2 * Copyright IBM Corporation, 2015
3 * Author Aneesh Kumar K.V <aneesh.kumar@linux.vnet.ibm.com>
5 * This program is free software; you can redistribute it and/or modify it
6 * under the terms of version 2 of the GNU Lesser General Public License
7 * as published by the Free Software Foundation.
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.
16 #include <asm/machdep.h>
21 bool __rpte_sub_valid(real_pte_t rpte, unsigned long index)
24 unsigned long ptev = pte_val(rpte.pte);
26 g_idx = (ptev & _PAGE_COMBO_VALID) >> _PAGE_F_GIX_SHIFT;
28 if (g_idx & (0x1 << index))
36 static unsigned long mark_subptegroup_valid(unsigned long ptev, unsigned long index)
40 if (!(ptev & _PAGE_COMBO))
45 return ptev | (g_idx << _PAGE_F_GIX_SHIFT);
48 int __hash_page_4K(unsigned long ea, unsigned long access, unsigned long vsid,
49 pte_t *ptep, unsigned long trap, unsigned long flags,
50 int ssize, int subpg_prot)
54 unsigned long hpte_group;
55 unsigned int subpg_index;
56 unsigned long rflags, pa, hidx;
57 unsigned long old_pte, new_pte, subpg_pte;
58 unsigned long vpn, hash, slot;
59 unsigned long shift = mmu_psize_defs[MMU_PAGE_4K].shift;
62 * atomically mark the linux large page PTE busy and dirty
65 pte_t pte = READ_ONCE(*ptep);
67 old_pte = pte_val(pte);
68 /* If PTE busy, retry the access */
69 if (unlikely(old_pte & _PAGE_BUSY))
71 /* If PTE permissions don't match, take page fault */
72 if (unlikely(access & ~old_pte))
75 * Try to lock the PTE, add ACCESSED and DIRTY if it was
76 * a write access. Since this is 4K insert of 64K page size
77 * also add _PAGE_COMBO
79 new_pte = old_pte | _PAGE_BUSY | _PAGE_ACCESSED | _PAGE_COMBO;
80 if (access & _PAGE_RW)
81 new_pte |= _PAGE_DIRTY;
82 } while (old_pte != __cmpxchg_u64((unsigned long *)ptep,
85 * Handle the subpage protection bits
87 subpg_pte = new_pte & ~subpg_prot;
88 rflags = htab_convert_pte_flags(subpg_pte);
90 if (!cpu_has_feature(CPU_FTR_NOEXECUTE) &&
91 !cpu_has_feature(CPU_FTR_COHERENT_ICACHE)) {
94 * No CPU has hugepages but lacks no execute, so we
95 * don't need to worry about that case
97 rflags = hash_page_do_lazy_icache(rflags, __pte(old_pte), trap);
100 subpg_index = (ea & (PAGE_SIZE - 1)) >> shift;
101 vpn = hpt_vpn(ea, vsid, ssize);
102 rpte = __real_pte(__pte(old_pte), ptep);
104 *None of the sub 4k page is hashed
106 if (!(old_pte & _PAGE_HASHPTE))
107 goto htab_insert_hpte;
109 * Check if the pte was already inserted into the hash table
110 * as a 64k HW page, and invalidate the 64k HPTE if so.
112 if (!(old_pte & _PAGE_COMBO)) {
113 flush_hash_page(vpn, rpte, MMU_PAGE_64K, ssize, flags);
114 old_pte &= ~_PAGE_HASHPTE | _PAGE_F_GIX | _PAGE_F_SECOND;
115 goto htab_insert_hpte;
118 * Check for sub page valid and update
120 if (__rpte_sub_valid(rpte, subpg_index)) {
123 hash = hpt_hash(vpn, shift, ssize);
124 hidx = __rpte_to_hidx(rpte, subpg_index);
125 if (hidx & _PTEIDX_SECONDARY)
127 slot = (hash & htab_hash_mask) * HPTES_PER_GROUP;
128 slot += hidx & _PTEIDX_GROUP_IX;
130 ret = ppc_md.hpte_updatepp(slot, rflags, vpn,
131 MMU_PAGE_4K, MMU_PAGE_4K,
134 *if we failed because typically the HPTE wasn't really here
135 * we try an insertion.
138 goto htab_insert_hpte;
140 *ptep = __pte(new_pte & ~_PAGE_BUSY);
146 * handle _PAGE_4K_PFN case
148 if (old_pte & _PAGE_4K_PFN) {
150 * All the sub 4k page have the same
153 pa = pte_pfn(__pte(old_pte)) << HW_PAGE_SHIFT;
155 pa = pte_pfn(__pte(old_pte)) << PAGE_SHIFT;
156 pa += (subpg_index << shift);
158 hash = hpt_hash(vpn, shift, ssize);
160 hpte_group = ((hash & htab_hash_mask) * HPTES_PER_GROUP) & ~0x7UL;
162 /* Insert into the hash table, primary slot */
163 slot = ppc_md.hpte_insert(hpte_group, vpn, pa, rflags, 0,
164 MMU_PAGE_4K, MMU_PAGE_4K, ssize);
166 * Primary is full, try the secondary
168 if (unlikely(slot == -1)) {
169 hpte_group = ((~hash & htab_hash_mask) * HPTES_PER_GROUP) & ~0x7UL;
170 slot = ppc_md.hpte_insert(hpte_group, vpn, pa,
171 rflags, HPTE_V_SECONDARY,
172 MMU_PAGE_4K, MMU_PAGE_4K, ssize);
175 hpte_group = ((hash & htab_hash_mask) *
176 HPTES_PER_GROUP) & ~0x7UL;
177 ppc_md.hpte_remove(hpte_group);
179 * FIXME!! Should be try the group from which we removed ?
185 * Hypervisor failure. Restore old pmd and return -1
186 * similar to __hash_page_*
188 if (unlikely(slot == -2)) {
189 *ptep = __pte(old_pte);
190 hash_failure_debug(ea, access, vsid, trap, ssize,
191 MMU_PAGE_4K, MMU_PAGE_4K, old_pte);
195 * Insert slot number & secondary bit in PTE second half,
196 * clear _PAGE_BUSY and set appropriate HPTE slot bit
197 * Since we have _PAGE_BUSY set on ptep, we can be sure
198 * nobody is undating hidx.
200 hidxp = (unsigned long *)(ptep + PTRS_PER_PTE);
201 rpte.hidx &= ~(0xfUL << (subpg_index << 2));
202 *hidxp = rpte.hidx | (slot << (subpg_index << 2));
203 new_pte = mark_subptegroup_valid(new_pte, subpg_index);
204 new_pte |= _PAGE_HASHPTE;
206 * check __real_pte for details on matching smp_rmb()
209 *ptep = __pte(new_pte & ~_PAGE_BUSY);
213 int __hash_page_64K(unsigned long ea, unsigned long access,
214 unsigned long vsid, pte_t *ptep, unsigned long trap,
215 unsigned long flags, int ssize)
218 unsigned long hpte_group;
219 unsigned long rflags, pa;
220 unsigned long old_pte, new_pte;
221 unsigned long vpn, hash, slot;
222 unsigned long shift = mmu_psize_defs[MMU_PAGE_64K].shift;
225 * atomically mark the linux large page PTE busy and dirty
228 pte_t pte = READ_ONCE(*ptep);
230 old_pte = pte_val(pte);
231 /* If PTE busy, retry the access */
232 if (unlikely(old_pte & _PAGE_BUSY))
234 /* If PTE permissions don't match, take page fault */
235 if (unlikely(access & ~old_pte))
238 * Check if PTE has the cache-inhibit bit set
239 * If so, bail out and refault as a 4k page
241 if (!mmu_has_feature(MMU_FTR_CI_LARGE_PAGE) &&
242 unlikely(old_pte & _PAGE_NO_CACHE))
245 * Try to lock the PTE, add ACCESSED and DIRTY if it was
246 * a write access. Since this is 4K insert of 64K page size
247 * also add _PAGE_COMBO
249 new_pte = old_pte | _PAGE_BUSY | _PAGE_ACCESSED;
250 if (access & _PAGE_RW)
251 new_pte |= _PAGE_DIRTY;
252 } while (old_pte != __cmpxchg_u64((unsigned long *)ptep,
255 rflags = htab_convert_pte_flags(new_pte);
257 if (!cpu_has_feature(CPU_FTR_NOEXECUTE) &&
258 !cpu_has_feature(CPU_FTR_COHERENT_ICACHE))
259 rflags = hash_page_do_lazy_icache(rflags, __pte(old_pte), trap);
261 vpn = hpt_vpn(ea, vsid, ssize);
262 if (unlikely(old_pte & _PAGE_HASHPTE)) {
264 * There MIGHT be an HPTE for this pte
266 hash = hpt_hash(vpn, shift, ssize);
267 if (old_pte & _PAGE_F_SECOND)
269 slot = (hash & htab_hash_mask) * HPTES_PER_GROUP;
270 slot += (old_pte & _PAGE_F_GIX) >> _PAGE_F_GIX_SHIFT;
272 if (ppc_md.hpte_updatepp(slot, rflags, vpn, MMU_PAGE_64K,
273 MMU_PAGE_64K, ssize, flags) == -1)
274 old_pte &= ~_PAGE_HPTEFLAGS;
277 if (likely(!(old_pte & _PAGE_HASHPTE))) {
279 pa = pte_pfn(__pte(old_pte)) << PAGE_SHIFT;
280 hash = hpt_hash(vpn, shift, ssize);
283 hpte_group = ((hash & htab_hash_mask) * HPTES_PER_GROUP) & ~0x7UL;
285 /* Insert into the hash table, primary slot */
286 slot = ppc_md.hpte_insert(hpte_group, vpn, pa, rflags, 0,
287 MMU_PAGE_64K, MMU_PAGE_64K, ssize);
289 * Primary is full, try the secondary
291 if (unlikely(slot == -1)) {
292 hpte_group = ((~hash & htab_hash_mask) * HPTES_PER_GROUP) & ~0x7UL;
293 slot = ppc_md.hpte_insert(hpte_group, vpn, pa,
294 rflags, HPTE_V_SECONDARY,
295 MMU_PAGE_64K, MMU_PAGE_64K, ssize);
298 hpte_group = ((hash & htab_hash_mask) *
299 HPTES_PER_GROUP) & ~0x7UL;
300 ppc_md.hpte_remove(hpte_group);
302 * FIXME!! Should be try the group from which we removed ?
308 * Hypervisor failure. Restore old pmd and return -1
309 * similar to __hash_page_*
311 if (unlikely(slot == -2)) {
312 *ptep = __pte(old_pte);
313 hash_failure_debug(ea, access, vsid, trap, ssize,
314 MMU_PAGE_64K, MMU_PAGE_64K, old_pte);
317 new_pte = (new_pte & ~_PAGE_HPTEFLAGS) | _PAGE_HASHPTE;
318 new_pte |= (slot << _PAGE_F_GIX_SHIFT) & (_PAGE_F_SECOND | _PAGE_F_GIX);
320 *ptep = __pte(new_pte & ~_PAGE_BUSY);