1 /* 2 * Page table allocation functions 3 * 4 * Copyright IBM Corp. 2016 5 * Author(s): Martin Schwidefsky <schwidefsky@de.ibm.com> 6 */ 7 8 #include <linux/mm.h> 9 #include <linux/sysctl.h> 10 #include <asm/mmu_context.h> 11 #include <asm/pgalloc.h> 12 #include <asm/gmap.h> 13 #include <asm/tlb.h> 14 #include <asm/tlbflush.h> 15 16 #ifdef CONFIG_PGSTE 17 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); 22 23 static struct ctl_table page_table_sysctl[] = { 24 { 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, 32 }, 33 { } 34 }; 35 36 static struct ctl_table page_table_sysctl_dir[] = { 37 { 38 .procname = "vm", 39 .maxlen = 0, 40 .mode = 0555, 41 .child = page_table_sysctl, 42 }, 43 { } 44 }; 45 46 static int __init page_table_register_sysctl(void) 47 { 48 return register_sysctl_table(page_table_sysctl_dir) ? 0 : -ENOMEM; 49 } 50 __initcall(page_table_register_sysctl); 51 52 #endif /* CONFIG_PGSTE */ 53 54 unsigned long *crst_table_alloc(struct mm_struct *mm) 55 { 56 struct page *page = alloc_pages(GFP_KERNEL, 2); 57 58 if (!page) 59 return NULL; 60 return (unsigned long *) page_to_phys(page); 61 } 62 63 void crst_table_free(struct mm_struct *mm, unsigned long *table) 64 { 65 free_pages((unsigned long) table, 2); 66 } 67 68 static void __crst_table_upgrade(void *arg) 69 { 70 struct mm_struct *mm = arg; 71 72 if (current->active_mm == mm) { 73 clear_user_asce(); 74 set_user_asce(mm); 75 } 76 __tlb_flush_local(); 77 } 78 79 int crst_table_upgrade(struct mm_struct *mm) 80 { 81 unsigned long *table, *pgd; 82 83 /* upgrade should only happen from 3 to 4 levels */ 84 BUG_ON(mm->context.asce_limit != (1UL << 42)); 85 86 table = crst_table_alloc(mm); 87 if (!table) 88 return -ENOMEM; 89 90 spin_lock_bh(&mm->page_table_lock); 91 pgd = (unsigned long *) mm->pgd; 92 crst_table_init(table, _REGION2_ENTRY_EMPTY); 93 pgd_populate(mm, (pgd_t *) table, (pud_t *) pgd); 94 mm->pgd = (pgd_t *) table; 95 mm->context.asce_limit = 1UL << 53; 96 mm->context.asce = __pa(mm->pgd) | _ASCE_TABLE_LENGTH | 97 _ASCE_USER_BITS | _ASCE_TYPE_REGION2; 98 mm->task_size = mm->context.asce_limit; 99 spin_unlock_bh(&mm->page_table_lock); 100 101 on_each_cpu(__crst_table_upgrade, mm, 0); 102 return 0; 103 } 104 105 void crst_table_downgrade(struct mm_struct *mm) 106 { 107 pgd_t *pgd; 108 109 /* downgrade should only happen from 3 to 2 levels (compat only) */ 110 BUG_ON(mm->context.asce_limit != (1UL << 42)); 111 112 if (current->active_mm == mm) { 113 clear_user_asce(); 114 __tlb_flush_mm(mm); 115 } 116 117 pgd = mm->pgd; 118 mm->pgd = (pgd_t *) (pgd_val(*pgd) & _REGION_ENTRY_ORIGIN); 119 mm->context.asce_limit = 1UL << 31; 120 mm->context.asce = __pa(mm->pgd) | _ASCE_TABLE_LENGTH | 121 _ASCE_USER_BITS | _ASCE_TYPE_SEGMENT; 122 mm->task_size = mm->context.asce_limit; 123 crst_table_free(mm, (unsigned long *) pgd); 124 125 if (current->active_mm == mm) 126 set_user_asce(mm); 127 } 128 129 static inline unsigned int atomic_xor_bits(atomic_t *v, unsigned int bits) 130 { 131 unsigned int old, new; 132 133 do { 134 old = atomic_read(v); 135 new = old ^ bits; 136 } while (atomic_cmpxchg(v, old, new) != old); 137 return new; 138 } 139 140 #ifdef CONFIG_PGSTE 141 142 struct page *page_table_alloc_pgste(struct mm_struct *mm) 143 { 144 struct page *page; 145 unsigned long *table; 146 147 page = alloc_page(GFP_KERNEL|__GFP_REPEAT); 148 if (page) { 149 table = (unsigned long *) page_to_phys(page); 150 clear_table(table, _PAGE_INVALID, PAGE_SIZE/2); 151 clear_table(table + PTRS_PER_PTE, 0, PAGE_SIZE/2); 152 } 153 return page; 154 } 155 156 void page_table_free_pgste(struct page *page) 157 { 158 __free_page(page); 159 } 160 161 #endif /* CONFIG_PGSTE */ 162 163 /* 164 * page table entry allocation/free routines. 165 */ 166 unsigned long *page_table_alloc(struct mm_struct *mm) 167 { 168 unsigned long *table; 169 struct page *page; 170 unsigned int mask, bit; 171 172 /* Try to get a fragment of a 4K page as a 2K page table */ 173 if (!mm_alloc_pgste(mm)) { 174 table = NULL; 175 spin_lock_bh(&mm->context.pgtable_lock); 176 if (!list_empty(&mm->context.pgtable_list)) { 177 page = list_first_entry(&mm->context.pgtable_list, 178 struct page, lru); 179 mask = atomic_read(&page->_mapcount); 180 mask = (mask | (mask >> 4)) & 3; 181 if (mask != 3) { 182 table = (unsigned long *) page_to_phys(page); 183 bit = mask & 1; /* =1 -> second 2K */ 184 if (bit) 185 table += PTRS_PER_PTE; 186 atomic_xor_bits(&page->_mapcount, 1U << bit); 187 list_del(&page->lru); 188 } 189 } 190 spin_unlock_bh(&mm->context.pgtable_lock); 191 if (table) 192 return table; 193 } 194 /* Allocate a fresh page */ 195 page = alloc_page(GFP_KERNEL); 196 if (!page) 197 return NULL; 198 if (!pgtable_page_ctor(page)) { 199 __free_page(page); 200 return NULL; 201 } 202 /* Initialize page table */ 203 table = (unsigned long *) page_to_phys(page); 204 if (mm_alloc_pgste(mm)) { 205 /* Return 4K page table with PGSTEs */ 206 atomic_set(&page->_mapcount, 3); 207 clear_table(table, _PAGE_INVALID, PAGE_SIZE/2); 208 clear_table(table + PTRS_PER_PTE, 0, PAGE_SIZE/2); 209 } else { 210 /* Return the first 2K fragment of the page */ 211 atomic_set(&page->_mapcount, 1); 212 clear_table(table, _PAGE_INVALID, PAGE_SIZE); 213 spin_lock_bh(&mm->context.pgtable_lock); 214 list_add(&page->lru, &mm->context.pgtable_list); 215 spin_unlock_bh(&mm->context.pgtable_lock); 216 } 217 return table; 218 } 219 220 void page_table_free(struct mm_struct *mm, unsigned long *table) 221 { 222 struct page *page; 223 unsigned int bit, mask; 224 225 page = pfn_to_page(__pa(table) >> PAGE_SHIFT); 226 if (!mm_alloc_pgste(mm)) { 227 /* Free 2K page table fragment of a 4K page */ 228 bit = (__pa(table) & ~PAGE_MASK)/(PTRS_PER_PTE*sizeof(pte_t)); 229 spin_lock_bh(&mm->context.pgtable_lock); 230 mask = atomic_xor_bits(&page->_mapcount, 1U << bit); 231 if (mask & 3) 232 list_add(&page->lru, &mm->context.pgtable_list); 233 else 234 list_del(&page->lru); 235 spin_unlock_bh(&mm->context.pgtable_lock); 236 if (mask != 0) 237 return; 238 } 239 240 pgtable_page_dtor(page); 241 atomic_set(&page->_mapcount, -1); 242 __free_page(page); 243 } 244 245 void page_table_free_rcu(struct mmu_gather *tlb, unsigned long *table, 246 unsigned long vmaddr) 247 { 248 struct mm_struct *mm; 249 struct page *page; 250 unsigned int bit, mask; 251 252 mm = tlb->mm; 253 page = pfn_to_page(__pa(table) >> PAGE_SHIFT); 254 if (mm_alloc_pgste(mm)) { 255 gmap_unlink(mm, table, vmaddr); 256 table = (unsigned long *) (__pa(table) | 3); 257 tlb_remove_table(tlb, table); 258 return; 259 } 260 bit = (__pa(table) & ~PAGE_MASK) / (PTRS_PER_PTE*sizeof(pte_t)); 261 spin_lock_bh(&mm->context.pgtable_lock); 262 mask = atomic_xor_bits(&page->_mapcount, 0x11U << bit); 263 if (mask & 3) 264 list_add_tail(&page->lru, &mm->context.pgtable_list); 265 else 266 list_del(&page->lru); 267 spin_unlock_bh(&mm->context.pgtable_lock); 268 table = (unsigned long *) (__pa(table) | (1U << bit)); 269 tlb_remove_table(tlb, table); 270 } 271 272 static void __tlb_remove_table(void *_table) 273 { 274 unsigned int mask = (unsigned long) _table & 3; 275 void *table = (void *)((unsigned long) _table ^ mask); 276 struct page *page = pfn_to_page(__pa(table) >> PAGE_SHIFT); 277 278 switch (mask) { 279 case 0: /* pmd or pud */ 280 free_pages((unsigned long) table, 2); 281 break; 282 case 1: /* lower 2K of a 4K page table */ 283 case 2: /* higher 2K of a 4K page table */ 284 if (atomic_xor_bits(&page->_mapcount, mask << 4) != 0) 285 break; 286 /* fallthrough */ 287 case 3: /* 4K page table with pgstes */ 288 pgtable_page_dtor(page); 289 atomic_set(&page->_mapcount, -1); 290 __free_page(page); 291 break; 292 } 293 } 294 295 static void tlb_remove_table_smp_sync(void *arg) 296 { 297 /* Simply deliver the interrupt */ 298 } 299 300 static void tlb_remove_table_one(void *table) 301 { 302 /* 303 * This isn't an RCU grace period and hence the page-tables cannot be 304 * assumed to be actually RCU-freed. 305 * 306 * It is however sufficient for software page-table walkers that rely 307 * on IRQ disabling. See the comment near struct mmu_table_batch. 308 */ 309 smp_call_function(tlb_remove_table_smp_sync, NULL, 1); 310 __tlb_remove_table(table); 311 } 312 313 static void tlb_remove_table_rcu(struct rcu_head *head) 314 { 315 struct mmu_table_batch *batch; 316 int i; 317 318 batch = container_of(head, struct mmu_table_batch, rcu); 319 320 for (i = 0; i < batch->nr; i++) 321 __tlb_remove_table(batch->tables[i]); 322 323 free_page((unsigned long)batch); 324 } 325 326 void tlb_table_flush(struct mmu_gather *tlb) 327 { 328 struct mmu_table_batch **batch = &tlb->batch; 329 330 if (*batch) { 331 call_rcu_sched(&(*batch)->rcu, tlb_remove_table_rcu); 332 *batch = NULL; 333 } 334 } 335 336 void tlb_remove_table(struct mmu_gather *tlb, void *table) 337 { 338 struct mmu_table_batch **batch = &tlb->batch; 339 340 tlb->mm->context.flush_mm = 1; 341 if (*batch == NULL) { 342 *batch = (struct mmu_table_batch *) 343 __get_free_page(GFP_NOWAIT | __GFP_NOWARN); 344 if (*batch == NULL) { 345 __tlb_flush_mm_lazy(tlb->mm); 346 tlb_remove_table_one(table); 347 return; 348 } 349 (*batch)->nr = 0; 350 } 351 (*batch)->tables[(*batch)->nr++] = table; 352 if ((*batch)->nr == MAX_TABLE_BATCH) 353 tlb_flush_mmu(tlb); 354 } 355