2 * This file contains the routines for flushing entries from the
3 * TLB and MMU hash table.
5 * Derived from arch/ppc64/mm/init.c:
10 * Copyright (C) 1996 Paul Mackerras
12 * Derived from "arch/i386/mm/init.c"
13 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
16 * Rework for PPC64 port.
18 * This program is free software; you can redistribute it and/or
19 * modify it under the terms of the GNU General Public License
20 * as published by the Free Software Foundation; either version
21 * 2 of the License, or (at your option) any later version.
24 #include <linux/kernel.h>
26 #include <linux/percpu.h>
27 #include <linux/hardirq.h>
28 #include <asm/pgalloc.h>
29 #include <asm/tlbflush.h>
32 #include <asm/pte-walk.h>
35 #include <trace/events/thp.h>
37 DEFINE_PER_CPU(struct ppc64_tlb_batch, ppc64_tlb_batch);
40 * A linux PTE was changed and the corresponding hash table entry
41 * neesd to be flushed. This function will either perform the flush
42 * immediately or will batch it up if the current CPU has an active
45 void hpte_need_flush(struct mm_struct *mm, unsigned long addr,
46 pte_t *ptep, unsigned long pte, int huge)
49 struct ppc64_tlb_batch *batch = &get_cpu_var(ppc64_tlb_batch);
58 /* Get page size (maybe move back to caller).
60 * NOTE: when using special 64K mappings in 4K environment like
61 * for SPEs, we obtain the page size from the slice, which thus
62 * must still exist (and thus the VMA not reused) at the time
66 #ifdef CONFIG_HUGETLB_PAGE
67 psize = get_slice_psize(mm, addr);
68 /* Mask the address for the correct page size */
69 addr &= ~((1UL << mmu_psize_defs[psize].shift) - 1);
70 if (unlikely(psize == MMU_PAGE_16G))
71 offset = PTRS_PER_PUD;
73 offset = PTRS_PER_PMD;
76 psize = pte_pagesize_index(mm, addr, pte); /* shutup gcc */
79 psize = pte_pagesize_index(mm, addr, pte);
80 /* Mask the address for the standard page size. If we
81 * have a 64k page kernel, but the hardware does not
82 * support 64k pages, this might be different from the
83 * hardware page size encoded in the slice table. */
85 offset = PTRS_PER_PTE;
89 /* Build full vaddr */
90 if (!is_kernel_addr(addr)) {
91 ssize = user_segment_size(addr);
92 vsid = get_user_vsid(&mm->context, addr, ssize);
94 vsid = get_kernel_vsid(addr, mmu_kernel_ssize);
95 ssize = mmu_kernel_ssize;
98 vpn = hpt_vpn(addr, vsid, ssize);
99 rpte = __real_pte(__pte(pte), ptep, offset);
102 * Check if we have an active batch on this CPU. If not, just
103 * flush now and return.
105 if (!batch->active) {
106 flush_hash_page(vpn, rpte, psize, ssize, mm_is_thread_local(mm));
107 put_cpu_var(ppc64_tlb_batch);
112 * This can happen when we are in the middle of a TLB batch and
113 * we encounter memory pressure (eg copy_page_range when it tries
114 * to allocate a new pte). If we have to reclaim memory and end
115 * up scanning and resetting referenced bits then our batch context
116 * will change mid stream.
118 * We also need to ensure only one page size is present in a given
121 if (i != 0 && (mm != batch->mm || batch->psize != psize ||
122 batch->ssize != ssize)) {
123 __flush_tlb_pending(batch);
128 batch->psize = psize;
129 batch->ssize = ssize;
131 batch->pte[i] = rpte;
134 if (i >= PPC64_TLB_BATCH_NR)
135 __flush_tlb_pending(batch);
136 put_cpu_var(ppc64_tlb_batch);
140 * This function is called when terminating an mmu batch or when a batch
141 * is full. It will perform the flush of all the entries currently stored
144 * Must be called from within some kind of spinlock/non-preempt region...
146 void __flush_tlb_pending(struct ppc64_tlb_batch *batch)
151 local = mm_is_thread_local(batch->mm);
153 flush_hash_page(batch->vpn[0], batch->pte[0],
154 batch->psize, batch->ssize, local);
156 flush_hash_range(i, local);
160 void hash__tlb_flush(struct mmu_gather *tlb)
162 struct ppc64_tlb_batch *tlbbatch = &get_cpu_var(ppc64_tlb_batch);
164 /* If there's a TLB batch pending, then we must flush it because the
165 * pages are going to be freed and we really don't want to have a CPU
166 * access a freed page because it has a stale TLB
169 __flush_tlb_pending(tlbbatch);
171 put_cpu_var(ppc64_tlb_batch);
175 * __flush_hash_table_range - Flush all HPTEs for a given address range
176 * from the hash table (and the TLB). But keeps
177 * the linux PTEs intact.
179 * @mm : mm_struct of the target address space (generally init_mm)
180 * @start : starting address
181 * @end : ending address (not included in the flush)
183 * This function is mostly to be used by some IO hotplug code in order
184 * to remove all hash entries from a given address range used to map IO
185 * space on a removed PCI-PCI bidge without tearing down the full mapping
186 * since 64K pages may overlap with other bridges when using 64K pages
187 * with 4K HW pages on IO space.
189 * Because of that usage pattern, it is implemented for small size rather
192 void __flush_hash_table_range(struct mm_struct *mm, unsigned long start,
199 start = _ALIGN_DOWN(start, PAGE_SIZE);
200 end = _ALIGN_UP(end, PAGE_SIZE);
204 /* Note: Normally, we should only ever use a batch within a
205 * PTE locked section. This violates the rule, but will work
206 * since we don't actually modify the PTEs, we just flush the
207 * hash while leaving the PTEs intact (including their reference
208 * to being hashed). This is not the most performance oriented
209 * way to do things but is fine for our needs here.
211 local_irq_save(flags);
212 arch_enter_lazy_mmu_mode();
213 for (; start < end; start += PAGE_SIZE) {
214 pte_t *ptep = find_current_mm_pte(mm->pgd, start, &is_thp,
220 pte = pte_val(*ptep);
222 trace_hugepage_invalidate(start, pte);
223 if (!(pte & H_PAGE_HASHPTE))
225 if (unlikely(is_thp))
226 hpte_do_hugepage_flush(mm, start, (pmd_t *)ptep, pte);
228 hpte_need_flush(mm, start, ptep, pte, hugepage_shift);
230 arch_leave_lazy_mmu_mode();
231 local_irq_restore(flags);
234 void flush_tlb_pmd_range(struct mm_struct *mm, pmd_t *pmd, unsigned long addr)
240 addr = _ALIGN_DOWN(addr, PMD_SIZE);
241 /* Note: Normally, we should only ever use a batch within a
242 * PTE locked section. This violates the rule, but will work
243 * since we don't actually modify the PTEs, we just flush the
244 * hash while leaving the PTEs intact (including their reference
245 * to being hashed). This is not the most performance oriented
246 * way to do things but is fine for our needs here.
248 local_irq_save(flags);
249 arch_enter_lazy_mmu_mode();
250 start_pte = pte_offset_map(pmd, addr);
251 for (pte = start_pte; pte < start_pte + PTRS_PER_PTE; pte++) {
252 unsigned long pteval = pte_val(*pte);
253 if (pteval & H_PAGE_HASHPTE)
254 hpte_need_flush(mm, addr, pte, pteval, 0);
257 arch_leave_lazy_mmu_mode();
258 local_irq_restore(flags);