1 /*
2 * Copyright IBM Corporation, 2013
3 * Author Aneesh Kumar K.V <aneesh.kumar@linux.vnet.ibm.com>
4 *
5 * This program is free software; you can redistribute it and/or modify it
6 * under the terms of version 2.1 of the GNU Lesser General Public License
7 * as published by the Free Software Foundation.
8 *
9 * This program is distributed in the hope that it would be useful, but
10 * WITHOUT ANY WARRANTY; without even the implied warranty of
11 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.
12 *
13 */
14
15 /*
16 * PPC64 THP Support for hash based MMUs
17 */
18 #include <linux/mm.h>
19 #include <asm/machdep.h>
20
invalidate_old_hpte(unsigned long vsid,unsigned long addr,pmd_t * pmdp,unsigned int psize,int ssize)21 static void invalidate_old_hpte(unsigned long vsid, unsigned long addr,
22 pmd_t *pmdp, unsigned int psize, int ssize)
23 {
24 int i, max_hpte_count, valid;
25 unsigned long s_addr;
26 unsigned char *hpte_slot_array;
27 unsigned long hidx, shift, vpn, hash, slot;
28
29 s_addr = addr & HPAGE_PMD_MASK;
30 hpte_slot_array = get_hpte_slot_array(pmdp);
31 /*
32 * IF we try to do a HUGE PTE update after a withdraw is done.
33 * we will find the below NULL. This happens when we do
34 * split_huge_page_pmd
35 */
36 if (!hpte_slot_array)
37 return;
38
39 if (ppc_md.hugepage_invalidate)
40 return ppc_md.hugepage_invalidate(vsid, s_addr, hpte_slot_array,
41 psize, ssize);
42 /*
43 * No bluk hpte removal support, invalidate each entry
44 */
45 shift = mmu_psize_defs[psize].shift;
46 max_hpte_count = HPAGE_PMD_SIZE >> shift;
47 for (i = 0; i < max_hpte_count; i++) {
48 /*
49 * 8 bits per each hpte entries
50 * 000| [ secondary group (one bit) | hidx (3 bits) | valid bit]
51 */
52 valid = hpte_valid(hpte_slot_array, i);
53 if (!valid)
54 continue;
55 hidx = hpte_hash_index(hpte_slot_array, i);
56
57 /* get the vpn */
58 addr = s_addr + (i * (1ul << shift));
59 vpn = hpt_vpn(addr, vsid, ssize);
60 hash = hpt_hash(vpn, shift, ssize);
61 if (hidx & _PTEIDX_SECONDARY)
62 hash = ~hash;
63
64 slot = (hash & htab_hash_mask) * HPTES_PER_GROUP;
65 slot += hidx & _PTEIDX_GROUP_IX;
66 ppc_md.hpte_invalidate(slot, vpn, psize,
67 MMU_PAGE_16M, ssize, 0);
68 }
69 }
70
71
__hash_page_thp(unsigned long ea,unsigned long access,unsigned long vsid,pmd_t * pmdp,unsigned long trap,int local,int ssize,unsigned int psize)72 int __hash_page_thp(unsigned long ea, unsigned long access, unsigned long vsid,
73 pmd_t *pmdp, unsigned long trap, int local, int ssize,
74 unsigned int psize)
75 {
76 unsigned int index, valid;
77 unsigned char *hpte_slot_array;
78 unsigned long rflags, pa, hidx;
79 unsigned long old_pmd, new_pmd;
80 int ret, lpsize = MMU_PAGE_16M;
81 unsigned long vpn, hash, shift, slot;
82
83 /*
84 * atomically mark the linux large page PMD busy and dirty
85 */
86 do {
87 pmd_t pmd = ACCESS_ONCE(*pmdp);
88
89 old_pmd = pmd_val(pmd);
90 /* If PMD busy, retry the access */
91 if (unlikely(old_pmd & _PAGE_BUSY))
92 return 0;
93 /* If PMD is trans splitting retry the access */
94 if (unlikely(old_pmd & _PAGE_SPLITTING))
95 return 0;
96 /* If PMD permissions don't match, take page fault */
97 if (unlikely(access & ~old_pmd))
98 return 1;
99 /*
100 * Try to lock the PTE, add ACCESSED and DIRTY if it was
101 * a write access
102 */
103 new_pmd = old_pmd | _PAGE_BUSY | _PAGE_ACCESSED;
104 if (access & _PAGE_RW)
105 new_pmd |= _PAGE_DIRTY;
106 } while (old_pmd != __cmpxchg_u64((unsigned long *)pmdp,
107 old_pmd, new_pmd));
108 /*
109 * PP bits. _PAGE_USER is already PP bit 0x2, so we only
110 * need to add in 0x1 if it's a read-only user page
111 */
112 rflags = new_pmd & _PAGE_USER;
113 if ((new_pmd & _PAGE_USER) && !((new_pmd & _PAGE_RW) &&
114 (new_pmd & _PAGE_DIRTY)))
115 rflags |= 0x1;
116 /*
117 * _PAGE_EXEC -> HW_NO_EXEC since it's inverted
118 */
119 rflags |= ((new_pmd & _PAGE_EXEC) ? 0 : HPTE_R_N);
120
121 #if 0
122 if (!cpu_has_feature(CPU_FTR_COHERENT_ICACHE)) {
123
124 /*
125 * No CPU has hugepages but lacks no execute, so we
126 * don't need to worry about that case
127 */
128 rflags = hash_page_do_lazy_icache(rflags, __pte(old_pte), trap);
129 }
130 #endif
131 /*
132 * Find the slot index details for this ea, using base page size.
133 */
134 shift = mmu_psize_defs[psize].shift;
135 index = (ea & ~HPAGE_PMD_MASK) >> shift;
136 BUG_ON(index >= 4096);
137
138 vpn = hpt_vpn(ea, vsid, ssize);
139 hpte_slot_array = get_hpte_slot_array(pmdp);
140 if (psize == MMU_PAGE_4K) {
141 /*
142 * invalidate the old hpte entry if we have that mapped via 64K
143 * base page size. This is because demote_segment won't flush
144 * hash page table entries.
145 */
146 if ((old_pmd & _PAGE_HASHPTE) && !(old_pmd & _PAGE_COMBO))
147 invalidate_old_hpte(vsid, ea, pmdp, MMU_PAGE_64K, ssize);
148 }
149
150 valid = hpte_valid(hpte_slot_array, index);
151 if (valid) {
152 /* update the hpte bits */
153 hash = hpt_hash(vpn, shift, ssize);
154 hidx = hpte_hash_index(hpte_slot_array, index);
155 if (hidx & _PTEIDX_SECONDARY)
156 hash = ~hash;
157 slot = (hash & htab_hash_mask) * HPTES_PER_GROUP;
158 slot += hidx & _PTEIDX_GROUP_IX;
159
160 ret = ppc_md.hpte_updatepp(slot, rflags, vpn,
161 psize, lpsize, ssize, local);
162 /*
163 * We failed to update, try to insert a new entry.
164 */
165 if (ret == -1) {
166 /*
167 * large pte is marked busy, so we can be sure
168 * nobody is looking at hpte_slot_array. hence we can
169 * safely update this here.
170 */
171 valid = 0;
172 hpte_slot_array[index] = 0;
173 }
174 }
175
176 if (!valid) {
177 unsigned long hpte_group;
178
179 hash = hpt_hash(vpn, shift, ssize);
180 /* insert new entry */
181 pa = pmd_pfn(__pmd(old_pmd)) << PAGE_SHIFT;
182 new_pmd |= _PAGE_HASHPTE;
183
184 /* Add in WIMG bits */
185 rflags |= (new_pmd & (_PAGE_WRITETHRU | _PAGE_NO_CACHE |
186 _PAGE_GUARDED));
187 /*
188 * enable the memory coherence always
189 */
190 rflags |= HPTE_R_M;
191 repeat:
192 hpte_group = ((hash & htab_hash_mask) * HPTES_PER_GROUP) & ~0x7UL;
193
194 /* Insert into the hash table, primary slot */
195 slot = ppc_md.hpte_insert(hpte_group, vpn, pa, rflags, 0,
196 psize, lpsize, ssize);
197 /*
198 * Primary is full, try the secondary
199 */
200 if (unlikely(slot == -1)) {
201 hpte_group = ((~hash & htab_hash_mask) *
202 HPTES_PER_GROUP) & ~0x7UL;
203 slot = ppc_md.hpte_insert(hpte_group, vpn, pa,
204 rflags, HPTE_V_SECONDARY,
205 psize, lpsize, ssize);
206 if (slot == -1) {
207 if (mftb() & 0x1)
208 hpte_group = ((hash & htab_hash_mask) *
209 HPTES_PER_GROUP) & ~0x7UL;
210
211 ppc_md.hpte_remove(hpte_group);
212 goto repeat;
213 }
214 }
215 /*
216 * Hypervisor failure. Restore old pmd and return -1
217 * similar to __hash_page_*
218 */
219 if (unlikely(slot == -2)) {
220 *pmdp = __pmd(old_pmd);
221 hash_failure_debug(ea, access, vsid, trap, ssize,
222 psize, lpsize, old_pmd);
223 return -1;
224 }
225 /*
226 * large pte is marked busy, so we can be sure
227 * nobody is looking at hpte_slot_array. hence we can
228 * safely update this here.
229 */
230 mark_hpte_slot_valid(hpte_slot_array, index, slot);
231 }
232 /*
233 * Mark the pte with _PAGE_COMBO, if we are trying to hash it with
234 * base page size 4k.
235 */
236 if (psize == MMU_PAGE_4K)
237 new_pmd |= _PAGE_COMBO;
238 /*
239 * The hpte valid is stored in the pgtable whose address is in the
240 * second half of the PMD. Order this against clearing of the busy bit in
241 * huge pmd.
242 */
243 smp_wmb();
244 *pmdp = __pmd(new_pmd & ~_PAGE_BUSY);
245 return 0;
246 }
247