1// SPDX-License-Identifier: GPL-2.0
  2
  3/*
  4 *  Handling Page Tables through page fragments
  5 *
  6 */
  7
  8#include <linux/kernel.h>
  9#include <linux/gfp.h>
 10#include <linux/mm.h>
 11#include <linux/percpu.h>
 12#include <linux/hardirq.h>
 13#include <linux/hugetlb.h>
 14#include <asm/pgalloc.h>
 15#include <asm/tlbflush.h>
 16#include <asm/tlb.h>
 17
 18void pte_frag_destroy(void *pte_frag)
 19{
 20	int count;
 21	struct ptdesc *ptdesc;
 22
 23	ptdesc = virt_to_ptdesc(pte_frag);
 24	/* drop all the pending references */
 25	count = ((unsigned long)pte_frag & ~PAGE_MASK) >> PTE_FRAG_SIZE_SHIFT;
 26	/* We allow PTE_FRAG_NR fragments from a PTE page */
 27	if (atomic_sub_and_test(PTE_FRAG_NR - count, &ptdesc->pt_frag_refcount)) {
 28		pagetable_pte_dtor(ptdesc);
 29		pagetable_free(ptdesc);
 30	}
 31}
 32
 33static pte_t *get_pte_from_cache(struct mm_struct *mm)
 34{
 35	void *pte_frag, *ret;
 36
 37	if (PTE_FRAG_NR == 1)
 38		return NULL;
 39
 40	spin_lock(&mm->page_table_lock);
 41	ret = pte_frag_get(&mm->context);
 42	if (ret) {
 43		pte_frag = ret + PTE_FRAG_SIZE;
 44		/*
 45		 * If we have taken up all the fragments mark PTE page NULL
 46		 */
 47		if (((unsigned long)pte_frag & ~PAGE_MASK) == 0)
 48			pte_frag = NULL;
 49		pte_frag_set(&mm->context, pte_frag);
 50	}
 51	spin_unlock(&mm->page_table_lock);
 52	return (pte_t *)ret;
 53}
 54
 55static pte_t *__alloc_for_ptecache(struct mm_struct *mm, int kernel)
 56{
 57	void *ret = NULL;
 58	struct ptdesc *ptdesc;
 59
 60	if (!kernel) {
 61		ptdesc = pagetable_alloc(PGALLOC_GFP | __GFP_ACCOUNT, 0);
 62		if (!ptdesc)
 63			return NULL;
 64		if (!pagetable_pte_ctor(ptdesc)) {
 65			pagetable_free(ptdesc);
 66			return NULL;
 67		}
 68	} else {
 69		ptdesc = pagetable_alloc(PGALLOC_GFP, 0);
 70		if (!ptdesc)
 71			return NULL;
 72	}
 73
 74	atomic_set(&ptdesc->pt_frag_refcount, 1);
 75
 76	ret = ptdesc_address(ptdesc);
 77	/*
 78	 * if we support only one fragment just return the
 79	 * allocated page.
 80	 */
 81	if (PTE_FRAG_NR == 1)
 82		return ret;
 83	spin_lock(&mm->page_table_lock);
 84	/*
 85	 * If we find ptdesc_page set, we return
 86	 * the allocated page with single fragment
 87	 * count.
 88	 */
 89	if (likely(!pte_frag_get(&mm->context))) {
 90		atomic_set(&ptdesc->pt_frag_refcount, PTE_FRAG_NR);
 91		pte_frag_set(&mm->context, ret + PTE_FRAG_SIZE);
 92	}
 93	spin_unlock(&mm->page_table_lock);
 94
 95	return (pte_t *)ret;
 96}
 97
 98pte_t *pte_fragment_alloc(struct mm_struct *mm, int kernel)
 99{
100	pte_t *pte;
101
102	pte = get_pte_from_cache(mm);
103	if (pte)
104		return pte;
105
106	return __alloc_for_ptecache(mm, kernel);
107}
108
109static void pte_free_now(struct rcu_head *head)
110{
111	struct ptdesc *ptdesc;
112
113	ptdesc = container_of(head, struct ptdesc, pt_rcu_head);
114	pagetable_pte_dtor(ptdesc);
115	pagetable_free(ptdesc);
116}
117
118void pte_fragment_free(unsigned long *table, int kernel)
119{
120	struct ptdesc *ptdesc = virt_to_ptdesc(table);
121
122	if (pagetable_is_reserved(ptdesc))
123		return free_reserved_ptdesc(ptdesc);
124
125	BUG_ON(atomic_read(&ptdesc->pt_frag_refcount) <= 0);
126	if (atomic_dec_and_test(&ptdesc->pt_frag_refcount)) {
127		if (kernel)
128			pagetable_free(ptdesc);
129		else if (folio_test_clear_active(ptdesc_folio(ptdesc)))
130			call_rcu(&ptdesc->pt_rcu_head, pte_free_now);
131		else
132			pte_free_now(&ptdesc->pt_rcu_head);
133	}
134}
135
136#ifdef CONFIG_TRANSPARENT_HUGEPAGE
137void pte_free_defer(struct mm_struct *mm, pgtable_t pgtable)
138{
139	struct page *page;
140
141	page = virt_to_page(pgtable);
142	SetPageActive(page);
143	pte_fragment_free((unsigned long *)pgtable, 0);
144}
145#endif /* CONFIG_TRANSPARENT_HUGEPAGE */