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:
6 * Copyright (C) 1995-1996 Gary Thomas (gdt@linuxppc.org)
8 * Modifications by Paul Mackerras (PowerMac) (paulus@cs.anu.edu.au)
9 * and Cort Dougan (PReP) (cort@cs.nmt.edu)
10 * Copyright (C) 1996 Paul Mackerras
12 * Derived from "arch/i386/mm/init.c"
13 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
15 * Dave Engebretsen <engebret@us.ibm.com>
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/init.h>
27 #include <linux/percpu.h>
28 #include <linux/hardirq.h>
29 #include <asm/pgalloc.h>
30 #include <asm/tlbflush.h>
34 DEFINE_PER_CPU(struct ppc64_tlb_batch, ppc64_tlb_batch);
36 /* This is declared as we are using the more or less generic
37 * include/asm-powerpc/tlb.h file -- tgall
39 DEFINE_PER_CPU(struct mmu_gather, mmu_gathers);
40 DEFINE_PER_CPU(struct pte_freelist_batch *, pte_freelist_cur);
41 unsigned long pte_freelist_forced_free;
43 struct pte_freelist_batch
47 pgtable_free_t tables[0];
50 DEFINE_PER_CPU(struct pte_freelist_batch *, pte_freelist_cur);
51 unsigned long pte_freelist_forced_free;
53 #define PTE_FREELIST_SIZE \
54 ((PAGE_SIZE - sizeof(struct pte_freelist_batch)) \
55 / sizeof(pgtable_free_t))
58 static void pte_free_smp_sync(void *arg)
60 /* Do nothing, just ensure we sync with all CPUs */
64 /* This is only called when we are critically out of memory
65 * (and fail to get a page in pte_free_tlb).
67 static void pgtable_free_now(pgtable_free_t pgf)
69 pte_freelist_forced_free++;
71 smp_call_function(pte_free_smp_sync, NULL, 0, 1);
76 static void pte_free_rcu_callback(struct rcu_head *head)
78 struct pte_freelist_batch *batch =
79 container_of(head, struct pte_freelist_batch, rcu);
82 for (i = 0; i < batch->index; i++)
83 pgtable_free(batch->tables[i]);
85 free_page((unsigned long)batch);
88 static void pte_free_submit(struct pte_freelist_batch *batch)
90 INIT_RCU_HEAD(&batch->rcu);
91 call_rcu(&batch->rcu, pte_free_rcu_callback);
94 void pgtable_free_tlb(struct mmu_gather *tlb, pgtable_free_t pgf)
96 /* This is safe since tlb_gather_mmu has disabled preemption */
97 cpumask_t local_cpumask = cpumask_of_cpu(smp_processor_id());
98 struct pte_freelist_batch **batchp = &__get_cpu_var(pte_freelist_cur);
100 if (atomic_read(&tlb->mm->mm_users) < 2 ||
101 cpus_equal(tlb->mm->cpu_vm_mask, local_cpumask)) {
106 if (*batchp == NULL) {
107 *batchp = (struct pte_freelist_batch *)__get_free_page(GFP_ATOMIC);
108 if (*batchp == NULL) {
109 pgtable_free_now(pgf);
112 (*batchp)->index = 0;
114 (*batchp)->tables[(*batchp)->index++] = pgf;
115 if ((*batchp)->index == PTE_FREELIST_SIZE) {
116 pte_free_submit(*batchp);
122 * A linux PTE was changed and the corresponding hash table entry
123 * neesd to be flushed. This function will either perform the flush
124 * immediately or will batch it up if the current CPU has an active
127 * Must be called from within some kind of spinlock/non-preempt region...
129 void hpte_need_flush(struct mm_struct *mm, unsigned long addr,
130 pte_t *ptep, unsigned long pte, int huge)
132 struct ppc64_tlb_batch *batch = &__get_cpu_var(ppc64_tlb_batch);
133 unsigned long vsid, vaddr;
141 /* We mask the address for the base page size. Huge pages will
142 * have applied their own masking already
146 /* Get page size (maybe move back to caller).
148 * NOTE: when using special 64K mappings in 4K environment like
149 * for SPEs, we obtain the page size from the slice, which thus
150 * must still exist (and thus the VMA not reused) at the time
154 #ifdef CONFIG_HUGETLB_PAGE
155 psize = mmu_huge_psize;
158 psize = pte_pagesize_index(mm, addr, pte); /* shutup gcc */
161 psize = pte_pagesize_index(mm, addr, pte);
163 /* Build full vaddr */
164 if (!is_kernel_addr(addr)) {
165 ssize = user_segment_size(addr);
166 vsid = get_vsid(mm->context.id, addr, ssize);
169 vsid = get_kernel_vsid(addr, mmu_kernel_ssize);
170 ssize = mmu_kernel_ssize;
172 vaddr = hpt_va(addr, vsid, ssize);
173 rpte = __real_pte(__pte(pte), ptep);
176 * Check if we have an active batch on this CPU. If not, just
177 * flush now and return. For now, we don global invalidates
178 * in that case, might be worth testing the mm cpu mask though
179 * and decide to use local invalidates instead...
181 if (!batch->active) {
182 flush_hash_page(vaddr, rpte, psize, ssize, 0);
187 * This can happen when we are in the middle of a TLB batch and
188 * we encounter memory pressure (eg copy_page_range when it tries
189 * to allocate a new pte). If we have to reclaim memory and end
190 * up scanning and resetting referenced bits then our batch context
191 * will change mid stream.
193 * We also need to ensure only one page size is present in a given
196 if (i != 0 && (mm != batch->mm || batch->psize != psize ||
197 batch->ssize != ssize)) {
198 __flush_tlb_pending(batch);
203 batch->psize = psize;
204 batch->ssize = ssize;
206 batch->pte[i] = rpte;
207 batch->vaddr[i] = vaddr;
209 if (i >= PPC64_TLB_BATCH_NR)
210 __flush_tlb_pending(batch);
214 * This function is called when terminating an mmu batch or when a batch
215 * is full. It will perform the flush of all the entries currently stored
218 * Must be called from within some kind of spinlock/non-preempt region...
220 void __flush_tlb_pending(struct ppc64_tlb_batch *batch)
226 tmp = cpumask_of_cpu(smp_processor_id());
227 if (cpus_equal(batch->mm->cpu_vm_mask, tmp))
230 flush_hash_page(batch->vaddr[0], batch->pte[0],
231 batch->psize, batch->ssize, local);
233 flush_hash_range(i, local);
237 void pte_free_finish(void)
239 /* This is safe since tlb_gather_mmu has disabled preemption */
240 struct pte_freelist_batch **batchp = &__get_cpu_var(pte_freelist_cur);
244 pte_free_submit(*batchp);
249 * __flush_hash_table_range - Flush all HPTEs for a given address range
250 * from the hash table (and the TLB). But keeps
251 * the linux PTEs intact.
253 * @mm : mm_struct of the target address space (generally init_mm)
254 * @start : starting address
255 * @end : ending address (not included in the flush)
257 * This function is mostly to be used by some IO hotplug code in order
258 * to remove all hash entries from a given address range used to map IO
259 * space on a removed PCI-PCI bidge without tearing down the full mapping
260 * since 64K pages may overlap with other bridges when using 64K pages
261 * with 4K HW pages on IO space.
263 * Because of that usage pattern, it's only available with CONFIG_HOTPLUG
264 * and is implemented for small size rather than speed.
266 #ifdef CONFIG_HOTPLUG
268 void __flush_hash_table_range(struct mm_struct *mm, unsigned long start,
273 start = _ALIGN_DOWN(start, PAGE_SIZE);
274 end = _ALIGN_UP(end, PAGE_SIZE);
278 /* Note: Normally, we should only ever use a batch within a
279 * PTE locked section. This violates the rule, but will work
280 * since we don't actually modify the PTEs, we just flush the
281 * hash while leaving the PTEs intact (including their reference
282 * to being hashed). This is not the most performance oriented
283 * way to do things but is fine for our needs here.
285 local_irq_save(flags);
286 arch_enter_lazy_mmu_mode();
287 for (; start < end; start += PAGE_SIZE) {
288 pte_t *ptep = find_linux_pte(mm->pgd, start);
293 pte = pte_val(*ptep);
294 if (!(pte & _PAGE_HASHPTE))
296 hpte_need_flush(mm, start, ptep, pte, 0);
298 arch_leave_lazy_mmu_mode();
299 local_irq_restore(flags);
302 #endif /* CONFIG_HOTPLUG */