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 42 /* change all active ASCEs to avoid the creation of new TLBs */ 43 if (current->active_mm == mm) { 44 get_lowcore()->user_asce.val = mm->context.asce; 45 local_ctl_load(7, &get_lowcore()->user_asce); 46 } 47 __tlb_flush_local(); 48 } 49 50 int crst_table_upgrade(struct mm_struct *mm, unsigned long end) 51 { 52 unsigned long *pgd = NULL, *p4d = NULL, *__pgd; 53 unsigned long asce_limit = mm->context.asce_limit; 54 55 /* upgrade should only happen from 3 to 4, 3 to 5, or 4 to 5 levels */ 56 VM_BUG_ON(asce_limit < _REGION2_SIZE); 57 58 if (end <= asce_limit) 59 return 0; 60 61 if (asce_limit == _REGION2_SIZE) { 62 p4d = crst_table_alloc(mm); 63 if (unlikely(!p4d)) 64 goto err_p4d; 65 crst_table_init(p4d, _REGION2_ENTRY_EMPTY); 66 pagetable_p4d_ctor(virt_to_ptdesc(p4d)); 67 } 68 if (end > _REGION1_SIZE) { 69 pgd = crst_table_alloc(mm); 70 if (unlikely(!pgd)) 71 goto err_pgd; 72 crst_table_init(pgd, _REGION1_ENTRY_EMPTY); 73 pagetable_pgd_ctor(virt_to_ptdesc(pgd)); 74 } 75 76 spin_lock_bh(&mm->page_table_lock); 77 78 /* 79 * This routine gets called with mmap_lock lock held and there is 80 * no reason to optimize for the case of otherwise. However, if 81 * that would ever change, the below check will let us know. 82 */ 83 VM_BUG_ON(asce_limit != mm->context.asce_limit); 84 85 if (p4d) { 86 __pgd = (unsigned long *) mm->pgd; 87 p4d_populate(mm, (p4d_t *) p4d, (pud_t *) __pgd); 88 mm->pgd = (pgd_t *) p4d; 89 mm->context.asce_limit = _REGION1_SIZE; 90 mm->context.asce = __pa(mm->pgd) | _ASCE_TABLE_LENGTH | 91 _ASCE_USER_BITS | _ASCE_TYPE_REGION2; 92 mm_inc_nr_puds(mm); 93 } 94 if (pgd) { 95 __pgd = (unsigned long *) mm->pgd; 96 pgd_populate(mm, (pgd_t *) pgd, (p4d_t *) __pgd); 97 mm->pgd = (pgd_t *) pgd; 98 mm->context.asce_limit = TASK_SIZE_MAX; 99 mm->context.asce = __pa(mm->pgd) | _ASCE_TABLE_LENGTH | 100 _ASCE_USER_BITS | _ASCE_TYPE_REGION1; 101 } 102 103 spin_unlock_bh(&mm->page_table_lock); 104 105 on_each_cpu(__crst_table_upgrade, mm, 0); 106 107 return 0; 108 109 err_pgd: 110 pagetable_dtor(virt_to_ptdesc(p4d)); 111 crst_table_free(mm, p4d); 112 err_p4d: 113 return -ENOMEM; 114 } 115 116 #ifdef CONFIG_PGSTE 117 118 struct ptdesc *page_table_alloc_pgste(struct mm_struct *mm) 119 { 120 struct ptdesc *ptdesc; 121 u64 *table; 122 123 ptdesc = pagetable_alloc(GFP_KERNEL, 0); 124 if (ptdesc) { 125 table = (u64 *)ptdesc_to_virt(ptdesc); 126 __arch_set_page_dat(table, 1); 127 memset64(table, _PAGE_INVALID, PTRS_PER_PTE); 128 memset64(table + PTRS_PER_PTE, 0, PTRS_PER_PTE); 129 } 130 return ptdesc; 131 } 132 133 void page_table_free_pgste(struct ptdesc *ptdesc) 134 { 135 pagetable_free(ptdesc); 136 } 137 138 #endif /* CONFIG_PGSTE */ 139 140 unsigned long *page_table_alloc(struct mm_struct *mm) 141 { 142 struct ptdesc *ptdesc; 143 unsigned long *table; 144 145 ptdesc = pagetable_alloc(GFP_KERNEL, 0); 146 if (!ptdesc) 147 return NULL; 148 if (!pagetable_pte_ctor(ptdesc)) { 149 pagetable_free(ptdesc); 150 return NULL; 151 } 152 table = ptdesc_to_virt(ptdesc); 153 __arch_set_page_dat(table, 1); 154 memset64((u64 *)table, _PAGE_INVALID, PTRS_PER_PTE); 155 memset64((u64 *)table + PTRS_PER_PTE, 0, PTRS_PER_PTE); 156 return table; 157 } 158 159 void page_table_free(struct mm_struct *mm, unsigned long *table) 160 { 161 struct ptdesc *ptdesc = virt_to_ptdesc(table); 162 163 pagetable_dtor_free(ptdesc); 164 } 165 166 #ifdef CONFIG_TRANSPARENT_HUGEPAGE 167 static void pte_free_now(struct rcu_head *head) 168 { 169 struct ptdesc *ptdesc = container_of(head, struct ptdesc, pt_rcu_head); 170 171 pagetable_dtor_free(ptdesc); 172 } 173 174 void pte_free_defer(struct mm_struct *mm, pgtable_t pgtable) 175 { 176 struct ptdesc *ptdesc = virt_to_ptdesc(pgtable); 177 178 call_rcu(&ptdesc->pt_rcu_head, pte_free_now); 179 /* 180 * THPs are not allowed for KVM guests. Warn if pgste ever reaches here. 181 * Turn to the generic pte_free_defer() version once gmap is removed. 182 */ 183 WARN_ON_ONCE(mm_has_pgste(mm)); 184 } 185 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */ 186 187 /* 188 * Base infrastructure required to generate basic asces, region, segment, 189 * and page tables that do not make use of enhanced features like EDAT1. 190 */ 191 192 static struct kmem_cache *base_pgt_cache; 193 194 static unsigned long *base_pgt_alloc(void) 195 { 196 unsigned long *table; 197 198 table = kmem_cache_alloc(base_pgt_cache, GFP_KERNEL); 199 if (table) 200 memset64((u64 *)table, _PAGE_INVALID, PTRS_PER_PTE); 201 return table; 202 } 203 204 static void base_pgt_free(unsigned long *table) 205 { 206 kmem_cache_free(base_pgt_cache, table); 207 } 208 209 static unsigned long *base_crst_alloc(unsigned long val) 210 { 211 unsigned long *table; 212 struct ptdesc *ptdesc; 213 214 ptdesc = pagetable_alloc(GFP_KERNEL, CRST_ALLOC_ORDER); 215 if (!ptdesc) 216 return NULL; 217 table = ptdesc_address(ptdesc); 218 crst_table_init(table, val); 219 return table; 220 } 221 222 static void base_crst_free(unsigned long *table) 223 { 224 if (!table) 225 return; 226 pagetable_free(virt_to_ptdesc(table)); 227 } 228 229 #define BASE_ADDR_END_FUNC(NAME, SIZE) \ 230 static inline unsigned long base_##NAME##_addr_end(unsigned long addr, \ 231 unsigned long end) \ 232 { \ 233 unsigned long next = (addr + (SIZE)) & ~((SIZE) - 1); \ 234 \ 235 return (next - 1) < (end - 1) ? next : end; \ 236 } 237 238 BASE_ADDR_END_FUNC(page, PAGE_SIZE) 239 BASE_ADDR_END_FUNC(segment, _SEGMENT_SIZE) 240 BASE_ADDR_END_FUNC(region3, _REGION3_SIZE) 241 BASE_ADDR_END_FUNC(region2, _REGION2_SIZE) 242 BASE_ADDR_END_FUNC(region1, _REGION1_SIZE) 243 244 static inline unsigned long base_lra(unsigned long address) 245 { 246 unsigned long real; 247 248 asm volatile( 249 " lra %0,0(%1)\n" 250 : "=d" (real) : "a" (address) : "cc"); 251 return real; 252 } 253 254 static int base_page_walk(unsigned long *origin, unsigned long addr, 255 unsigned long end, int alloc) 256 { 257 unsigned long *pte, next; 258 259 if (!alloc) 260 return 0; 261 pte = origin; 262 pte += (addr & _PAGE_INDEX) >> PAGE_SHIFT; 263 do { 264 next = base_page_addr_end(addr, end); 265 *pte = base_lra(addr); 266 } while (pte++, addr = next, addr < end); 267 return 0; 268 } 269 270 static int base_segment_walk(unsigned long *origin, unsigned long addr, 271 unsigned long end, int alloc) 272 { 273 unsigned long *ste, next, *table; 274 int rc; 275 276 ste = origin; 277 ste += (addr & _SEGMENT_INDEX) >> _SEGMENT_SHIFT; 278 do { 279 next = base_segment_addr_end(addr, end); 280 if (*ste & _SEGMENT_ENTRY_INVALID) { 281 if (!alloc) 282 continue; 283 table = base_pgt_alloc(); 284 if (!table) 285 return -ENOMEM; 286 *ste = __pa(table) | _SEGMENT_ENTRY; 287 } 288 table = __va(*ste & _SEGMENT_ENTRY_ORIGIN); 289 rc = base_page_walk(table, addr, next, alloc); 290 if (rc) 291 return rc; 292 if (!alloc) 293 base_pgt_free(table); 294 cond_resched(); 295 } while (ste++, addr = next, addr < end); 296 return 0; 297 } 298 299 static int base_region3_walk(unsigned long *origin, unsigned long addr, 300 unsigned long end, int alloc) 301 { 302 unsigned long *rtte, next, *table; 303 int rc; 304 305 rtte = origin; 306 rtte += (addr & _REGION3_INDEX) >> _REGION3_SHIFT; 307 do { 308 next = base_region3_addr_end(addr, end); 309 if (*rtte & _REGION_ENTRY_INVALID) { 310 if (!alloc) 311 continue; 312 table = base_crst_alloc(_SEGMENT_ENTRY_EMPTY); 313 if (!table) 314 return -ENOMEM; 315 *rtte = __pa(table) | _REGION3_ENTRY; 316 } 317 table = __va(*rtte & _REGION_ENTRY_ORIGIN); 318 rc = base_segment_walk(table, addr, next, alloc); 319 if (rc) 320 return rc; 321 if (!alloc) 322 base_crst_free(table); 323 } while (rtte++, addr = next, addr < end); 324 return 0; 325 } 326 327 static int base_region2_walk(unsigned long *origin, unsigned long addr, 328 unsigned long end, int alloc) 329 { 330 unsigned long *rste, next, *table; 331 int rc; 332 333 rste = origin; 334 rste += (addr & _REGION2_INDEX) >> _REGION2_SHIFT; 335 do { 336 next = base_region2_addr_end(addr, end); 337 if (*rste & _REGION_ENTRY_INVALID) { 338 if (!alloc) 339 continue; 340 table = base_crst_alloc(_REGION3_ENTRY_EMPTY); 341 if (!table) 342 return -ENOMEM; 343 *rste = __pa(table) | _REGION2_ENTRY; 344 } 345 table = __va(*rste & _REGION_ENTRY_ORIGIN); 346 rc = base_region3_walk(table, addr, next, alloc); 347 if (rc) 348 return rc; 349 if (!alloc) 350 base_crst_free(table); 351 } while (rste++, addr = next, addr < end); 352 return 0; 353 } 354 355 static int base_region1_walk(unsigned long *origin, unsigned long addr, 356 unsigned long end, int alloc) 357 { 358 unsigned long *rfte, next, *table; 359 int rc; 360 361 rfte = origin; 362 rfte += (addr & _REGION1_INDEX) >> _REGION1_SHIFT; 363 do { 364 next = base_region1_addr_end(addr, end); 365 if (*rfte & _REGION_ENTRY_INVALID) { 366 if (!alloc) 367 continue; 368 table = base_crst_alloc(_REGION2_ENTRY_EMPTY); 369 if (!table) 370 return -ENOMEM; 371 *rfte = __pa(table) | _REGION1_ENTRY; 372 } 373 table = __va(*rfte & _REGION_ENTRY_ORIGIN); 374 rc = base_region2_walk(table, addr, next, alloc); 375 if (rc) 376 return rc; 377 if (!alloc) 378 base_crst_free(table); 379 } while (rfte++, addr = next, addr < end); 380 return 0; 381 } 382 383 /** 384 * base_asce_free - free asce and tables returned from base_asce_alloc() 385 * @asce: asce to be freed 386 * 387 * Frees all region, segment, and page tables that were allocated with a 388 * corresponding base_asce_alloc() call. 389 */ 390 void base_asce_free(unsigned long asce) 391 { 392 unsigned long *table = __va(asce & _ASCE_ORIGIN); 393 394 if (!asce) 395 return; 396 switch (asce & _ASCE_TYPE_MASK) { 397 case _ASCE_TYPE_SEGMENT: 398 base_segment_walk(table, 0, _REGION3_SIZE, 0); 399 break; 400 case _ASCE_TYPE_REGION3: 401 base_region3_walk(table, 0, _REGION2_SIZE, 0); 402 break; 403 case _ASCE_TYPE_REGION2: 404 base_region2_walk(table, 0, _REGION1_SIZE, 0); 405 break; 406 case _ASCE_TYPE_REGION1: 407 base_region1_walk(table, 0, TASK_SIZE_MAX, 0); 408 break; 409 } 410 base_crst_free(table); 411 } 412 413 static int base_pgt_cache_init(void) 414 { 415 static DEFINE_MUTEX(base_pgt_cache_mutex); 416 unsigned long sz = _PAGE_TABLE_SIZE; 417 418 if (base_pgt_cache) 419 return 0; 420 mutex_lock(&base_pgt_cache_mutex); 421 if (!base_pgt_cache) 422 base_pgt_cache = kmem_cache_create("base_pgt", sz, sz, 0, NULL); 423 mutex_unlock(&base_pgt_cache_mutex); 424 return base_pgt_cache ? 0 : -ENOMEM; 425 } 426 427 /** 428 * base_asce_alloc - create kernel mapping without enhanced DAT features 429 * @addr: virtual start address of kernel mapping 430 * @num_pages: number of consecutive pages 431 * 432 * Generate an asce, including all required region, segment and page tables, 433 * that can be used to access the virtual kernel mapping. The difference is 434 * that the returned asce does not make use of any enhanced DAT features like 435 * e.g. large pages. This is required for some I/O functions that pass an 436 * asce, like e.g. some service call requests. 437 * 438 * Note: the returned asce may NEVER be attached to any cpu. It may only be 439 * used for I/O requests. tlb entries that might result because the 440 * asce was attached to a cpu won't be cleared. 441 */ 442 unsigned long base_asce_alloc(unsigned long addr, unsigned long num_pages) 443 { 444 unsigned long asce, *table, end; 445 int rc; 446 447 if (base_pgt_cache_init()) 448 return 0; 449 end = addr + num_pages * PAGE_SIZE; 450 if (end <= _REGION3_SIZE) { 451 table = base_crst_alloc(_SEGMENT_ENTRY_EMPTY); 452 if (!table) 453 return 0; 454 rc = base_segment_walk(table, addr, end, 1); 455 asce = __pa(table) | _ASCE_TYPE_SEGMENT | _ASCE_TABLE_LENGTH; 456 } else if (end <= _REGION2_SIZE) { 457 table = base_crst_alloc(_REGION3_ENTRY_EMPTY); 458 if (!table) 459 return 0; 460 rc = base_region3_walk(table, addr, end, 1); 461 asce = __pa(table) | _ASCE_TYPE_REGION3 | _ASCE_TABLE_LENGTH; 462 } else if (end <= _REGION1_SIZE) { 463 table = base_crst_alloc(_REGION2_ENTRY_EMPTY); 464 if (!table) 465 return 0; 466 rc = base_region2_walk(table, addr, end, 1); 467 asce = __pa(table) | _ASCE_TYPE_REGION2 | _ASCE_TABLE_LENGTH; 468 } else { 469 table = base_crst_alloc(_REGION1_ENTRY_EMPTY); 470 if (!table) 471 return 0; 472 rc = base_region1_walk(table, addr, end, 1); 473 asce = __pa(table) | _ASCE_TYPE_REGION1 | _ASCE_TABLE_LENGTH; 474 } 475 if (rc) { 476 base_asce_free(asce); 477 asce = 0; 478 } 479 return asce; 480 } 481