1 // SPDX-License-Identifier: GPL-2.0 2 /* 3 * Page table allocation functions 4 * 5 * Copyright IBM Corp. 2016 6 * Author(s): Martin Schwidefsky <schwidefsky@de.ibm.com> 7 */ 8 9 #include <linux/sysctl.h> 10 #include <linux/slab.h> 11 #include <linux/mm.h> 12 #include <asm/mmu_context.h> 13 #include <asm/page-states.h> 14 #include <asm/pgalloc.h> 15 #include <asm/gmap.h> 16 #include <asm/tlb.h> 17 #include <asm/tlbflush.h> 18 19 unsigned long *crst_table_alloc(struct mm_struct *mm) 20 { 21 struct ptdesc *ptdesc = pagetable_alloc(GFP_KERNEL, CRST_ALLOC_ORDER); 22 unsigned long *table; 23 24 if (!ptdesc) 25 return NULL; 26 table = ptdesc_to_virt(ptdesc); 27 __arch_set_page_dat(table, 1UL << CRST_ALLOC_ORDER); 28 return table; 29 } 30 31 void crst_table_free(struct mm_struct *mm, unsigned long *table) 32 { 33 if (!table) 34 return; 35 pagetable_free(virt_to_ptdesc(table)); 36 } 37 38 static void __crst_table_upgrade(void *arg) 39 { 40 struct mm_struct *mm = arg; 41 struct ctlreg asce; 42 43 /* change all active ASCEs to avoid the creation of new TLBs */ 44 if (current->active_mm == mm) { 45 asce.val = mm->context.asce; 46 get_lowcore()->user_asce = asce; 47 local_ctl_load(7, &asce); 48 if (!test_thread_flag(TIF_ASCE_PRIMARY)) 49 local_ctl_load(1, &asce); 50 } 51 __tlb_flush_local(); 52 } 53 54 int crst_table_upgrade(struct mm_struct *mm, unsigned long end) 55 { 56 unsigned long *pgd = NULL, *p4d = NULL, *__pgd; 57 unsigned long asce_limit = mm->context.asce_limit; 58 59 mmap_assert_write_locked(mm); 60 61 /* upgrade should only happen from 3 to 4, 3 to 5, or 4 to 5 levels */ 62 VM_BUG_ON(asce_limit < _REGION2_SIZE); 63 64 if (end <= asce_limit) 65 return 0; 66 67 if (asce_limit == _REGION2_SIZE) { 68 p4d = crst_table_alloc(mm); 69 if (unlikely(!p4d)) 70 goto err_p4d; 71 crst_table_init(p4d, _REGION2_ENTRY_EMPTY); 72 pagetable_p4d_ctor(virt_to_ptdesc(p4d)); 73 } 74 if (end > _REGION1_SIZE) { 75 pgd = crst_table_alloc(mm); 76 if (unlikely(!pgd)) 77 goto err_pgd; 78 crst_table_init(pgd, _REGION1_ENTRY_EMPTY); 79 pagetable_pgd_ctor(virt_to_ptdesc(pgd)); 80 } 81 82 spin_lock_bh(&mm->page_table_lock); 83 84 if (p4d) { 85 __pgd = (unsigned long *) mm->pgd; 86 p4d_populate(mm, (p4d_t *) p4d, (pud_t *) __pgd); 87 mm->pgd = (pgd_t *) p4d; 88 mm->context.asce_limit = _REGION1_SIZE; 89 mm->context.asce = __pa(mm->pgd) | _ASCE_TABLE_LENGTH | 90 _ASCE_USER_BITS | _ASCE_TYPE_REGION2; 91 mm_inc_nr_puds(mm); 92 } 93 if (pgd) { 94 __pgd = (unsigned long *) mm->pgd; 95 pgd_populate(mm, (pgd_t *) pgd, (p4d_t *) __pgd); 96 mm->pgd = (pgd_t *) pgd; 97 mm->context.asce_limit = TASK_SIZE_MAX; 98 mm->context.asce = __pa(mm->pgd) | _ASCE_TABLE_LENGTH | 99 _ASCE_USER_BITS | _ASCE_TYPE_REGION1; 100 } 101 102 spin_unlock_bh(&mm->page_table_lock); 103 104 on_each_cpu(__crst_table_upgrade, mm, 0); 105 106 return 0; 107 108 err_pgd: 109 pagetable_dtor(virt_to_ptdesc(p4d)); 110 crst_table_free(mm, p4d); 111 err_p4d: 112 return -ENOMEM; 113 } 114 115 #ifdef CONFIG_PGSTE 116 117 struct ptdesc *page_table_alloc_pgste(struct mm_struct *mm) 118 { 119 struct ptdesc *ptdesc; 120 u64 *table; 121 122 ptdesc = pagetable_alloc(GFP_KERNEL, 0); 123 if (ptdesc) { 124 table = (u64 *)ptdesc_to_virt(ptdesc); 125 __arch_set_page_dat(table, 1); 126 memset64(table, _PAGE_INVALID, PTRS_PER_PTE); 127 memset64(table + PTRS_PER_PTE, 0, PTRS_PER_PTE); 128 } 129 return ptdesc; 130 } 131 132 void page_table_free_pgste(struct ptdesc *ptdesc) 133 { 134 pagetable_free(ptdesc); 135 } 136 137 #endif /* CONFIG_PGSTE */ 138 139 unsigned long *page_table_alloc(struct mm_struct *mm) 140 { 141 struct ptdesc *ptdesc; 142 unsigned long *table; 143 144 ptdesc = pagetable_alloc(GFP_KERNEL, 0); 145 if (!ptdesc) 146 return NULL; 147 if (!pagetable_pte_ctor(ptdesc)) { 148 pagetable_free(ptdesc); 149 return NULL; 150 } 151 table = ptdesc_to_virt(ptdesc); 152 __arch_set_page_dat(table, 1); 153 memset64((u64 *)table, _PAGE_INVALID, PTRS_PER_PTE); 154 memset64((u64 *)table + PTRS_PER_PTE, 0, PTRS_PER_PTE); 155 return table; 156 } 157 158 void page_table_free(struct mm_struct *mm, unsigned long *table) 159 { 160 struct ptdesc *ptdesc = virt_to_ptdesc(table); 161 162 pagetable_dtor_free(ptdesc); 163 } 164 165 #ifdef CONFIG_TRANSPARENT_HUGEPAGE 166 static void pte_free_now(struct rcu_head *head) 167 { 168 struct ptdesc *ptdesc = container_of(head, struct ptdesc, pt_rcu_head); 169 170 pagetable_dtor_free(ptdesc); 171 } 172 173 void pte_free_defer(struct mm_struct *mm, pgtable_t pgtable) 174 { 175 struct ptdesc *ptdesc = virt_to_ptdesc(pgtable); 176 177 call_rcu(&ptdesc->pt_rcu_head, pte_free_now); 178 /* 179 * THPs are not allowed for KVM guests. Warn if pgste ever reaches here. 180 * Turn to the generic pte_free_defer() version once gmap is removed. 181 */ 182 WARN_ON_ONCE(mm_has_pgste(mm)); 183 } 184 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */ 185 186 /* 187 * Base infrastructure required to generate basic asces, region, segment, 188 * and page tables that do not make use of enhanced features like EDAT1. 189 */ 190 191 static struct kmem_cache *base_pgt_cache; 192 193 static unsigned long *base_pgt_alloc(void) 194 { 195 unsigned long *table; 196 197 table = kmem_cache_alloc(base_pgt_cache, GFP_KERNEL); 198 if (table) 199 memset64((u64 *)table, _PAGE_INVALID, PTRS_PER_PTE); 200 return table; 201 } 202 203 static void base_pgt_free(unsigned long *table) 204 { 205 kmem_cache_free(base_pgt_cache, table); 206 } 207 208 static unsigned long *base_crst_alloc(unsigned long val) 209 { 210 unsigned long *table; 211 struct ptdesc *ptdesc; 212 213 ptdesc = pagetable_alloc(GFP_KERNEL, CRST_ALLOC_ORDER); 214 if (!ptdesc) 215 return NULL; 216 table = ptdesc_address(ptdesc); 217 crst_table_init(table, val); 218 return table; 219 } 220 221 static void base_crst_free(unsigned long *table) 222 { 223 if (!table) 224 return; 225 pagetable_free(virt_to_ptdesc(table)); 226 } 227 228 #define BASE_ADDR_END_FUNC(NAME, SIZE) \ 229 static inline unsigned long base_##NAME##_addr_end(unsigned long addr, \ 230 unsigned long end) \ 231 { \ 232 unsigned long next = (addr + (SIZE)) & ~((SIZE) - 1); \ 233 \ 234 return (next - 1) < (end - 1) ? next : end; \ 235 } 236 237 BASE_ADDR_END_FUNC(page, PAGE_SIZE) 238 BASE_ADDR_END_FUNC(segment, _SEGMENT_SIZE) 239 BASE_ADDR_END_FUNC(region3, _REGION3_SIZE) 240 BASE_ADDR_END_FUNC(region2, _REGION2_SIZE) 241 BASE_ADDR_END_FUNC(region1, _REGION1_SIZE) 242 243 static inline unsigned long base_lra(unsigned long address) 244 { 245 unsigned long real; 246 247 asm volatile( 248 " lra %0,0(%1)\n" 249 : "=d" (real) : "a" (address) : "cc"); 250 return real; 251 } 252 253 static int base_page_walk(unsigned long *origin, unsigned long addr, 254 unsigned long end, int alloc) 255 { 256 unsigned long *pte, next; 257 258 if (!alloc) 259 return 0; 260 pte = origin; 261 pte += (addr & _PAGE_INDEX) >> PAGE_SHIFT; 262 do { 263 next = base_page_addr_end(addr, end); 264 *pte = base_lra(addr); 265 } while (pte++, addr = next, addr < end); 266 return 0; 267 } 268 269 static int base_segment_walk(unsigned long *origin, unsigned long addr, 270 unsigned long end, int alloc) 271 { 272 unsigned long *ste, next, *table; 273 int rc; 274 275 ste = origin; 276 ste += (addr & _SEGMENT_INDEX) >> _SEGMENT_SHIFT; 277 do { 278 next = base_segment_addr_end(addr, end); 279 if (*ste & _SEGMENT_ENTRY_INVALID) { 280 if (!alloc) 281 continue; 282 table = base_pgt_alloc(); 283 if (!table) 284 return -ENOMEM; 285 *ste = __pa(table) | _SEGMENT_ENTRY; 286 } 287 table = __va(*ste & _SEGMENT_ENTRY_ORIGIN); 288 rc = base_page_walk(table, addr, next, alloc); 289 if (rc) 290 return rc; 291 if (!alloc) 292 base_pgt_free(table); 293 cond_resched(); 294 } while (ste++, addr = next, addr < end); 295 return 0; 296 } 297 298 static int base_region3_walk(unsigned long *origin, unsigned long addr, 299 unsigned long end, int alloc) 300 { 301 unsigned long *rtte, next, *table; 302 int rc; 303 304 rtte = origin; 305 rtte += (addr & _REGION3_INDEX) >> _REGION3_SHIFT; 306 do { 307 next = base_region3_addr_end(addr, end); 308 if (*rtte & _REGION_ENTRY_INVALID) { 309 if (!alloc) 310 continue; 311 table = base_crst_alloc(_SEGMENT_ENTRY_EMPTY); 312 if (!table) 313 return -ENOMEM; 314 *rtte = __pa(table) | _REGION3_ENTRY; 315 } 316 table = __va(*rtte & _REGION_ENTRY_ORIGIN); 317 rc = base_segment_walk(table, addr, next, alloc); 318 if (rc) 319 return rc; 320 if (!alloc) 321 base_crst_free(table); 322 } while (rtte++, addr = next, addr < end); 323 return 0; 324 } 325 326 static int base_region2_walk(unsigned long *origin, unsigned long addr, 327 unsigned long end, int alloc) 328 { 329 unsigned long *rste, next, *table; 330 int rc; 331 332 rste = origin; 333 rste += (addr & _REGION2_INDEX) >> _REGION2_SHIFT; 334 do { 335 next = base_region2_addr_end(addr, end); 336 if (*rste & _REGION_ENTRY_INVALID) { 337 if (!alloc) 338 continue; 339 table = base_crst_alloc(_REGION3_ENTRY_EMPTY); 340 if (!table) 341 return -ENOMEM; 342 *rste = __pa(table) | _REGION2_ENTRY; 343 } 344 table = __va(*rste & _REGION_ENTRY_ORIGIN); 345 rc = base_region3_walk(table, addr, next, alloc); 346 if (rc) 347 return rc; 348 if (!alloc) 349 base_crst_free(table); 350 } while (rste++, addr = next, addr < end); 351 return 0; 352 } 353 354 static int base_region1_walk(unsigned long *origin, unsigned long addr, 355 unsigned long end, int alloc) 356 { 357 unsigned long *rfte, next, *table; 358 int rc; 359 360 rfte = origin; 361 rfte += (addr & _REGION1_INDEX) >> _REGION1_SHIFT; 362 do { 363 next = base_region1_addr_end(addr, end); 364 if (*rfte & _REGION_ENTRY_INVALID) { 365 if (!alloc) 366 continue; 367 table = base_crst_alloc(_REGION2_ENTRY_EMPTY); 368 if (!table) 369 return -ENOMEM; 370 *rfte = __pa(table) | _REGION1_ENTRY; 371 } 372 table = __va(*rfte & _REGION_ENTRY_ORIGIN); 373 rc = base_region2_walk(table, addr, next, alloc); 374 if (rc) 375 return rc; 376 if (!alloc) 377 base_crst_free(table); 378 } while (rfte++, addr = next, addr < end); 379 return 0; 380 } 381 382 /** 383 * base_asce_free - free asce and tables returned from base_asce_alloc() 384 * @asce: asce to be freed 385 * 386 * Frees all region, segment, and page tables that were allocated with a 387 * corresponding base_asce_alloc() call. 388 */ 389 void base_asce_free(unsigned long asce) 390 { 391 unsigned long *table = __va(asce & _ASCE_ORIGIN); 392 393 if (!asce) 394 return; 395 switch (asce & _ASCE_TYPE_MASK) { 396 case _ASCE_TYPE_SEGMENT: 397 base_segment_walk(table, 0, _REGION3_SIZE, 0); 398 break; 399 case _ASCE_TYPE_REGION3: 400 base_region3_walk(table, 0, _REGION2_SIZE, 0); 401 break; 402 case _ASCE_TYPE_REGION2: 403 base_region2_walk(table, 0, _REGION1_SIZE, 0); 404 break; 405 case _ASCE_TYPE_REGION1: 406 base_region1_walk(table, 0, TASK_SIZE_MAX, 0); 407 break; 408 } 409 base_crst_free(table); 410 } 411 412 static int base_pgt_cache_init(void) 413 { 414 static DEFINE_MUTEX(base_pgt_cache_mutex); 415 unsigned long sz = _PAGE_TABLE_SIZE; 416 417 if (base_pgt_cache) 418 return 0; 419 mutex_lock(&base_pgt_cache_mutex); 420 if (!base_pgt_cache) 421 base_pgt_cache = kmem_cache_create("base_pgt", sz, sz, 0, NULL); 422 mutex_unlock(&base_pgt_cache_mutex); 423 return base_pgt_cache ? 0 : -ENOMEM; 424 } 425 426 /** 427 * base_asce_alloc - create kernel mapping without enhanced DAT features 428 * @addr: virtual start address of kernel mapping 429 * @num_pages: number of consecutive pages 430 * 431 * Generate an asce, including all required region, segment and page tables, 432 * that can be used to access the virtual kernel mapping. The difference is 433 * that the returned asce does not make use of any enhanced DAT features like 434 * e.g. large pages. This is required for some I/O functions that pass an 435 * asce, like e.g. some service call requests. 436 * 437 * Note: the returned asce may NEVER be attached to any cpu. It may only be 438 * used for I/O requests. tlb entries that might result because the 439 * asce was attached to a cpu won't be cleared. 440 */ 441 unsigned long base_asce_alloc(unsigned long addr, unsigned long num_pages) 442 { 443 unsigned long asce, *table, end; 444 int rc; 445 446 if (base_pgt_cache_init()) 447 return 0; 448 end = addr + num_pages * PAGE_SIZE; 449 if (end <= _REGION3_SIZE) { 450 table = base_crst_alloc(_SEGMENT_ENTRY_EMPTY); 451 if (!table) 452 return 0; 453 rc = base_segment_walk(table, addr, end, 1); 454 asce = __pa(table) | _ASCE_TYPE_SEGMENT | _ASCE_TABLE_LENGTH; 455 } else if (end <= _REGION2_SIZE) { 456 table = base_crst_alloc(_REGION3_ENTRY_EMPTY); 457 if (!table) 458 return 0; 459 rc = base_region3_walk(table, addr, end, 1); 460 asce = __pa(table) | _ASCE_TYPE_REGION3 | _ASCE_TABLE_LENGTH; 461 } else if (end <= _REGION1_SIZE) { 462 table = base_crst_alloc(_REGION2_ENTRY_EMPTY); 463 if (!table) 464 return 0; 465 rc = base_region2_walk(table, addr, end, 1); 466 asce = __pa(table) | _ASCE_TYPE_REGION2 | _ASCE_TABLE_LENGTH; 467 } else { 468 table = base_crst_alloc(_REGION1_ENTRY_EMPTY); 469 if (!table) 470 return 0; 471 rc = base_region1_walk(table, addr, end, 1); 472 asce = __pa(table) | _ASCE_TYPE_REGION1 | _ASCE_TABLE_LENGTH; 473 } 474 if (rc) { 475 base_asce_free(asce); 476 asce = 0; 477 } 478 return asce; 479 } 480