Martin Schwidefsky | 1e133ab | 2016-03-08 11:49:57 +0100 | [diff] [blame] | 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 | |
Martin Schwidefsky | 1aea9b3 | 2017-04-24 18:19:10 +0200 | [diff] [blame] | 79 | int crst_table_upgrade(struct mm_struct *mm, unsigned long end) |
Martin Schwidefsky | 1e133ab | 2016-03-08 11:49:57 +0100 | [diff] [blame] | 80 | { |
| 81 | unsigned long *table, *pgd; |
Martin Schwidefsky | 1aea9b3 | 2017-04-24 18:19:10 +0200 | [diff] [blame] | 82 | int rc, notify; |
Martin Schwidefsky | 1e133ab | 2016-03-08 11:49:57 +0100 | [diff] [blame] | 83 | |
Martin Schwidefsky | 1aea9b3 | 2017-04-24 18:19:10 +0200 | [diff] [blame] | 84 | /* upgrade should only happen from 3 to 4, 3 to 5, or 4 to 5 levels */ |
| 85 | BUG_ON(mm->context.asce_limit < (1UL << 42)); |
| 86 | if (end >= TASK_SIZE_MAX) |
Martin Schwidefsky | 1e133ab | 2016-03-08 11:49:57 +0100 | [diff] [blame] | 87 | return -ENOMEM; |
Martin Schwidefsky | 1aea9b3 | 2017-04-24 18:19:10 +0200 | [diff] [blame] | 88 | rc = 0; |
| 89 | notify = 0; |
| 90 | while (mm->context.asce_limit < end) { |
| 91 | table = crst_table_alloc(mm); |
| 92 | if (!table) { |
| 93 | rc = -ENOMEM; |
| 94 | break; |
| 95 | } |
| 96 | spin_lock_bh(&mm->page_table_lock); |
| 97 | pgd = (unsigned long *) mm->pgd; |
| 98 | if (mm->context.asce_limit == (1UL << 42)) { |
| 99 | crst_table_init(table, _REGION2_ENTRY_EMPTY); |
| 100 | p4d_populate(mm, (p4d_t *) table, (pud_t *) pgd); |
| 101 | mm->pgd = (pgd_t *) table; |
| 102 | mm->context.asce_limit = 1UL << 53; |
| 103 | mm->context.asce = __pa(mm->pgd) | _ASCE_TABLE_LENGTH | |
| 104 | _ASCE_USER_BITS | _ASCE_TYPE_REGION2; |
| 105 | } else { |
| 106 | crst_table_init(table, _REGION1_ENTRY_EMPTY); |
| 107 | pgd_populate(mm, (pgd_t *) table, (p4d_t *) pgd); |
| 108 | mm->pgd = (pgd_t *) table; |
| 109 | mm->context.asce_limit = -PAGE_SIZE; |
| 110 | mm->context.asce = __pa(mm->pgd) | _ASCE_TABLE_LENGTH | |
| 111 | _ASCE_USER_BITS | _ASCE_TYPE_REGION1; |
| 112 | } |
| 113 | notify = 1; |
| 114 | spin_unlock_bh(&mm->page_table_lock); |
| 115 | } |
| 116 | if (notify) |
| 117 | on_each_cpu(__crst_table_upgrade, mm, 0); |
| 118 | return rc; |
Martin Schwidefsky | 1e133ab | 2016-03-08 11:49:57 +0100 | [diff] [blame] | 119 | } |
| 120 | |
Gerald Schaefer | 723cacb | 2016-04-15 16:38:40 +0200 | [diff] [blame] | 121 | void crst_table_downgrade(struct mm_struct *mm) |
Martin Schwidefsky | 1e133ab | 2016-03-08 11:49:57 +0100 | [diff] [blame] | 122 | { |
| 123 | pgd_t *pgd; |
| 124 | |
Gerald Schaefer | 723cacb | 2016-04-15 16:38:40 +0200 | [diff] [blame] | 125 | /* downgrade should only happen from 3 to 2 levels (compat only) */ |
| 126 | BUG_ON(mm->context.asce_limit != (1UL << 42)); |
| 127 | |
Martin Schwidefsky | 1e133ab | 2016-03-08 11:49:57 +0100 | [diff] [blame] | 128 | if (current->active_mm == mm) { |
| 129 | clear_user_asce(); |
| 130 | __tlb_flush_mm(mm); |
| 131 | } |
Gerald Schaefer | 723cacb | 2016-04-15 16:38:40 +0200 | [diff] [blame] | 132 | |
| 133 | pgd = mm->pgd; |
| 134 | mm->pgd = (pgd_t *) (pgd_val(*pgd) & _REGION_ENTRY_ORIGIN); |
| 135 | mm->context.asce_limit = 1UL << 31; |
| 136 | mm->context.asce = __pa(mm->pgd) | _ASCE_TABLE_LENGTH | |
| 137 | _ASCE_USER_BITS | _ASCE_TYPE_SEGMENT; |
Gerald Schaefer | 723cacb | 2016-04-15 16:38:40 +0200 | [diff] [blame] | 138 | crst_table_free(mm, (unsigned long *) pgd); |
| 139 | |
Martin Schwidefsky | 1e133ab | 2016-03-08 11:49:57 +0100 | [diff] [blame] | 140 | if (current->active_mm == mm) |
| 141 | set_user_asce(mm); |
| 142 | } |
| 143 | |
| 144 | static inline unsigned int atomic_xor_bits(atomic_t *v, unsigned int bits) |
| 145 | { |
| 146 | unsigned int old, new; |
| 147 | |
| 148 | do { |
| 149 | old = atomic_read(v); |
| 150 | new = old ^ bits; |
| 151 | } while (atomic_cmpxchg(v, old, new) != old); |
| 152 | return new; |
| 153 | } |
| 154 | |
Martin Schwidefsky | 4be130a | 2016-03-08 12:12:18 +0100 | [diff] [blame] | 155 | #ifdef CONFIG_PGSTE |
| 156 | |
| 157 | struct page *page_table_alloc_pgste(struct mm_struct *mm) |
| 158 | { |
| 159 | struct page *page; |
| 160 | unsigned long *table; |
| 161 | |
Michal Hocko | faee35a | 2017-03-07 16:48:40 +0100 | [diff] [blame] | 162 | page = alloc_page(GFP_KERNEL); |
Martin Schwidefsky | 4be130a | 2016-03-08 12:12:18 +0100 | [diff] [blame] | 163 | if (page) { |
| 164 | table = (unsigned long *) page_to_phys(page); |
| 165 | clear_table(table, _PAGE_INVALID, PAGE_SIZE/2); |
| 166 | clear_table(table + PTRS_PER_PTE, 0, PAGE_SIZE/2); |
| 167 | } |
| 168 | return page; |
| 169 | } |
| 170 | |
| 171 | void page_table_free_pgste(struct page *page) |
| 172 | { |
| 173 | __free_page(page); |
| 174 | } |
| 175 | |
| 176 | #endif /* CONFIG_PGSTE */ |
| 177 | |
Martin Schwidefsky | 1e133ab | 2016-03-08 11:49:57 +0100 | [diff] [blame] | 178 | /* |
| 179 | * page table entry allocation/free routines. |
| 180 | */ |
| 181 | unsigned long *page_table_alloc(struct mm_struct *mm) |
| 182 | { |
| 183 | unsigned long *table; |
| 184 | struct page *page; |
| 185 | unsigned int mask, bit; |
| 186 | |
| 187 | /* Try to get a fragment of a 4K page as a 2K page table */ |
| 188 | if (!mm_alloc_pgste(mm)) { |
| 189 | table = NULL; |
Martin Schwidefsky | 8ecb1a5 | 2016-03-08 11:54:14 +0100 | [diff] [blame] | 190 | spin_lock_bh(&mm->context.pgtable_lock); |
Martin Schwidefsky | 1e133ab | 2016-03-08 11:49:57 +0100 | [diff] [blame] | 191 | if (!list_empty(&mm->context.pgtable_list)) { |
| 192 | page = list_first_entry(&mm->context.pgtable_list, |
| 193 | struct page, lru); |
| 194 | mask = atomic_read(&page->_mapcount); |
| 195 | mask = (mask | (mask >> 4)) & 3; |
| 196 | if (mask != 3) { |
| 197 | table = (unsigned long *) page_to_phys(page); |
| 198 | bit = mask & 1; /* =1 -> second 2K */ |
| 199 | if (bit) |
| 200 | table += PTRS_PER_PTE; |
| 201 | atomic_xor_bits(&page->_mapcount, 1U << bit); |
| 202 | list_del(&page->lru); |
| 203 | } |
| 204 | } |
Martin Schwidefsky | 8ecb1a5 | 2016-03-08 11:54:14 +0100 | [diff] [blame] | 205 | spin_unlock_bh(&mm->context.pgtable_lock); |
Martin Schwidefsky | 1e133ab | 2016-03-08 11:49:57 +0100 | [diff] [blame] | 206 | if (table) |
| 207 | return table; |
| 208 | } |
| 209 | /* Allocate a fresh page */ |
Michal Hocko | 10d58bf | 2016-06-24 14:49:17 -0700 | [diff] [blame] | 210 | page = alloc_page(GFP_KERNEL); |
Martin Schwidefsky | 1e133ab | 2016-03-08 11:49:57 +0100 | [diff] [blame] | 211 | if (!page) |
| 212 | return NULL; |
| 213 | if (!pgtable_page_ctor(page)) { |
| 214 | __free_page(page); |
| 215 | return NULL; |
| 216 | } |
| 217 | /* Initialize page table */ |
| 218 | table = (unsigned long *) page_to_phys(page); |
| 219 | if (mm_alloc_pgste(mm)) { |
| 220 | /* Return 4K page table with PGSTEs */ |
| 221 | atomic_set(&page->_mapcount, 3); |
| 222 | clear_table(table, _PAGE_INVALID, PAGE_SIZE/2); |
| 223 | clear_table(table + PTRS_PER_PTE, 0, PAGE_SIZE/2); |
| 224 | } else { |
| 225 | /* Return the first 2K fragment of the page */ |
| 226 | atomic_set(&page->_mapcount, 1); |
| 227 | clear_table(table, _PAGE_INVALID, PAGE_SIZE); |
Martin Schwidefsky | 8ecb1a5 | 2016-03-08 11:54:14 +0100 | [diff] [blame] | 228 | spin_lock_bh(&mm->context.pgtable_lock); |
Martin Schwidefsky | 1e133ab | 2016-03-08 11:49:57 +0100 | [diff] [blame] | 229 | list_add(&page->lru, &mm->context.pgtable_list); |
Martin Schwidefsky | 8ecb1a5 | 2016-03-08 11:54:14 +0100 | [diff] [blame] | 230 | spin_unlock_bh(&mm->context.pgtable_lock); |
Martin Schwidefsky | 1e133ab | 2016-03-08 11:49:57 +0100 | [diff] [blame] | 231 | } |
| 232 | return table; |
| 233 | } |
| 234 | |
| 235 | void page_table_free(struct mm_struct *mm, unsigned long *table) |
| 236 | { |
| 237 | struct page *page; |
| 238 | unsigned int bit, mask; |
| 239 | |
| 240 | page = pfn_to_page(__pa(table) >> PAGE_SHIFT); |
| 241 | if (!mm_alloc_pgste(mm)) { |
| 242 | /* Free 2K page table fragment of a 4K page */ |
| 243 | bit = (__pa(table) & ~PAGE_MASK)/(PTRS_PER_PTE*sizeof(pte_t)); |
Martin Schwidefsky | 8ecb1a5 | 2016-03-08 11:54:14 +0100 | [diff] [blame] | 244 | spin_lock_bh(&mm->context.pgtable_lock); |
Martin Schwidefsky | 1e133ab | 2016-03-08 11:49:57 +0100 | [diff] [blame] | 245 | mask = atomic_xor_bits(&page->_mapcount, 1U << bit); |
| 246 | if (mask & 3) |
| 247 | list_add(&page->lru, &mm->context.pgtable_list); |
| 248 | else |
| 249 | list_del(&page->lru); |
Martin Schwidefsky | 8ecb1a5 | 2016-03-08 11:54:14 +0100 | [diff] [blame] | 250 | spin_unlock_bh(&mm->context.pgtable_lock); |
Martin Schwidefsky | 1e133ab | 2016-03-08 11:49:57 +0100 | [diff] [blame] | 251 | if (mask != 0) |
| 252 | return; |
| 253 | } |
| 254 | |
| 255 | pgtable_page_dtor(page); |
| 256 | atomic_set(&page->_mapcount, -1); |
| 257 | __free_page(page); |
| 258 | } |
| 259 | |
| 260 | void page_table_free_rcu(struct mmu_gather *tlb, unsigned long *table, |
| 261 | unsigned long vmaddr) |
| 262 | { |
| 263 | struct mm_struct *mm; |
| 264 | struct page *page; |
| 265 | unsigned int bit, mask; |
| 266 | |
| 267 | mm = tlb->mm; |
| 268 | page = pfn_to_page(__pa(table) >> PAGE_SHIFT); |
| 269 | if (mm_alloc_pgste(mm)) { |
| 270 | gmap_unlink(mm, table, vmaddr); |
| 271 | table = (unsigned long *) (__pa(table) | 3); |
| 272 | tlb_remove_table(tlb, table); |
| 273 | return; |
| 274 | } |
| 275 | bit = (__pa(table) & ~PAGE_MASK) / (PTRS_PER_PTE*sizeof(pte_t)); |
Martin Schwidefsky | 8ecb1a5 | 2016-03-08 11:54:14 +0100 | [diff] [blame] | 276 | spin_lock_bh(&mm->context.pgtable_lock); |
Martin Schwidefsky | 1e133ab | 2016-03-08 11:49:57 +0100 | [diff] [blame] | 277 | mask = atomic_xor_bits(&page->_mapcount, 0x11U << bit); |
| 278 | if (mask & 3) |
| 279 | list_add_tail(&page->lru, &mm->context.pgtable_list); |
| 280 | else |
| 281 | list_del(&page->lru); |
Martin Schwidefsky | 8ecb1a5 | 2016-03-08 11:54:14 +0100 | [diff] [blame] | 282 | spin_unlock_bh(&mm->context.pgtable_lock); |
Martin Schwidefsky | 1e133ab | 2016-03-08 11:49:57 +0100 | [diff] [blame] | 283 | table = (unsigned long *) (__pa(table) | (1U << bit)); |
| 284 | tlb_remove_table(tlb, table); |
| 285 | } |
| 286 | |
| 287 | static void __tlb_remove_table(void *_table) |
| 288 | { |
| 289 | unsigned int mask = (unsigned long) _table & 3; |
| 290 | void *table = (void *)((unsigned long) _table ^ mask); |
| 291 | struct page *page = pfn_to_page(__pa(table) >> PAGE_SHIFT); |
| 292 | |
| 293 | switch (mask) { |
Martin Schwidefsky | 1aea9b3 | 2017-04-24 18:19:10 +0200 | [diff] [blame] | 294 | case 0: /* pmd, pud, or p4d */ |
Martin Schwidefsky | 1e133ab | 2016-03-08 11:49:57 +0100 | [diff] [blame] | 295 | free_pages((unsigned long) table, 2); |
| 296 | break; |
| 297 | case 1: /* lower 2K of a 4K page table */ |
| 298 | case 2: /* higher 2K of a 4K page table */ |
| 299 | if (atomic_xor_bits(&page->_mapcount, mask << 4) != 0) |
| 300 | break; |
| 301 | /* fallthrough */ |
| 302 | case 3: /* 4K page table with pgstes */ |
| 303 | pgtable_page_dtor(page); |
| 304 | atomic_set(&page->_mapcount, -1); |
| 305 | __free_page(page); |
| 306 | break; |
| 307 | } |
| 308 | } |
| 309 | |
| 310 | static void tlb_remove_table_smp_sync(void *arg) |
| 311 | { |
| 312 | /* Simply deliver the interrupt */ |
| 313 | } |
| 314 | |
| 315 | static void tlb_remove_table_one(void *table) |
| 316 | { |
| 317 | /* |
| 318 | * This isn't an RCU grace period and hence the page-tables cannot be |
| 319 | * assumed to be actually RCU-freed. |
| 320 | * |
| 321 | * It is however sufficient for software page-table walkers that rely |
| 322 | * on IRQ disabling. See the comment near struct mmu_table_batch. |
| 323 | */ |
| 324 | smp_call_function(tlb_remove_table_smp_sync, NULL, 1); |
| 325 | __tlb_remove_table(table); |
| 326 | } |
| 327 | |
| 328 | static void tlb_remove_table_rcu(struct rcu_head *head) |
| 329 | { |
| 330 | struct mmu_table_batch *batch; |
| 331 | int i; |
| 332 | |
| 333 | batch = container_of(head, struct mmu_table_batch, rcu); |
| 334 | |
| 335 | for (i = 0; i < batch->nr; i++) |
| 336 | __tlb_remove_table(batch->tables[i]); |
| 337 | |
| 338 | free_page((unsigned long)batch); |
| 339 | } |
| 340 | |
| 341 | void tlb_table_flush(struct mmu_gather *tlb) |
| 342 | { |
| 343 | struct mmu_table_batch **batch = &tlb->batch; |
| 344 | |
| 345 | if (*batch) { |
| 346 | call_rcu_sched(&(*batch)->rcu, tlb_remove_table_rcu); |
| 347 | *batch = NULL; |
| 348 | } |
| 349 | } |
| 350 | |
| 351 | void tlb_remove_table(struct mmu_gather *tlb, void *table) |
| 352 | { |
| 353 | struct mmu_table_batch **batch = &tlb->batch; |
| 354 | |
| 355 | tlb->mm->context.flush_mm = 1; |
| 356 | if (*batch == NULL) { |
| 357 | *batch = (struct mmu_table_batch *) |
| 358 | __get_free_page(GFP_NOWAIT | __GFP_NOWARN); |
| 359 | if (*batch == NULL) { |
| 360 | __tlb_flush_mm_lazy(tlb->mm); |
| 361 | tlb_remove_table_one(table); |
| 362 | return; |
| 363 | } |
| 364 | (*batch)->nr = 0; |
| 365 | } |
| 366 | (*batch)->tables[(*batch)->nr++] = table; |
| 367 | if ((*batch)->nr == MAX_TABLE_BATCH) |
| 368 | tlb_flush_mmu(tlb); |
| 369 | } |