mm/book3s64/hash_hugepage.c

0001 /*
0002  * Copyright IBM Corporation, 2013
0003  * Author Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
0004  *
0005  * This program is free software; you can redistribute it and/or modify it
0006  * under the terms of version 2.1 of the GNU Lesser General Public License
0007  * as published by the Free Software Foundation.
0008  *
0009  * This program is distributed in the hope that it would be useful, but
0010  * WITHOUT ANY WARRANTY; without even the implied warranty of
0011  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.
0012  *
0013  */
0014
0015 /*
0016  * PPC64 THP Support for hash based MMUs
0017  */
0018 #include <linux/mm.h>
0019 #include <asm/machdep.h>
0020
0021 int __hash_page_thp(unsigned long ea, unsigned long access, unsigned long vsid,
0022             pmd_t *pmdp, unsigned long trap, unsigned long flags,
0023             int ssize, unsigned int psize)
0024 {
0025     unsigned int index, valid;
0026     unsigned char *hpte_slot_array;
0027     unsigned long rflags, pa, hidx;
0028     unsigned long old_pmd, new_pmd;
0029     int ret, lpsize = MMU_PAGE_16M;
0030     unsigned long vpn, hash, shift, slot;
0031
0032     /*
0033      * atomically mark the linux large page PMD busy and dirty
0034      */
0035     do {
0036         pmd_t pmd = READ_ONCE(*pmdp);
0037
0038         old_pmd = pmd_val(pmd);
0039         /* If PMD busy, retry the access */
0040         if (unlikely(old_pmd & H_PAGE_BUSY))
0041             return 0;
0042         /* If PMD permissions don't match, take page fault */
0043         if (unlikely(!check_pte_access(access, old_pmd)))
0044             return 1;
0045         /*
0046          * Try to lock the PTE, add ACCESSED and DIRTY if it was
0047          * a write access
0048          */
0049         new_pmd = old_pmd | H_PAGE_BUSY | _PAGE_ACCESSED;
0050         if (access & _PAGE_WRITE)
0051             new_pmd |= _PAGE_DIRTY;
0052     } while (!pmd_xchg(pmdp, __pmd(old_pmd), __pmd(new_pmd)));
0053
0054     /*
0055      * Make sure this is thp or devmap entry
0056      */
0057     if (!(old_pmd & (H_PAGE_THP_HUGE | _PAGE_DEVMAP)))
0058         return 0;
0059
0060     rflags = htab_convert_pte_flags(new_pmd, flags);
0061
0062 #if 0
0063     if (!cpu_has_feature(CPU_FTR_COHERENT_ICACHE)) {
0064
0065         /*
0066          * No CPU has hugepages but lacks no execute, so we
0067          * don't need to worry about that case
0068          */
0069         rflags = hash_page_do_lazy_icache(rflags, __pte(old_pte), trap);
0070     }
0071 #endif
0072     /*
0073      * Find the slot index details for this ea, using base page size.
0074      */
0075     shift = mmu_psize_defs[psize].shift;
0076     index = (ea & ~HPAGE_PMD_MASK) >> shift;
0077     BUG_ON(index >= PTE_FRAG_SIZE);
0078
0079     vpn = hpt_vpn(ea, vsid, ssize);
0080     hpte_slot_array = get_hpte_slot_array(pmdp);
0081     if (psize == MMU_PAGE_4K) {
0082         /*
0083          * invalidate the old hpte entry if we have that mapped via 64K
0084          * base page size. This is because demote_segment won't flush
0085          * hash page table entries.
0086          */
0087         if ((old_pmd & H_PAGE_HASHPTE) && !(old_pmd & H_PAGE_COMBO)) {
0088             flush_hash_hugepage(vsid, ea, pmdp, MMU_PAGE_64K,
0089                         ssize, flags);
0090             /*
0091              * With THP, we also clear the slot information with
0092              * respect to all the 64K hash pte mapping the 16MB
0093              * page. They are all invalid now. This make sure we
0094              * don't find the slot valid when we fault with 4k
0095              * base page size.
0096              *
0097              */
0098             memset(hpte_slot_array, 0, PTE_FRAG_SIZE);
0099         }
0100     }
0101
0102     valid = hpte_valid(hpte_slot_array, index);
0103     if (valid) {
0104         /* update the hpte bits */
0105         hash = hpt_hash(vpn, shift, ssize);
0106         hidx =  hpte_hash_index(hpte_slot_array, index);
0107         if (hidx & _PTEIDX_SECONDARY)
0108             hash = ~hash;
0109         slot = (hash & htab_hash_mask) * HPTES_PER_GROUP;
0110         slot += hidx & _PTEIDX_GROUP_IX;
0111
0112         ret = mmu_hash_ops.hpte_updatepp(slot, rflags, vpn,
0113                          psize, lpsize, ssize, flags);
0114         /*
0115          * We failed to update, try to insert a new entry.
0116          */
0117         if (ret == -1) {
0118             /*
0119              * large pte is marked busy, so we can be sure
0120              * nobody is looking at hpte_slot_array. hence we can
0121              * safely update this here.
0122              */
0123             valid = 0;
0124             hpte_slot_array[index] = 0;
0125         }
0126     }
0127
0128     if (!valid) {
0129         unsigned long hpte_group;
0130
0131         hash = hpt_hash(vpn, shift, ssize);
0132         /* insert new entry */
0133         pa = pmd_pfn(__pmd(old_pmd)) << PAGE_SHIFT;
0134         new_pmd |= H_PAGE_HASHPTE;
0135
0136 repeat:
0137         hpte_group = (hash & htab_hash_mask) * HPTES_PER_GROUP;
0138
0139         /* Insert into the hash table, primary slot */
0140         slot = mmu_hash_ops.hpte_insert(hpte_group, vpn, pa, rflags, 0,
0141                         psize, lpsize, ssize);
0142         /*
0143          * Primary is full, try the secondary
0144          */
0145         if (unlikely(slot == -1)) {
0146             hpte_group = (~hash & htab_hash_mask) * HPTES_PER_GROUP;
0147             slot = mmu_hash_ops.hpte_insert(hpte_group, vpn, pa,
0148                             rflags,
0149                             HPTE_V_SECONDARY,
0150                             psize, lpsize, ssize);
0151             if (slot == -1) {
0152                 if (mftb() & 0x1)
0153                     hpte_group = (hash & htab_hash_mask) *
0154                             HPTES_PER_GROUP;
0155
0156                 mmu_hash_ops.hpte_remove(hpte_group);
0157                 goto repeat;
0158             }
0159         }
0160         /*
0161          * Hypervisor failure. Restore old pmd and return -1
0162          * similar to __hash_page_*
0163          */
0164         if (unlikely(slot == -2)) {
0165             *pmdp = __pmd(old_pmd);
0166             hash_failure_debug(ea, access, vsid, trap, ssize,
0167                        psize, lpsize, old_pmd);
0168             return -1;
0169         }
0170         /*
0171          * large pte is marked busy, so we can be sure
0172          * nobody is looking at hpte_slot_array. hence we can
0173          * safely update this here.
0174          */
0175         mark_hpte_slot_valid(hpte_slot_array, index, slot);
0176     }
0177     /*
0178      * Mark the pte with H_PAGE_COMBO, if we are trying to hash it with
0179      * base page size 4k.
0180      */
0181     if (psize == MMU_PAGE_4K)
0182         new_pmd |= H_PAGE_COMBO;
0183     /*
0184      * The hpte valid is stored in the pgtable whose address is in the
0185      * second half of the PMD. Order this against clearing of the busy bit in
0186      * huge pmd.
0187      */
0188     smp_wmb();
0189     *pmdp = __pmd(new_pmd & ~H_PAGE_BUSY);
0190     return 0;
0191 }