xref: /linux/arch/powerpc/mm/book3s64/hugetlbpage.c (revision f43d2ffb47c9e86f5ec24e1de6ce6da6808634a2)
1 // SPDX-License-Identifier: GPL-2.0
2 /*
3  * PPC64 Huge TLB Page Support for hash based MMUs (POWER4 and later)
4  *
5  * Copyright (C) 2003 David Gibson, IBM Corporation.
6  *
7  * Based on the IA-32 version:
8  * Copyright (C) 2002, Rohit Seth <rohit.seth@intel.com>
9  */
10 
11 #include <linux/mm.h>
12 #include <linux/hugetlb.h>
13 #include <asm/cacheflush.h>
14 #include <asm/machdep.h>
15 
16 unsigned int hpage_shift;
17 EXPORT_SYMBOL(hpage_shift);
18 
19 int __hash_page_huge(unsigned long ea, unsigned long access, unsigned long vsid,
20 		     pte_t *ptep, unsigned long trap, unsigned long flags,
21 		     int ssize, unsigned int shift, unsigned int mmu_psize)
22 {
23 	real_pte_t rpte;
24 	unsigned long vpn;
25 	unsigned long old_pte, new_pte;
26 	unsigned long rflags, pa;
27 	long slot, offset;
28 
29 	BUG_ON(shift != mmu_psize_defs[mmu_psize].shift);
30 
31 	/* Search the Linux page table for a match with va */
32 	vpn = hpt_vpn(ea, vsid, ssize);
33 
34 	/*
35 	 * At this point, we have a pte (old_pte) which can be used to build
36 	 * or update an HPTE. There are 2 cases:
37 	 *
38 	 * 1. There is a valid (present) pte with no associated HPTE (this is
39 	 *	the most common case)
40 	 * 2. There is a valid (present) pte with an associated HPTE. The
41 	 *	current values of the pp bits in the HPTE prevent access
42 	 *	because we are doing software DIRTY bit management and the
43 	 *	page is currently not DIRTY.
44 	 */
45 
46 
47 	do {
48 		old_pte = pte_val(*ptep);
49 		/* If PTE busy, retry the access */
50 		if (unlikely(old_pte & H_PAGE_BUSY))
51 			return 0;
52 		/* If PTE permissions don't match, take page fault */
53 		if (unlikely(!check_pte_access(access, old_pte)))
54 			return 1;
55 
56 		/*
57 		 * Try to lock the PTE, add ACCESSED and DIRTY if it was
58 		 * a write access
59 		 */
60 		new_pte = old_pte | H_PAGE_BUSY | _PAGE_ACCESSED;
61 		if (access & _PAGE_WRITE)
62 			new_pte |= _PAGE_DIRTY;
63 	} while(!pte_xchg(ptep, __pte(old_pte), __pte(new_pte)));
64 
65 	/* Make sure this is a hugetlb entry */
66 	if (old_pte & (H_PAGE_THP_HUGE | _PAGE_DEVMAP))
67 		return 0;
68 
69 	rflags = htab_convert_pte_flags(new_pte, flags);
70 	if (unlikely(mmu_psize == MMU_PAGE_16G))
71 		offset = PTRS_PER_PUD;
72 	else
73 		offset = PTRS_PER_PMD;
74 	rpte = __real_pte(__pte(old_pte), ptep, offset);
75 
76 	if (!cpu_has_feature(CPU_FTR_COHERENT_ICACHE))
77 		/*
78 		 * No CPU has hugepages but lacks no execute, so we
79 		 * don't need to worry about that case
80 		 */
81 		rflags = hash_page_do_lazy_icache(rflags, __pte(old_pte), trap);
82 
83 	/* Check if pte already has an hpte (case 2) */
84 	if (unlikely(old_pte & H_PAGE_HASHPTE)) {
85 		/* There MIGHT be an HPTE for this pte */
86 		unsigned long gslot;
87 
88 		gslot = pte_get_hash_gslot(vpn, shift, ssize, rpte, 0);
89 		if (mmu_hash_ops.hpte_updatepp(gslot, rflags, vpn, mmu_psize,
90 					       mmu_psize, ssize, flags) == -1)
91 			old_pte &= ~_PAGE_HPTEFLAGS;
92 	}
93 
94 	if (likely(!(old_pte & H_PAGE_HASHPTE))) {
95 		unsigned long hash = hpt_hash(vpn, shift, ssize);
96 
97 		pa = pte_pfn(__pte(old_pte)) << PAGE_SHIFT;
98 
99 		/* clear HPTE slot informations in new PTE */
100 		new_pte = (new_pte & ~_PAGE_HPTEFLAGS) | H_PAGE_HASHPTE;
101 
102 		slot = hpte_insert_repeating(hash, vpn, pa, rflags, 0,
103 					     mmu_psize, ssize);
104 
105 		/*
106 		 * Hypervisor failure. Restore old pte and return -1
107 		 * similar to __hash_page_*
108 		 */
109 		if (unlikely(slot == -2)) {
110 			*ptep = __pte(old_pte);
111 			hash_failure_debug(ea, access, vsid, trap, ssize,
112 					   mmu_psize, mmu_psize, old_pte);
113 			return -1;
114 		}
115 
116 		new_pte |= pte_set_hidx(ptep, rpte, 0, slot, offset);
117 	}
118 
119 	/*
120 	 * No need to use ldarx/stdcx here
121 	 */
122 	*ptep = __pte(new_pte & ~H_PAGE_BUSY);
123 	return 0;
124 }
125 
126 pte_t huge_ptep_modify_prot_start(struct vm_area_struct *vma,
127 				  unsigned long addr, pte_t *ptep)
128 {
129 	unsigned long pte_val;
130 	/*
131 	 * Clear the _PAGE_PRESENT so that no hardware parallel update is
132 	 * possible. Also keep the pte_present true so that we don't take
133 	 * wrong fault.
134 	 */
135 	pte_val = pte_update(vma->vm_mm, addr, ptep,
136 			     _PAGE_PRESENT, _PAGE_INVALID, 1);
137 
138 	return __pte(pte_val);
139 }
140 
141 void huge_ptep_modify_prot_commit(struct vm_area_struct *vma, unsigned long addr,
142 				  pte_t *ptep, pte_t old_pte, pte_t pte)
143 {
144 
145 	if (radix_enabled())
146 		return radix__huge_ptep_modify_prot_commit(vma, addr, ptep,
147 							   old_pte, pte);
148 	set_huge_pte_at(vma->vm_mm, addr, ptep, pte);
149 }
150 
151 void hugetlbpage_init_default(void)
152 {
153 	/* Set default large page size. Currently, we pick 16M or 1M
154 	 * depending on what is available
155 	 */
156 	if (mmu_psize_defs[MMU_PAGE_16M].shift)
157 		hpage_shift = mmu_psize_defs[MMU_PAGE_16M].shift;
158 	else if (mmu_psize_defs[MMU_PAGE_1M].shift)
159 		hpage_shift = mmu_psize_defs[MMU_PAGE_1M].shift;
160 	else if (mmu_psize_defs[MMU_PAGE_2M].shift)
161 		hpage_shift = mmu_psize_defs[MMU_PAGE_2M].shift;
162 }
163