hash_hugepage.c

  1/*
  2 * Copyright IBM Corporation, 2013
  3 * Author Aneesh Kumar K.V <aneesh.kumar@linux.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
 21int __hash_page_thp(unsigned long ea, unsigned long access, unsigned long vsid,
 22		    pmd_t *pmdp, unsigned long trap, unsigned long flags,
 23		    int ssize, unsigned int psize)
 24{
 25	unsigned int index, valid;
 26	unsigned char *hpte_slot_array;
 27	unsigned long rflags, pa, hidx;
 28	unsigned long old_pmd, new_pmd;
 29	int ret, lpsize = MMU_PAGE_16M;
 30	unsigned long vpn, hash, shift, slot;
 31
 32	/*
 33	 * atomically mark the linux large page PMD busy and dirty
 34	 */
 35	do {
 36		pmd_t pmd = READ_ONCE(*pmdp);
 37
 38		old_pmd = pmd_val(pmd);
 39		/* If PMD busy, retry the access */
 40		if (unlikely(old_pmd & H_PAGE_BUSY))
 41			return 0;
 42		/* If PMD permissions don't match, take page fault */
 43		if (unlikely(!check_pte_access(access, old_pmd)))
 44			return 1;
 45		/*
 46		 * Try to lock the PTE, add ACCESSED and DIRTY if it was
 47		 * a write access
 48		 */
 49		new_pmd = old_pmd | H_PAGE_BUSY | _PAGE_ACCESSED;
 50		if (access & _PAGE_WRITE)
 51			new_pmd |= _PAGE_DIRTY;
 52	} while (!pmd_xchg(pmdp, __pmd(old_pmd), __pmd(new_pmd)));
 53
 54	/*
 55	 * Make sure this is thp or devmap entry
 56	 */
 57	if (!(old_pmd & H_PAGE_THP_HUGE))
 58		return 0;
 59
 60	rflags = htab_convert_pte_flags(new_pmd, flags);
 61
 62	/*
 63	 * THPs are only supported on platforms that can do mixed page size
 64	 * segments (MPSS) and all such platforms have coherent icache. Hence we
 65	 * don't need to do lazy icache flush (hash_page_do_lazy_icache()) on
 66	 * noexecute fault.
 67	 */
 68
 69	/*
 70	 * Find the slot index details for this ea, using base page size.
 71	 */
 72	shift = mmu_psize_defs[psize].shift;
 73	index = (ea & ~HPAGE_PMD_MASK) >> shift;
 74	BUG_ON(index >= PTE_FRAG_SIZE);
 75
 76	vpn = hpt_vpn(ea, vsid, ssize);
 77	hpte_slot_array = get_hpte_slot_array(pmdp);
 78	if (psize == MMU_PAGE_4K) {
 79		/*
 80		 * invalidate the old hpte entry if we have that mapped via 64K
 81		 * base page size. This is because demote_segment won't flush
 82		 * hash page table entries.
 83		 */
 84		if ((old_pmd & H_PAGE_HASHPTE) && !(old_pmd & H_PAGE_COMBO)) {
 85			flush_hash_hugepage(vsid, ea, pmdp, MMU_PAGE_64K,
 86					    ssize, flags);
 87			/*
 88			 * With THP, we also clear the slot information with
 89			 * respect to all the 64K hash pte mapping the 16MB
 90			 * page. They are all invalid now. This make sure we
 91			 * don't find the slot valid when we fault with 4k
 92			 * base page size.
 93			 *
 94			 */
 95			memset(hpte_slot_array, 0, PTE_FRAG_SIZE);
 96		}
 97	}
 98
 99	valid = hpte_valid(hpte_slot_array, index);
100	if (valid) {
101		/* update the hpte bits */
102		hash = hpt_hash(vpn, shift, ssize);
103		hidx =  hpte_hash_index(hpte_slot_array, index);
104		if (hidx & _PTEIDX_SECONDARY)
105			hash = ~hash;
106		slot = (hash & htab_hash_mask) * HPTES_PER_GROUP;
107		slot += hidx & _PTEIDX_GROUP_IX;
108
109		ret = mmu_hash_ops.hpte_updatepp(slot, rflags, vpn,
110						 psize, lpsize, ssize, flags);
111		/*
112		 * We failed to update, try to insert a new entry.
113		 */
114		if (ret == -1) {
115			/*
116			 * large pte is marked busy, so we can be sure
117			 * nobody is looking at hpte_slot_array. hence we can
118			 * safely update this here.
119			 */
120			valid = 0;
121			hpte_slot_array[index] = 0;
122		}
123	}
124
125	if (!valid) {
126		unsigned long hpte_group;
127
128		hash = hpt_hash(vpn, shift, ssize);
129		/* insert new entry */
130		pa = pmd_pfn(__pmd(old_pmd)) << PAGE_SHIFT;
131		new_pmd |= H_PAGE_HASHPTE;
132
133repeat:
134		hpte_group = (hash & htab_hash_mask) * HPTES_PER_GROUP;
135
136		/* Insert into the hash table, primary slot */
137		slot = mmu_hash_ops.hpte_insert(hpte_group, vpn, pa, rflags, 0,
138						psize, lpsize, ssize);
139		/*
140		 * Primary is full, try the secondary
141		 */
142		if (unlikely(slot == -1)) {
143			hpte_group = (~hash & htab_hash_mask) * HPTES_PER_GROUP;
144			slot = mmu_hash_ops.hpte_insert(hpte_group, vpn, pa,
145							rflags,
146							HPTE_V_SECONDARY,
147							psize, lpsize, ssize);
148			if (slot == -1) {
149				if (mftb() & 0x1)
150					hpte_group = (hash & htab_hash_mask) *
151							HPTES_PER_GROUP;
152
153				mmu_hash_ops.hpte_remove(hpte_group);
154				goto repeat;
155			}
156		}
157		/*
158		 * Hypervisor failure. Restore old pmd and return -1
159		 * similar to __hash_page_*
160		 */
161		if (unlikely(slot == -2)) {
162			*pmdp = __pmd(old_pmd);
163			hash_failure_debug(ea, access, vsid, trap, ssize,
164					   psize, lpsize, old_pmd);
165			return -1;
166		}
167		/*
168		 * large pte is marked busy, so we can be sure
169		 * nobody is looking at hpte_slot_array. hence we can
170		 * safely update this here.
171		 */
172		mark_hpte_slot_valid(hpte_slot_array, index, slot);
173	}
174	/*
175	 * Mark the pte with H_PAGE_COMBO, if we are trying to hash it with
176	 * base page size 4k.
177	 */
178	if (psize == MMU_PAGE_4K)
179		new_pmd |= H_PAGE_COMBO;
180	/*
181	 * The hpte valid is stored in the pgtable whose address is in the
182	 * second half of the PMD. Order this against clearing of the busy bit in
183	 * huge pmd.
184	 */
185	smp_wmb();
186	*pmdp = __pmd(new_pmd & ~H_PAGE_BUSY);
187	return 0;
188}