2 * Page table allocation functions
4 * Copyright IBM Corp. 2016
5 * Author(s): Martin Schwidefsky <schwidefsky@de.ibm.com>
9 #include <linux/sysctl.h>
10 #include <asm/mmu_context.h>
11 #include <asm/pgalloc.h>
14 #include <asm/tlbflush.h>
18 static int page_table_allocate_pgste_min = 0;
19 static int page_table_allocate_pgste_max = 1;
20 int page_table_allocate_pgste = 0;
21 EXPORT_SYMBOL(page_table_allocate_pgste);
23 static struct ctl_table page_table_sysctl[] = {
25 .procname = "allocate_pgste",
26 .data = &page_table_allocate_pgste,
27 .maxlen = sizeof(int),
28 .mode = S_IRUGO | S_IWUSR,
29 .proc_handler = proc_dointvec,
30 .extra1 = &page_table_allocate_pgste_min,
31 .extra2 = &page_table_allocate_pgste_max,
36 static struct ctl_table page_table_sysctl_dir[] = {
41 .child = page_table_sysctl,
46 static int __init page_table_register_sysctl(void)
48 return register_sysctl_table(page_table_sysctl_dir) ? 0 : -ENOMEM;
50 __initcall(page_table_register_sysctl);
52 #endif /* CONFIG_PGSTE */
54 unsigned long *crst_table_alloc(struct mm_struct *mm)
56 struct page *page = alloc_pages(GFP_KERNEL, 2);
60 return (unsigned long *) page_to_phys(page);
63 void crst_table_free(struct mm_struct *mm, unsigned long *table)
65 free_pages((unsigned long) table, 2);
68 static void __crst_table_upgrade(void *arg)
70 struct mm_struct *mm = arg;
72 if (current->active_mm == mm) {
79 int crst_table_upgrade(struct mm_struct *mm, unsigned long limit)
81 unsigned long *table, *pgd;
85 BUG_ON(limit > TASK_MAX_SIZE);
88 table = crst_table_alloc(mm);
91 spin_lock_bh(&mm->page_table_lock);
92 if (mm->context.asce_limit < limit) {
93 pgd = (unsigned long *) mm->pgd;
94 if (mm->context.asce_limit <= (1UL << 31)) {
95 entry = _REGION3_ENTRY_EMPTY;
96 mm->context.asce_limit = 1UL << 42;
97 mm->context.asce_bits = _ASCE_TABLE_LENGTH |
101 entry = _REGION2_ENTRY_EMPTY;
102 mm->context.asce_limit = 1UL << 53;
103 mm->context.asce_bits = _ASCE_TABLE_LENGTH |
107 crst_table_init(table, entry);
108 pgd_populate(mm, (pgd_t *) table, (pud_t *) pgd);
109 mm->pgd = (pgd_t *) table;
110 mm->task_size = mm->context.asce_limit;
114 spin_unlock_bh(&mm->page_table_lock);
116 crst_table_free(mm, table);
117 if (mm->context.asce_limit < limit)
120 on_each_cpu(__crst_table_upgrade, mm, 0);
124 void crst_table_downgrade(struct mm_struct *mm, unsigned long limit)
128 if (current->active_mm == mm) {
132 while (mm->context.asce_limit > limit) {
134 switch (pgd_val(*pgd) & _REGION_ENTRY_TYPE_MASK) {
135 case _REGION_ENTRY_TYPE_R2:
136 mm->context.asce_limit = 1UL << 42;
137 mm->context.asce_bits = _ASCE_TABLE_LENGTH |
141 case _REGION_ENTRY_TYPE_R3:
142 mm->context.asce_limit = 1UL << 31;
143 mm->context.asce_bits = _ASCE_TABLE_LENGTH |
150 mm->pgd = (pgd_t *) (pgd_val(*pgd) & _REGION_ENTRY_ORIGIN);
151 mm->task_size = mm->context.asce_limit;
152 crst_table_free(mm, (unsigned long *) pgd);
154 if (current->active_mm == mm)
158 static inline unsigned int atomic_xor_bits(atomic_t *v, unsigned int bits)
160 unsigned int old, new;
163 old = atomic_read(v);
165 } while (atomic_cmpxchg(v, old, new) != old);
170 * page table entry allocation/free routines.
172 unsigned long *page_table_alloc(struct mm_struct *mm)
174 unsigned long *table;
176 unsigned int mask, bit;
178 /* Try to get a fragment of a 4K page as a 2K page table */
179 if (!mm_alloc_pgste(mm)) {
181 spin_lock_bh(&mm->context.list_lock);
182 if (!list_empty(&mm->context.pgtable_list)) {
183 page = list_first_entry(&mm->context.pgtable_list,
185 mask = atomic_read(&page->_mapcount);
186 mask = (mask | (mask >> 4)) & 3;
188 table = (unsigned long *) page_to_phys(page);
189 bit = mask & 1; /* =1 -> second 2K */
191 table += PTRS_PER_PTE;
192 atomic_xor_bits(&page->_mapcount, 1U << bit);
193 list_del(&page->lru);
196 spin_unlock_bh(&mm->context.list_lock);
200 /* Allocate a fresh page */
201 page = alloc_page(GFP_KERNEL|__GFP_REPEAT);
204 if (!pgtable_page_ctor(page)) {
208 /* Initialize page table */
209 table = (unsigned long *) page_to_phys(page);
210 if (mm_alloc_pgste(mm)) {
211 /* Return 4K page table with PGSTEs */
212 atomic_set(&page->_mapcount, 3);
213 clear_table(table, _PAGE_INVALID, PAGE_SIZE/2);
214 clear_table(table + PTRS_PER_PTE, 0, PAGE_SIZE/2);
216 /* Return the first 2K fragment of the page */
217 atomic_set(&page->_mapcount, 1);
218 clear_table(table, _PAGE_INVALID, PAGE_SIZE);
219 spin_lock_bh(&mm->context.list_lock);
220 list_add(&page->lru, &mm->context.pgtable_list);
221 spin_unlock_bh(&mm->context.list_lock);
226 void page_table_free(struct mm_struct *mm, unsigned long *table)
229 unsigned int bit, mask;
231 page = pfn_to_page(__pa(table) >> PAGE_SHIFT);
232 if (!mm_alloc_pgste(mm)) {
233 /* Free 2K page table fragment of a 4K page */
234 bit = (__pa(table) & ~PAGE_MASK)/(PTRS_PER_PTE*sizeof(pte_t));
235 spin_lock_bh(&mm->context.list_lock);
236 mask = atomic_xor_bits(&page->_mapcount, 1U << bit);
238 list_add(&page->lru, &mm->context.pgtable_list);
240 list_del(&page->lru);
241 spin_unlock_bh(&mm->context.list_lock);
246 pgtable_page_dtor(page);
247 atomic_set(&page->_mapcount, -1);
251 void page_table_free_rcu(struct mmu_gather *tlb, unsigned long *table,
252 unsigned long vmaddr)
254 struct mm_struct *mm;
256 unsigned int bit, mask;
259 page = pfn_to_page(__pa(table) >> PAGE_SHIFT);
260 if (mm_alloc_pgste(mm)) {
261 gmap_unlink(mm, table, vmaddr);
262 table = (unsigned long *) (__pa(table) | 3);
263 tlb_remove_table(tlb, table);
266 bit = (__pa(table) & ~PAGE_MASK) / (PTRS_PER_PTE*sizeof(pte_t));
267 spin_lock_bh(&mm->context.list_lock);
268 mask = atomic_xor_bits(&page->_mapcount, 0x11U << bit);
270 list_add_tail(&page->lru, &mm->context.pgtable_list);
272 list_del(&page->lru);
273 spin_unlock_bh(&mm->context.list_lock);
274 table = (unsigned long *) (__pa(table) | (1U << bit));
275 tlb_remove_table(tlb, table);
278 static void __tlb_remove_table(void *_table)
280 unsigned int mask = (unsigned long) _table & 3;
281 void *table = (void *)((unsigned long) _table ^ mask);
282 struct page *page = pfn_to_page(__pa(table) >> PAGE_SHIFT);
285 case 0: /* pmd or pud */
286 free_pages((unsigned long) table, 2);
288 case 1: /* lower 2K of a 4K page table */
289 case 2: /* higher 2K of a 4K page table */
290 if (atomic_xor_bits(&page->_mapcount, mask << 4) != 0)
293 case 3: /* 4K page table with pgstes */
294 pgtable_page_dtor(page);
295 atomic_set(&page->_mapcount, -1);
301 static void tlb_remove_table_smp_sync(void *arg)
303 /* Simply deliver the interrupt */
306 static void tlb_remove_table_one(void *table)
309 * This isn't an RCU grace period and hence the page-tables cannot be
310 * assumed to be actually RCU-freed.
312 * It is however sufficient for software page-table walkers that rely
313 * on IRQ disabling. See the comment near struct mmu_table_batch.
315 smp_call_function(tlb_remove_table_smp_sync, NULL, 1);
316 __tlb_remove_table(table);
319 static void tlb_remove_table_rcu(struct rcu_head *head)
321 struct mmu_table_batch *batch;
324 batch = container_of(head, struct mmu_table_batch, rcu);
326 for (i = 0; i < batch->nr; i++)
327 __tlb_remove_table(batch->tables[i]);
329 free_page((unsigned long)batch);
332 void tlb_table_flush(struct mmu_gather *tlb)
334 struct mmu_table_batch **batch = &tlb->batch;
337 call_rcu_sched(&(*batch)->rcu, tlb_remove_table_rcu);
342 void tlb_remove_table(struct mmu_gather *tlb, void *table)
344 struct mmu_table_batch **batch = &tlb->batch;
346 tlb->mm->context.flush_mm = 1;
347 if (*batch == NULL) {
348 *batch = (struct mmu_table_batch *)
349 __get_free_page(GFP_NOWAIT | __GFP_NOWARN);
350 if (*batch == NULL) {
351 __tlb_flush_mm_lazy(tlb->mm);
352 tlb_remove_table_one(table);
357 (*batch)->tables[(*batch)->nr++] = table;
358 if ((*batch)->nr == MAX_TABLE_BATCH)