hugetlb_vmemmap.c

  1// SPDX-License-Identifier: GPL-2.0
  2/*
  3 * HugeTLB Vmemmap Optimization (HVO)
  4 *
  5 * Copyright (c) 2020, ByteDance. All rights reserved.
  6 *
  7 *     Author: Muchun Song <songmuchun@bytedance.com>
  8 *
  9 * See Documentation/mm/vmemmap_dedup.rst
 10 */
 11#define pr_fmt(fmt)	"HugeTLB: " fmt
 12
 13#include <linux/pgtable.h>
 14#include <linux/moduleparam.h>
 15#include <linux/bootmem_info.h>
 16#include <linux/mmdebug.h>
 17#include <linux/pagewalk.h>
 18#include <linux/pgalloc.h>
 19
 20#include <asm/tlbflush.h>
 21#include "hugetlb_vmemmap.h"
 22
 23/**
 24 * struct vmemmap_remap_walk - walk vmemmap page table
 25 *
 26 * @remap_pte:		called for each lowest-level entry (PTE).
 27 * @nr_walked:		the number of walked pte.
 28 * @reuse_page:		the page which is reused for the tail vmemmap pages.
 29 * @reuse_addr:		the virtual address of the @reuse_page page.
 30 * @vmemmap_pages:	the list head of the vmemmap pages that can be freed
 31 *			or is mapped from.
 32 * @flags:		used to modify behavior in vmemmap page table walking
 33 *			operations.
 34 */
 35struct vmemmap_remap_walk {
 36	void			(*remap_pte)(pte_t *pte, unsigned long addr,
 37					     struct vmemmap_remap_walk *walk);
 38	unsigned long		nr_walked;
 39	struct page		*reuse_page;
 40	unsigned long		reuse_addr;
 41	struct list_head	*vmemmap_pages;
 42
 43/* Skip the TLB flush when we split the PMD */
 44#define VMEMMAP_SPLIT_NO_TLB_FLUSH	BIT(0)
 45/* Skip the TLB flush when we remap the PTE */
 46#define VMEMMAP_REMAP_NO_TLB_FLUSH	BIT(1)
 47/* synchronize_rcu() to avoid writes from page_ref_add_unless() */
 48#define VMEMMAP_SYNCHRONIZE_RCU		BIT(2)
 49	unsigned long		flags;
 50};
 51
 52static int vmemmap_split_pmd(pmd_t *pmd, struct page *head, unsigned long start,
 53			     struct vmemmap_remap_walk *walk)
 54{
 55	pmd_t __pmd;
 56	int i;
 57	unsigned long addr = start;
 58	pte_t *pgtable;
 59
 60	pgtable = pte_alloc_one_kernel(&init_mm);
 61	if (!pgtable)
 62		return -ENOMEM;
 63
 64	pmd_populate_kernel(&init_mm, &__pmd, pgtable);
 65
 66	for (i = 0; i < PTRS_PER_PTE; i++, addr += PAGE_SIZE) {
 67		pte_t entry, *pte;
 68		pgprot_t pgprot = PAGE_KERNEL;
 69
 70		entry = mk_pte(head + i, pgprot);
 71		pte = pte_offset_kernel(&__pmd, addr);
 72		set_pte_at(&init_mm, addr, pte, entry);
 73	}
 74
 75	spin_lock(&init_mm.page_table_lock);
 76	if (likely(pmd_leaf(*pmd))) {
 77		/*
 78		 * Higher order allocations from buddy allocator must be able to
 79		 * be treated as independent small pages (as they can be freed
 80		 * individually).
 81		 */
 82		if (!PageReserved(head))
 83			split_page(head, get_order(PMD_SIZE));
 84
 85		/* Make pte visible before pmd. See comment in pmd_install(). */
 86		smp_wmb();
 87		pmd_populate_kernel(&init_mm, pmd, pgtable);
 88		if (!(walk->flags & VMEMMAP_SPLIT_NO_TLB_FLUSH))
 89			flush_tlb_kernel_range(start, start + PMD_SIZE);
 90	} else {
 91		pte_free_kernel(&init_mm, pgtable);
 92	}
 93	spin_unlock(&init_mm.page_table_lock);
 94
 95	return 0;
 96}
 97
 98static int vmemmap_pmd_entry(pmd_t *pmd, unsigned long addr,
 99			     unsigned long next, struct mm_walk *walk)
100{
101	int ret = 0;
102	struct page *head;
103	struct vmemmap_remap_walk *vmemmap_walk = walk->private;
104
105	/* Only splitting, not remapping the vmemmap pages. */
106	if (!vmemmap_walk->remap_pte)
107		walk->action = ACTION_CONTINUE;
108
109	spin_lock(&init_mm.page_table_lock);
110	head = pmd_leaf(*pmd) ? pmd_page(*pmd) : NULL;
111	/*
112	 * Due to HugeTLB alignment requirements and the vmemmap
113	 * pages being at the start of the hotplugged memory
114	 * region in memory_hotplug.memmap_on_memory case. Checking
115	 * the vmemmap page associated with the first vmemmap page
116	 * if it is self-hosted is sufficient.
117	 *
118	 * [                  hotplugged memory                  ]
119	 * [        section        ][...][        section        ]
120	 * [ vmemmap ][              usable memory               ]
121	 *   ^  | ^                        |
122	 *   +--+ |                        |
123	 *        +------------------------+
124	 */
125	if (IS_ENABLED(CONFIG_MEMORY_HOTPLUG) && unlikely(!vmemmap_walk->nr_walked)) {
126		struct page *page = head ? head + pte_index(addr) :
127				    pte_page(ptep_get(pte_offset_kernel(pmd, addr)));
128
129		if (PageVmemmapSelfHosted(page))
130			ret = -ENOTSUPP;
131	}
132	spin_unlock(&init_mm.page_table_lock);
133	if (!head || ret)
134		return ret;
135
136	return vmemmap_split_pmd(pmd, head, addr & PMD_MASK, vmemmap_walk);
137}
138
139static int vmemmap_pte_entry(pte_t *pte, unsigned long addr,
140			     unsigned long next, struct mm_walk *walk)
141{
142	struct vmemmap_remap_walk *vmemmap_walk = walk->private;
143
144	/*
145	 * The reuse_page is found 'first' in page table walking before
146	 * starting remapping.
147	 */
148	if (!vmemmap_walk->reuse_page)
149		vmemmap_walk->reuse_page = pte_page(ptep_get(pte));
150	else
151		vmemmap_walk->remap_pte(pte, addr, vmemmap_walk);
152	vmemmap_walk->nr_walked++;
153
154	return 0;
155}
156
157static const struct mm_walk_ops vmemmap_remap_ops = {
158	.pmd_entry	= vmemmap_pmd_entry,
159	.pte_entry	= vmemmap_pte_entry,
160};
161
162static int vmemmap_remap_range(unsigned long start, unsigned long end,
163			       struct vmemmap_remap_walk *walk)
164{
165	int ret;
166
167	VM_BUG_ON(!PAGE_ALIGNED(start | end));
168
169	mmap_read_lock(&init_mm);
170	ret = walk_kernel_page_table_range(start, end, &vmemmap_remap_ops,
171				    NULL, walk);
172	mmap_read_unlock(&init_mm);
173	if (ret)
174		return ret;
175
176	if (walk->remap_pte && !(walk->flags & VMEMMAP_REMAP_NO_TLB_FLUSH))
177		flush_tlb_kernel_range(start, end);
178
179	return 0;
180}
181
182/*
183 * Free a vmemmap page. A vmemmap page can be allocated from the memblock
184 * allocator or buddy allocator. If the PG_reserved flag is set, it means
185 * that it allocated from the memblock allocator, just free it via the
186 * free_bootmem_page(). Otherwise, use __free_page().
187 */
188static inline void free_vmemmap_page(struct page *page)
189{
190	if (PageReserved(page)) {
191		memmap_boot_pages_add(-1);
192		free_bootmem_page(page);
193	} else {
194		memmap_pages_add(-1);
195		__free_page(page);
196	}
197}
198
199/* Free a list of the vmemmap pages */
200static void free_vmemmap_page_list(struct list_head *list)
201{
202	struct page *page, *next;
203
204	list_for_each_entry_safe(page, next, list, lru)
205		free_vmemmap_page(page);
206}
207
208static void vmemmap_remap_pte(pte_t *pte, unsigned long addr,
209			      struct vmemmap_remap_walk *walk)
210{
211	/*
212	 * Remap the tail pages as read-only to catch illegal write operation
213	 * to the tail pages.
214	 */
215	pgprot_t pgprot = PAGE_KERNEL_RO;
216	struct page *page = pte_page(ptep_get(pte));
217	pte_t entry;
218
219	/* Remapping the head page requires r/w */
220	if (unlikely(addr == walk->reuse_addr)) {
221		pgprot = PAGE_KERNEL;
222		list_del(&walk->reuse_page->lru);
223
224		/*
225		 * Makes sure that preceding stores to the page contents from
226		 * vmemmap_remap_free() become visible before the set_pte_at()
227		 * write.
228		 */
229		smp_wmb();
230	}
231
232	entry = mk_pte(walk->reuse_page, pgprot);
233	list_add(&page->lru, walk->vmemmap_pages);
234	set_pte_at(&init_mm, addr, pte, entry);
235}
236
237/*
238 * How many struct page structs need to be reset. When we reuse the head
239 * struct page, the special metadata (e.g. page->flags or page->mapping)
240 * cannot copy to the tail struct page structs. The invalid value will be
241 * checked in the free_tail_page_prepare(). In order to avoid the message
242 * of "corrupted mapping in tail page". We need to reset at least 4 (one
243 * head struct page struct and three tail struct page structs) struct page
244 * structs.
245 */
246#define NR_RESET_STRUCT_PAGE		4
247
248static inline void reset_struct_pages(struct page *start)
249{
250	struct page *from = start + NR_RESET_STRUCT_PAGE;
251
252	BUILD_BUG_ON(NR_RESET_STRUCT_PAGE * 2 > PAGE_SIZE / sizeof(struct page));
253	memcpy(start, from, sizeof(*from) * NR_RESET_STRUCT_PAGE);
254}
255
256static void vmemmap_restore_pte(pte_t *pte, unsigned long addr,
257				struct vmemmap_remap_walk *walk)
258{
259	pgprot_t pgprot = PAGE_KERNEL;
260	struct page *page;
261	void *to;
262
263	BUG_ON(pte_page(ptep_get(pte)) != walk->reuse_page);
264
265	page = list_first_entry(walk->vmemmap_pages, struct page, lru);
266	list_del(&page->lru);
267	to = page_to_virt(page);
268	copy_page(to, (void *)walk->reuse_addr);
269	reset_struct_pages(to);
270
271	/*
272	 * Makes sure that preceding stores to the page contents become visible
273	 * before the set_pte_at() write.
274	 */
275	smp_wmb();
276	set_pte_at(&init_mm, addr, pte, mk_pte(page, pgprot));
277}
278
279/**
280 * vmemmap_remap_split - split the vmemmap virtual address range [@start, @end)
281 *                      backing PMDs of the directmap into PTEs
282 * @start:     start address of the vmemmap virtual address range that we want
283 *             to remap.
284 * @end:       end address of the vmemmap virtual address range that we want to
285 *             remap.
286 * @reuse:     reuse address.
287 *
288 * Return: %0 on success, negative error code otherwise.
289 */
290static int vmemmap_remap_split(unsigned long start, unsigned long end,
291			       unsigned long reuse)
292{
293	struct vmemmap_remap_walk walk = {
294		.remap_pte	= NULL,
295		.flags		= VMEMMAP_SPLIT_NO_TLB_FLUSH,
296	};
297
298	/* See the comment in the vmemmap_remap_free(). */
299	BUG_ON(start - reuse != PAGE_SIZE);
300
301	return vmemmap_remap_range(reuse, end, &walk);
302}
303
304/**
305 * vmemmap_remap_free - remap the vmemmap virtual address range [@start, @end)
306 *			to the page which @reuse is mapped to, then free vmemmap
307 *			which the range are mapped to.
308 * @start:	start address of the vmemmap virtual address range that we want
309 *		to remap.
310 * @end:	end address of the vmemmap virtual address range that we want to
311 *		remap.
312 * @reuse:	reuse address.
313 * @vmemmap_pages: list to deposit vmemmap pages to be freed.  It is callers
314 *		responsibility to free pages.
315 * @flags:	modifications to vmemmap_remap_walk flags
316 *
317 * Return: %0 on success, negative error code otherwise.
318 */
319static int vmemmap_remap_free(unsigned long start, unsigned long end,
320			      unsigned long reuse,
321			      struct list_head *vmemmap_pages,
322			      unsigned long flags)
323{
324	int ret;
325	struct vmemmap_remap_walk walk = {
326		.remap_pte	= vmemmap_remap_pte,
327		.reuse_addr	= reuse,
328		.vmemmap_pages	= vmemmap_pages,
329		.flags		= flags,
330	};
331	int nid = page_to_nid((struct page *)reuse);
332	gfp_t gfp_mask = GFP_KERNEL | __GFP_NORETRY | __GFP_NOWARN;
333
334	/*
335	 * Allocate a new head vmemmap page to avoid breaking a contiguous
336	 * block of struct page memory when freeing it back to page allocator
337	 * in free_vmemmap_page_list(). This will allow the likely contiguous
338	 * struct page backing memory to be kept contiguous and allowing for
339	 * more allocations of hugepages. Fallback to the currently
340	 * mapped head page in case should it fail to allocate.
341	 */
342	walk.reuse_page = alloc_pages_node(nid, gfp_mask, 0);
343	if (walk.reuse_page) {
344		copy_page(page_to_virt(walk.reuse_page),
345			  (void *)walk.reuse_addr);
346		list_add(&walk.reuse_page->lru, vmemmap_pages);
347		memmap_pages_add(1);
348	}
349
350	/*
351	 * In order to make remapping routine most efficient for the huge pages,
352	 * the routine of vmemmap page table walking has the following rules
353	 * (see more details from the vmemmap_pte_range()):
354	 *
355	 * - The range [@start, @end) and the range [@reuse, @reuse + PAGE_SIZE)
356	 *   should be continuous.
357	 * - The @reuse address is part of the range [@reuse, @end) that we are
358	 *   walking which is passed to vmemmap_remap_range().
359	 * - The @reuse address is the first in the complete range.
360	 *
361	 * So we need to make sure that @start and @reuse meet the above rules.
362	 */
363	BUG_ON(start - reuse != PAGE_SIZE);
364
365	ret = vmemmap_remap_range(reuse, end, &walk);
366	if (ret && walk.nr_walked) {
367		end = reuse + walk.nr_walked * PAGE_SIZE;
368		/*
369		 * vmemmap_pages contains pages from the previous
370		 * vmemmap_remap_range call which failed.  These
371		 * are pages which were removed from the vmemmap.
372		 * They will be restored in the following call.
373		 */
374		walk = (struct vmemmap_remap_walk) {
375			.remap_pte	= vmemmap_restore_pte,
376			.reuse_addr	= reuse,
377			.vmemmap_pages	= vmemmap_pages,
378			.flags		= 0,
379		};
380
381		vmemmap_remap_range(reuse, end, &walk);
382	}
383
384	return ret;
385}
386
387static int alloc_vmemmap_page_list(unsigned long start, unsigned long end,
388				   struct list_head *list)
389{
390	gfp_t gfp_mask = GFP_KERNEL | __GFP_RETRY_MAYFAIL;
391	unsigned long nr_pages = (end - start) >> PAGE_SHIFT;
392	int nid = page_to_nid((struct page *)start);
393	struct page *page, *next;
394	int i;
395
396	for (i = 0; i < nr_pages; i++) {
397		page = alloc_pages_node(nid, gfp_mask, 0);
398		if (!page)
399			goto out;
400		list_add(&page->lru, list);
401	}
402	memmap_pages_add(nr_pages);
403
404	return 0;
405out:
406	list_for_each_entry_safe(page, next, list, lru)
407		__free_page(page);
408	return -ENOMEM;
409}
410
411/**
412 * vmemmap_remap_alloc - remap the vmemmap virtual address range [@start, end)
413 *			 to the page which is from the @vmemmap_pages
414 *			 respectively.
415 * @start:	start address of the vmemmap virtual address range that we want
416 *		to remap.
417 * @end:	end address of the vmemmap virtual address range that we want to
418 *		remap.
419 * @reuse:	reuse address.
420 * @flags:	modifications to vmemmap_remap_walk flags
421 *
422 * Return: %0 on success, negative error code otherwise.
423 */
424static int vmemmap_remap_alloc(unsigned long start, unsigned long end,
425			       unsigned long reuse, unsigned long flags)
426{
427	LIST_HEAD(vmemmap_pages);
428	struct vmemmap_remap_walk walk = {
429		.remap_pte	= vmemmap_restore_pte,
430		.reuse_addr	= reuse,
431		.vmemmap_pages	= &vmemmap_pages,
432		.flags		= flags,
433	};
434
435	/* See the comment in the vmemmap_remap_free(). */
436	BUG_ON(start - reuse != PAGE_SIZE);
437
438	if (alloc_vmemmap_page_list(start, end, &vmemmap_pages))
439		return -ENOMEM;
440
441	return vmemmap_remap_range(reuse, end, &walk);
442}
443
444DEFINE_STATIC_KEY_FALSE(hugetlb_optimize_vmemmap_key);
445EXPORT_SYMBOL(hugetlb_optimize_vmemmap_key);
446
447static bool vmemmap_optimize_enabled = IS_ENABLED(CONFIG_HUGETLB_PAGE_OPTIMIZE_VMEMMAP_DEFAULT_ON);
448static int __init hugetlb_vmemmap_optimize_param(char *buf)
449{
450	return kstrtobool(buf, &vmemmap_optimize_enabled);
451}
452early_param("hugetlb_free_vmemmap", hugetlb_vmemmap_optimize_param);
453
454static int __hugetlb_vmemmap_restore_folio(const struct hstate *h,
455					   struct folio *folio, unsigned long flags)
456{
457	int ret;
458	unsigned long vmemmap_start = (unsigned long)&folio->page, vmemmap_end;
459	unsigned long vmemmap_reuse;
460
461	VM_WARN_ON_ONCE_FOLIO(!folio_test_hugetlb(folio), folio);
462	VM_WARN_ON_ONCE_FOLIO(folio_ref_count(folio), folio);
463
464	if (!folio_test_hugetlb_vmemmap_optimized(folio))
465		return 0;
466
467	if (flags & VMEMMAP_SYNCHRONIZE_RCU)
468		synchronize_rcu();
469
470	vmemmap_end	= vmemmap_start + hugetlb_vmemmap_size(h);
471	vmemmap_reuse	= vmemmap_start;
472	vmemmap_start	+= HUGETLB_VMEMMAP_RESERVE_SIZE;
473
474	/*
475	 * The pages which the vmemmap virtual address range [@vmemmap_start,
476	 * @vmemmap_end) are mapped to are freed to the buddy allocator, and
477	 * the range is mapped to the page which @vmemmap_reuse is mapped to.
478	 * When a HugeTLB page is freed to the buddy allocator, previously
479	 * discarded vmemmap pages must be allocated and remapping.
480	 */
481	ret = vmemmap_remap_alloc(vmemmap_start, vmemmap_end, vmemmap_reuse, flags);
482	if (!ret) {
483		folio_clear_hugetlb_vmemmap_optimized(folio);
484		static_branch_dec(&hugetlb_optimize_vmemmap_key);
485	}
486
487	return ret;
488}
489
490/**
491 * hugetlb_vmemmap_restore_folio - restore previously optimized (by
492 *				hugetlb_vmemmap_optimize_folio()) vmemmap pages which
493 *				will be reallocated and remapped.
494 * @h:		struct hstate.
495 * @folio:     the folio whose vmemmap pages will be restored.
496 *
497 * Return: %0 if @folio's vmemmap pages have been reallocated and remapped,
498 * negative error code otherwise.
499 */
500int hugetlb_vmemmap_restore_folio(const struct hstate *h, struct folio *folio)
501{
502	return __hugetlb_vmemmap_restore_folio(h, folio, VMEMMAP_SYNCHRONIZE_RCU);
503}
504
505/**
506 * hugetlb_vmemmap_restore_folios - restore vmemmap for every folio on the list.
507 * @h:			hstate.
508 * @folio_list:		list of folios.
509 * @non_hvo_folios:	Output list of folios for which vmemmap exists.
510 *
511 * Return: number of folios for which vmemmap was restored, or an error code
512 *		if an error was encountered restoring vmemmap for a folio.
513 *		Folios that have vmemmap are moved to the non_hvo_folios
514 *		list.  Processing of entries stops when the first error is
515 *		encountered. The folio that experienced the error and all
516 *		non-processed folios will remain on folio_list.
517 */
518long hugetlb_vmemmap_restore_folios(const struct hstate *h,
519					struct list_head *folio_list,
520					struct list_head *non_hvo_folios)
521{
522	struct folio *folio, *t_folio;
523	long restored = 0;
524	long ret = 0;
525	unsigned long flags = VMEMMAP_REMAP_NO_TLB_FLUSH | VMEMMAP_SYNCHRONIZE_RCU;
526
527	list_for_each_entry_safe(folio, t_folio, folio_list, lru) {
528		if (folio_test_hugetlb_vmemmap_optimized(folio)) {
529			ret = __hugetlb_vmemmap_restore_folio(h, folio, flags);
530			/* only need to synchronize_rcu() once for each batch */
531			flags &= ~VMEMMAP_SYNCHRONIZE_RCU;
532
533			if (ret)
534				break;
535			restored++;
536		}
537
538		/* Add non-optimized folios to output list */
539		list_move(&folio->lru, non_hvo_folios);
540	}
541
542	if (restored)
543		flush_tlb_all();
544	if (!ret)
545		ret = restored;
546	return ret;
547}
548
549/* Return true iff a HugeTLB whose vmemmap should and can be optimized. */
550static bool vmemmap_should_optimize_folio(const struct hstate *h, struct folio *folio)
551{
552	if (folio_test_hugetlb_vmemmap_optimized(folio))
553		return false;
554
555	if (!READ_ONCE(vmemmap_optimize_enabled))
556		return false;
557
558	if (!hugetlb_vmemmap_optimizable(h))
559		return false;
560
561	return true;
562}
563
564static int __hugetlb_vmemmap_optimize_folio(const struct hstate *h,
565					    struct folio *folio,
566					    struct list_head *vmemmap_pages,
567					    unsigned long flags)
568{
569	int ret = 0;
570	unsigned long vmemmap_start = (unsigned long)&folio->page, vmemmap_end;
571	unsigned long vmemmap_reuse;
572
573	VM_WARN_ON_ONCE_FOLIO(!folio_test_hugetlb(folio), folio);
574	VM_WARN_ON_ONCE_FOLIO(folio_ref_count(folio), folio);
575
576	if (!vmemmap_should_optimize_folio(h, folio))
577		return ret;
578
579	static_branch_inc(&hugetlb_optimize_vmemmap_key);
580
581	if (flags & VMEMMAP_SYNCHRONIZE_RCU)
582		synchronize_rcu();
583	/*
584	 * Very Subtle
585	 * If VMEMMAP_REMAP_NO_TLB_FLUSH is set, TLB flushing is not performed
586	 * immediately after remapping.  As a result, subsequent accesses
587	 * and modifications to struct pages associated with the hugetlb
588	 * page could be to the OLD struct pages.  Set the vmemmap optimized
589	 * flag here so that it is copied to the new head page.  This keeps
590	 * the old and new struct pages in sync.
591	 * If there is an error during optimization, we will immediately FLUSH
592	 * the TLB and clear the flag below.
593	 */
594	folio_set_hugetlb_vmemmap_optimized(folio);
595
596	vmemmap_end	= vmemmap_start + hugetlb_vmemmap_size(h);
597	vmemmap_reuse	= vmemmap_start;
598	vmemmap_start	+= HUGETLB_VMEMMAP_RESERVE_SIZE;
599
600	/*
601	 * Remap the vmemmap virtual address range [@vmemmap_start, @vmemmap_end)
602	 * to the page which @vmemmap_reuse is mapped to.  Add pages previously
603	 * mapping the range to vmemmap_pages list so that they can be freed by
604	 * the caller.
605	 */
606	ret = vmemmap_remap_free(vmemmap_start, vmemmap_end, vmemmap_reuse,
607				 vmemmap_pages, flags);
608	if (ret) {
609		static_branch_dec(&hugetlb_optimize_vmemmap_key);
610		folio_clear_hugetlb_vmemmap_optimized(folio);
611	}
612
613	return ret;
614}
615
616/**
617 * hugetlb_vmemmap_optimize_folio - optimize @folio's vmemmap pages.
618 * @h:		struct hstate.
619 * @folio:     the folio whose vmemmap pages will be optimized.
620 *
621 * This function only tries to optimize @folio's vmemmap pages and does not
622 * guarantee that the optimization will succeed after it returns. The caller
623 * can use folio_test_hugetlb_vmemmap_optimized(@folio) to detect if @folio's
624 * vmemmap pages have been optimized.
625 */
626void hugetlb_vmemmap_optimize_folio(const struct hstate *h, struct folio *folio)
627{
628	LIST_HEAD(vmemmap_pages);
629
630	__hugetlb_vmemmap_optimize_folio(h, folio, &vmemmap_pages, VMEMMAP_SYNCHRONIZE_RCU);
631	free_vmemmap_page_list(&vmemmap_pages);
632}
633
634static int hugetlb_vmemmap_split_folio(const struct hstate *h, struct folio *folio)
635{
636	unsigned long vmemmap_start = (unsigned long)&folio->page, vmemmap_end;
637	unsigned long vmemmap_reuse;
638
639	if (!vmemmap_should_optimize_folio(h, folio))
640		return 0;
641
642	vmemmap_end	= vmemmap_start + hugetlb_vmemmap_size(h);
643	vmemmap_reuse	= vmemmap_start;
644	vmemmap_start	+= HUGETLB_VMEMMAP_RESERVE_SIZE;
645
646	/*
647	 * Split PMDs on the vmemmap virtual address range [@vmemmap_start,
648	 * @vmemmap_end]
649	 */
650	return vmemmap_remap_split(vmemmap_start, vmemmap_end, vmemmap_reuse);
651}
652
653static void __hugetlb_vmemmap_optimize_folios(struct hstate *h,
654					      struct list_head *folio_list,
655					      bool boot)
656{
657	struct folio *folio;
658	int nr_to_optimize;
659	LIST_HEAD(vmemmap_pages);
660	unsigned long flags = VMEMMAP_REMAP_NO_TLB_FLUSH | VMEMMAP_SYNCHRONIZE_RCU;
661
662	nr_to_optimize = 0;
663	list_for_each_entry(folio, folio_list, lru) {
664		int ret;
665		unsigned long spfn, epfn;
666
667		if (boot && folio_test_hugetlb_vmemmap_optimized(folio)) {
668			/*
669			 * Already optimized by pre-HVO, just map the
670			 * mirrored tail page structs RO.
671			 */
672			spfn = (unsigned long)&folio->page;
673			epfn = spfn + pages_per_huge_page(h);
674			vmemmap_wrprotect_hvo(spfn, epfn, folio_nid(folio),
675					HUGETLB_VMEMMAP_RESERVE_SIZE);
676			register_page_bootmem_memmap(pfn_to_section_nr(spfn),
677					&folio->page,
678					HUGETLB_VMEMMAP_RESERVE_SIZE);
679			static_branch_inc(&hugetlb_optimize_vmemmap_key);
680			continue;
681		}
682
683		nr_to_optimize++;
684
685		ret = hugetlb_vmemmap_split_folio(h, folio);
686
687		/*
688		 * Splitting the PMD requires allocating a page, thus let's fail
689		 * early once we encounter the first OOM. No point in retrying
690		 * as it can be dynamically done on remap with the memory
691		 * we get back from the vmemmap deduplication.
692		 */
693		if (ret == -ENOMEM)
694			break;
695	}
696
697	if (!nr_to_optimize)
698		/*
699		 * All pre-HVO folios, nothing left to do. It's ok if
700		 * there is a mix of pre-HVO and not yet HVO-ed folios
701		 * here, as __hugetlb_vmemmap_optimize_folio() will
702		 * skip any folios that already have the optimized flag
703		 * set, see vmemmap_should_optimize_folio().
704		 */
705		goto out;
706
707	flush_tlb_all();
708
709	list_for_each_entry(folio, folio_list, lru) {
710		int ret;
711
712		ret = __hugetlb_vmemmap_optimize_folio(h, folio, &vmemmap_pages, flags);
713		/* only need to synchronize_rcu() once for each batch */
714		flags &= ~VMEMMAP_SYNCHRONIZE_RCU;
715
716		/*
717		 * Pages to be freed may have been accumulated.  If we
718		 * encounter an ENOMEM,  free what we have and try again.
719		 * This can occur in the case that both splitting fails
720		 * halfway and head page allocation also failed. In this
721		 * case __hugetlb_vmemmap_optimize_folio() would free memory
722		 * allowing more vmemmap remaps to occur.
723		 */
724		if (ret == -ENOMEM && !list_empty(&vmemmap_pages)) {
725			flush_tlb_all();
726			free_vmemmap_page_list(&vmemmap_pages);
727			INIT_LIST_HEAD(&vmemmap_pages);
728			__hugetlb_vmemmap_optimize_folio(h, folio, &vmemmap_pages, flags);
729		}
730	}
731
732out:
733	flush_tlb_all();
734	free_vmemmap_page_list(&vmemmap_pages);
735}
736
737void hugetlb_vmemmap_optimize_folios(struct hstate *h, struct list_head *folio_list)
738{
739	__hugetlb_vmemmap_optimize_folios(h, folio_list, false);
740}
741
742void hugetlb_vmemmap_optimize_bootmem_folios(struct hstate *h, struct list_head *folio_list)
743{
744	__hugetlb_vmemmap_optimize_folios(h, folio_list, true);
745}
746
747#ifdef CONFIG_SPARSEMEM_VMEMMAP_PREINIT
748
749/* Return true of a bootmem allocated HugeTLB page should be pre-HVO-ed */
750static bool vmemmap_should_optimize_bootmem_page(struct huge_bootmem_page *m)
751{
752	unsigned long section_size, psize, pmd_vmemmap_size;
753	phys_addr_t paddr;
754
755	if (!READ_ONCE(vmemmap_optimize_enabled))
756		return false;
757
758	if (!hugetlb_vmemmap_optimizable(m->hstate))
759		return false;
760
761	psize = huge_page_size(m->hstate);
762	paddr = virt_to_phys(m);
763
764	/*
765	 * Pre-HVO only works if the bootmem huge page
766	 * is aligned to the section size.
767	 */
768	section_size = (1UL << PA_SECTION_SHIFT);
769	if (!IS_ALIGNED(paddr, section_size) ||
770	    !IS_ALIGNED(psize, section_size))
771		return false;
772
773	/*
774	 * The pre-HVO code does not deal with splitting PMDS,
775	 * so the bootmem page must be aligned to the number
776	 * of base pages that can be mapped with one vmemmap PMD.
777	 */
778	pmd_vmemmap_size = (PMD_SIZE / (sizeof(struct page))) << PAGE_SHIFT;
779	if (!IS_ALIGNED(paddr, pmd_vmemmap_size) ||
780	    !IS_ALIGNED(psize, pmd_vmemmap_size))
781		return false;
782
783	return true;
784}
785
786/*
787 * Initialize memmap section for a gigantic page, HVO-style.
788 */
789void __init hugetlb_vmemmap_init_early(int nid)
790{
791	unsigned long psize, paddr, section_size;
792	unsigned long ns, i, pnum, pfn, nr_pages;
793	unsigned long start, end;
794	struct huge_bootmem_page *m = NULL;
795	void *map;
796
797	/*
798	 * Noting to do if bootmem pages were not allocated
799	 * early in boot, or if HVO wasn't enabled in the
800	 * first place.
801	 */
802	if (!hugetlb_bootmem_allocated())
803		return;
804
805	if (!READ_ONCE(vmemmap_optimize_enabled))
806		return;
807
808	section_size = (1UL << PA_SECTION_SHIFT);
809
810	list_for_each_entry(m, &huge_boot_pages[nid], list) {
811		if (!vmemmap_should_optimize_bootmem_page(m))
812			continue;
813
814		nr_pages = pages_per_huge_page(m->hstate);
815		psize = nr_pages << PAGE_SHIFT;
816		paddr = virt_to_phys(m);
817		pfn = PHYS_PFN(paddr);
818		map = pfn_to_page(pfn);
819		start = (unsigned long)map;
820		end = start + nr_pages * sizeof(struct page);
821
822		if (vmemmap_populate_hvo(start, end, nid,
823					HUGETLB_VMEMMAP_RESERVE_SIZE) < 0)
824			continue;
825
826		memmap_boot_pages_add(HUGETLB_VMEMMAP_RESERVE_SIZE / PAGE_SIZE);
827
828		pnum = pfn_to_section_nr(pfn);
829		ns = psize / section_size;
830
831		for (i = 0; i < ns; i++) {
832			sparse_init_early_section(nid, map, pnum,
833					SECTION_IS_VMEMMAP_PREINIT);
834			map += section_map_size();
835			pnum++;
836		}
837
838		m->flags |= HUGE_BOOTMEM_HVO;
839	}
840}
841
842void __init hugetlb_vmemmap_init_late(int nid)
843{
844	struct huge_bootmem_page *m, *tm;
845	unsigned long phys, nr_pages, start, end;
846	unsigned long pfn, nr_mmap;
847	struct hstate *h;
848	void *map;
849
850	if (!hugetlb_bootmem_allocated())
851		return;
852
853	if (!READ_ONCE(vmemmap_optimize_enabled))
854		return;
855
856	list_for_each_entry_safe(m, tm, &huge_boot_pages[nid], list) {
857		if (!(m->flags & HUGE_BOOTMEM_HVO))
858			continue;
859
860		phys = virt_to_phys(m);
861		h = m->hstate;
862		pfn = PHYS_PFN(phys);
863		nr_pages = pages_per_huge_page(h);
864
865		if (!hugetlb_bootmem_page_zones_valid(nid, m)) {
866			/*
867			 * Oops, the hugetlb page spans multiple zones.
868			 * Remove it from the list, and undo HVO.
869			 */
870			list_del(&m->list);
871
872			map = pfn_to_page(pfn);
873
874			start = (unsigned long)map;
875			end = start + nr_pages * sizeof(struct page);
876
877			vmemmap_undo_hvo(start, end, nid,
878					 HUGETLB_VMEMMAP_RESERVE_SIZE);
879			nr_mmap = end - start - HUGETLB_VMEMMAP_RESERVE_SIZE;
880			memmap_boot_pages_add(DIV_ROUND_UP(nr_mmap, PAGE_SIZE));
881
882			memblock_phys_free(phys, huge_page_size(h));
883			continue;
884		} else
885			m->flags |= HUGE_BOOTMEM_ZONES_VALID;
886	}
887}
888#endif
889
890static const struct ctl_table hugetlb_vmemmap_sysctls[] = {
891	{
892		.procname	= "hugetlb_optimize_vmemmap",
893		.data		= &vmemmap_optimize_enabled,
894		.maxlen		= sizeof(vmemmap_optimize_enabled),
895		.mode		= 0644,
896		.proc_handler	= proc_dobool,
897	},
898};
899
900static int __init hugetlb_vmemmap_init(void)
901{
902	const struct hstate *h;
903
904	/* HUGETLB_VMEMMAP_RESERVE_SIZE should cover all used struct pages */
905	BUILD_BUG_ON(__NR_USED_SUBPAGE > HUGETLB_VMEMMAP_RESERVE_PAGES);
906
907	for_each_hstate(h) {
908		if (hugetlb_vmemmap_optimizable(h)) {
909			register_sysctl_init("vm", hugetlb_vmemmap_sysctls);
910			break;
911		}
912	}
913	return 0;
914}
915late_initcall(hugetlb_vmemmap_init);