SHIFT;
333
334	/*
335	 * Clear mirrored mappings for tail page structs.
336	 */
337	for (maddr = addr + headsize; maddr < end; maddr += PAGE_SIZE) {
338		pte = virt_to_kpte(maddr);
339		pte_clear(&init_mm, maddr, pte);
340	}
341
342	/*
343	 * Clear and free mappings for head page and first tail page
344	 * structs.
345	 */
346	for (maddr = addr; headpages-- > 0; maddr += PAGE_SIZE) {
347		pte = virt_to_kpte(maddr);
348		pfn = pte_pfn(ptep_get(pte));
349		pte_clear(&init_mm, maddr, pte);
350		memblock_phys_free(PFN_PHYS(pfn), PAGE_SIZE);
351	}
352
353	flush_tlb_kernel_range(addr, end);
354
355	return vmemmap_populate(addr, end, node, NULL);
356}
357
358/*
359 * Write protect the mirrored tail page structs for HVO. This will be
360 * called from the hugetlb code when gathering and initializing the
361 * memblock allocated gigantic pages. The write protect can't be
362 * done earlier, since it can't be guaranteed that the reserved
363 * page structures will not be written to during initialization,
364 * even if CONFIG_DEFERRED_STRUCT_PAGE_INIT is enabled.
365 *
366 * The PTEs are known to exist, and nothing else should be touching
367 * these pages. The caller is responsible for any TLB flushing.
368 */
369void vmemmap_wrprotect_hvo(unsigned long addr, unsigned long end,
370				    int node, unsigned long headsize)
371{
372	unsigned long maddr;
373	pte_t *pte;
374
375	for (maddr = addr + headsize; maddr < end; maddr += PAGE_SIZE) {
376		pte = virt_to_kpte(maddr);
377		ptep_set_wrprotect(&init_mm, maddr, pte);
378	}
379}
380
381/*
382 * Populate vmemmap pages HVO-style. The first page contains the head
383 * page and needed tail pages, the other ones are mirrors of the first
384 * page.
385 */
386int __meminit vmemmap_populate_hvo(unsigned long addr, unsigned long end,
387				       int node, unsigned long headsize)
388{
389	pte_t *pte;
390	unsigned long maddr;
391
392	for (maddr = addr; maddr < addr + headsize; maddr += PAGE_SIZE) {
393		pte = vmemmap_populate_address(maddr, node, NULL, -1, 0);
394		if (!pte)
395			return -ENOMEM;
396	}
397
398	/*
399	 * Reuse the last page struct page mapped above for the rest.
400	 */
401	return vmemmap_populate_range(maddr, end, node, NULL,
402					pte_pfn(ptep_get(pte)), 0);
403}
404
405void __weak __meminit vmemmap_set_pmd(pmd_t *pmd, void *p, int node,
406				      unsigned long addr, unsigned long next)
407{
408}
409
410int __weak __meminit vmemmap_check_pmd(pmd_t *pmd, int node,
411				       unsigned long addr, unsigned long next)
412{
413	return 0;
414}
415
416int __meminit vmemmap_populate_hugepages(unsigned long start, unsigned long end,
417					 int node, struct vmem_altmap *altmap)
418{
419	unsigned long addr;
420	unsigned long next;
421	pgd_t *pgd;
422	p4d_t *p4d;
423	pud_t *pud;
424	pmd_t *pmd;
425
426	for (addr = start; addr < end; addr = next) {
427		next = pmd_addr_end(addr, end);
428
429		pgd = vmemmap_pgd_populate(addr, node);
430		if (!pgd)
431			return -ENOMEM;
432
433		p4d = vmemmap_p4d_populate(pgd, addr, node);
434		if (!p4d)
435			return -ENOMEM;
436
437		pud = vmemmap_pud_populate(p4d, addr, node);
438		if (!pud)
439			return -ENOMEM;
440
441		pmd = pmd_offset(pud, addr);
442		if (pmd_none(pmdp_get(pmd))) {
443			void *p;
444
445			p = vmemmap_alloc_block_buf(PMD_SIZE, node, altmap);
446			if (p) {
447				vmemmap_set_pmd(pmd, p, node, addr, next);
448				continue;
449			} else if (altmap) {
450				/*
451				 * No fallback: In any case we care about, the
452				 * altmap should be reasonably sized and aligned
453				 * such that vmemmap_alloc_block_buf() will always
454				 * succeed. For consistency with the PTE case,
455				 * return an error here as failure could indicate
456				 * a configuration issue with the size of the altmap.
457				 */
458				return -ENOMEM;
459			}
460		} else if (vmemmap_check_pmd(pmd, node, addr, next))
461			continue;
462		if (vmemmap_populate_basepages(addr, next, node, altmap))
463			return -ENOMEM;
464	}
465	return 0;
466}
467
468#ifndef vmemmap_populate_compound_pages
469/*
470 * For compound pages bigger than section size (e.g. x86 1G compound
471 * pages with 2M subsection size) fill the rest of sections as tail
472 * pages.
473 *
474 * Note that memremap_pages() resets @nr_range value and will increment
475 * it after each range successful onlining. Thus the value or @nr_range
476 * at section memmap populate corresponds to the in-progress range
477 * being onlined here.
478 */
479static bool __meminit reuse_compound_section(unsigned long start_pfn,
480					     struct dev_pagemap *pgmap)
481{
482	unsigned long nr_pages = pgmap_vmemmap_nr(pgmap);
483	unsigned long offset = start_pfn -
484		PHYS_PFN(pgmap->ranges[pgmap->nr_range].start);
485
486	return !IS_ALIGNED(offset, nr_pages) && nr_pages > PAGES_PER_SUBSECTION;
487}
488
489static pte_t * __meminit compound_section_tail_page(unsigned long addr)
490{
491	pte_t *pte;
492
493	addr -= PAGE_SIZE;
494
495	/*
496	 * Assuming sections are populated sequentially, the previous section's
497	 * page data can be reused.
498	 */
499	pte = pte_offset_kernel(pmd_off_k(addr), addr);
500	if (!pte)
501		return NULL;
502
503	return pte;
504}
505
506static int __meminit vmemmap_populate_compound_pages(unsigned long start_pfn,
507						     unsigned long start,
508						     unsigned long end, int node,
509						     struct dev_pagemap *pgmap)
510{
511	unsigned long size, addr;
512	pte_t *pte;
513	int rc;
514
515	if (reuse_compound_section(start_pfn, pgmap)) {
516		pte = compound_section_tail_page(start);
517		if (!pte)
518			return -ENOMEM;
519
520		/*
521		 * Reuse the page that was populated in the prior iteration
522		 * with just tail struct pages.
523		 */
524		return vmemmap_populate_range(start, end, node, NULL,
525					      pte_pfn(ptep_get(pte)),
526					      VMEMMAP_POPULATE_PAGEREF);
527	}
528
529	size = min(end - start, pgmap_vmemmap_nr(pgmap) * sizeof(struct page));
530	for (addr = start; addr < end; addr += size) {
531		unsigned long next, last = addr + size;
532
533		/* Populate the head page vmemmap page */
534		pte = vmemmap_populate_address(addr, node, NULL, -1, 0);
535		if (!pte)
536			return -ENOMEM;
537
538		/* Populate the tail pages vmemmap page */
539		next = addr + PAGE_SIZE;
540		pte = vmemmap_populate_address(next, node, NULL, -1, 0);
541		if (!pte)
542			return -ENOMEM;
543
544		/*
545		 * Reuse the previous page for the rest of tail pages
546		 * See layout diagram in Documentation/mm/vmemmap_dedup.rst
547		 */
548		next += PAGE_SIZE;
549		rc = vmemmap_populate_range(next, last, node, NULL,
550					    pte_pfn(ptep_get(pte)),
551					    VMEMMAP_POPULATE_PAGEREF);
552		if (rc)
553			return -ENOMEM;
554	}
555
556	return 0;
557}
558
559#endif
560
561struct page * __meminit __populate_section_memmap(unsigned long pfn,
562		unsigned long nr_pages, int nid, struct vmem_altmap *altmap,
563		struct dev_pagemap *pgmap)
564{
565	unsigned long start = (unsigned long) pfn_to_page(pfn);
566	unsigned long end = start + nr_pages * sizeof(struct page);
567	int r;
568
569	if (WARN_ON_ONCE(!IS_ALIGNED(pfn, PAGES_PER_SUBSECTION) ||
570		!IS_ALIGNED(nr_pages, PAGES_PER_SUBSECTION)))
571		return NULL;
572
573	if (vmemmap_can_optimize(altmap, pgmap))
574		r = vmemmap_populate_compound_pages(pfn, start, end, nid, pgmap);
575	else
576		r = vmemmap_populate(start, end, nid, altmap);
577
578	if (r < 0)
579		return NULL;
580
581	return pfn_to_page(pfn);
582}
583
584#ifdef CONFIG_SPARSEMEM_VMEMMAP_PREINIT
585/*
586 * This is called just before initializing sections for a NUMA node.
587 * Any special initialization that needs to be done before the
588 * generic initialization can be done from here. Sections that
589 * are initialized in hooks called from here will be skipped by
590 * the generic initialization.
591 */
592void __init sparse_vmemmap_init_nid_early(int nid)
593{
594	hugetlb_vmemmap_init_early(nid);
595}
596
597/*
598 * This is called just before the initialization of page structures
599 * through memmap_init. Zones are now initialized, so any work that
600 * needs to be done that needs zone information can be done from
601 * here.
602 */
603void __init sparse_vmemmap_init_nid_late(int nid)
604{
605	hugetlb_vmemmap_init_late(nid);
606}
607#endif