2 * MMU context allocation for 64-bit kernels.
4 * Copyright (C) 2004 Anton Blanchard, IBM Corp. <anton@samba.org>
6 * This program is free software; you can redistribute it and/or
7 * modify it under the terms of the GNU General Public License
8 * as published by the Free Software Foundation; either version
9 * 2 of the License, or (at your option) any later version.
13 #include <linux/sched.h>
14 #include <linux/kernel.h>
15 #include <linux/errno.h>
16 #include <linux/string.h>
17 #include <linux/types.h>
19 #include <linux/pkeys.h>
20 #include <linux/spinlock.h>
21 #include <linux/idr.h>
22 #include <linux/export.h>
23 #include <linux/gfp.h>
24 #include <linux/slab.h>
26 #include <asm/mmu_context.h>
27 #include <asm/pgalloc.h>
29 static DEFINE_SPINLOCK(mmu_context_lock);
30 static DEFINE_IDA(mmu_context_ida);
32 static int alloc_context_id(int min_id, int max_id)
37 if (!ida_pre_get(&mmu_context_ida, GFP_KERNEL))
40 spin_lock(&mmu_context_lock);
41 err = ida_get_new_above(&mmu_context_ida, min_id, &index);
42 spin_unlock(&mmu_context_lock);
50 spin_lock(&mmu_context_lock);
51 ida_remove(&mmu_context_ida, index);
52 spin_unlock(&mmu_context_lock);
59 void hash__reserve_context_id(int id)
64 if (!ida_pre_get(&mmu_context_ida, GFP_KERNEL))
67 spin_lock(&mmu_context_lock);
68 rc = ida_get_new_above(&mmu_context_ida, id, &result);
69 spin_unlock(&mmu_context_lock);
70 } while (rc == -EAGAIN);
72 WARN(result != id, "mmu: Failed to reserve context id %d (rc %d)\n", id, result);
75 int hash__alloc_context_id(void)
79 if (mmu_has_feature(MMU_FTR_68_BIT_VA))
80 max = MAX_USER_CONTEXT;
82 max = MAX_USER_CONTEXT_65BIT_VA;
84 return alloc_context_id(MIN_USER_CONTEXT, max);
86 EXPORT_SYMBOL_GPL(hash__alloc_context_id);
88 static int hash__init_new_context(struct mm_struct *mm)
92 index = hash__alloc_context_id();
97 * In the case of exec, use the default limit,
98 * otherwise inherit it from the mm we are duplicating.
100 if (!mm->context.slb_addr_limit)
101 mm->context.slb_addr_limit = DEFAULT_MAP_WINDOW_USER64;
104 * The old code would re-promote on fork, we don't do that when using
105 * slices as it could cause problem promoting slices that have been
108 * For book3s we have MMU_NO_CONTEXT set to be ~0. Hence check
109 * explicitly against context.id == 0. This ensures that we properly
110 * initialize context slice details for newly allocated mm's (which will
111 * have id == 0) and don't alter context slice inherited via fork (which
112 * will have id != 0).
114 * We should not be calling init_new_context() on init_mm. Hence a
115 * check against 0 is OK.
117 if (mm->context.id == 0)
118 slice_set_user_psize(mm, mmu_virtual_psize);
120 subpage_prot_init_new_context(mm);
126 static int radix__init_new_context(struct mm_struct *mm)
128 unsigned long rts_field;
131 max_id = (1 << mmu_pid_bits) - 1;
132 index = alloc_context_id(mmu_base_pid, max_id);
137 * set the process table entry,
139 rts_field = radix__get_tree_size();
140 process_tb[index].prtb0 = cpu_to_be64(rts_field | __pa(mm->pgd) | RADIX_PGD_INDEX_SIZE);
143 * Order the above store with subsequent update of the PID
144 * register (at which point HW can start loading/caching
145 * the entry) and the corresponding load by the MMU from
148 asm volatile("ptesync;isync" : : : "memory");
150 mm->context.npu_context = NULL;
155 int init_new_context(struct task_struct *tsk, struct mm_struct *mm)
160 index = radix__init_new_context(mm);
162 index = hash__init_new_context(mm);
167 mm->context.id = index;
169 #ifdef CONFIG_PPC_64K_PAGES
170 mm->context.pte_frag = NULL;
172 #ifdef CONFIG_SPAPR_TCE_IOMMU
175 atomic_set(&mm->context.active_cpus, 0);
176 atomic_set(&mm->context.copros, 0);
181 void __destroy_context(int context_id)
183 spin_lock(&mmu_context_lock);
184 ida_remove(&mmu_context_ida, context_id);
185 spin_unlock(&mmu_context_lock);
187 EXPORT_SYMBOL_GPL(__destroy_context);
189 #ifdef CONFIG_PPC_64K_PAGES
190 static void destroy_pagetable_page(struct mm_struct *mm)
196 pte_frag = mm->context.pte_frag;
200 page = virt_to_page(pte_frag);
201 /* drop all the pending references */
202 count = ((unsigned long)pte_frag & ~PAGE_MASK) >> PTE_FRAG_SIZE_SHIFT;
203 /* We allow PTE_FRAG_NR fragments from a PTE page */
204 if (page_ref_sub_and_test(page, PTE_FRAG_NR - count)) {
205 pgtable_page_dtor(page);
206 free_unref_page(page);
211 static inline void destroy_pagetable_page(struct mm_struct *mm)
217 void destroy_context(struct mm_struct *mm)
219 #ifdef CONFIG_SPAPR_TCE_IOMMU
220 WARN_ON_ONCE(!list_empty(&mm->context.iommu_group_mem_list));
223 WARN_ON(process_tb[mm->context.id].prtb0 != 0);
225 subpage_prot_free(mm);
226 destroy_pagetable_page(mm);
227 __destroy_context(mm->context.id);
228 mm->context.id = MMU_NO_CONTEXT;
231 void arch_exit_mmap(struct mm_struct *mm)
233 if (radix_enabled()) {
235 * Radix doesn't have a valid bit in the process table
236 * entries. However we know that at least P9 implementation
237 * will avoid caching an entry with an invalid RTS field,
238 * and 0 is invalid. So this will do.
240 * This runs before the "fullmm" tlb flush in exit_mmap,
241 * which does a RIC=2 tlbie to clear the process table
242 * entry. See the "fullmm" comments in tlb-radix.c.
244 * No barrier required here after the store because
245 * this process will do the invalidate, which starts with
248 process_tb[mm->context.id].prtb0 = 0;
252 #ifdef CONFIG_PPC_RADIX_MMU
253 void radix__switch_mmu_context(struct mm_struct *prev, struct mm_struct *next)
256 if (cpu_has_feature(CPU_FTR_POWER9_DD1)) {
258 mtspr(SPRN_PID, next->context.id);
260 asm volatile(PPC_INVALIDATE_ERAT : : :"memory");
262 mtspr(SPRN_PID, next->context.id);