1 // SPDX-License-Identifier: GPL-2.0 2 /* 3 * Copyright IBM Corp. 2006 4 */ 5 6 #include <linux/memory_hotplug.h> 7 #include <linux/memblock.h> 8 #include <linux/pfn.h> 9 #include <linux/mm.h> 10 #include <linux/init.h> 11 #include <linux/list.h> 12 #include <linux/hugetlb.h> 13 #include <linux/slab.h> 14 #include <asm/cacheflush.h> 15 #include <asm/nospec-branch.h> 16 #include <asm/pgalloc.h> 17 #include <asm/setup.h> 18 #include <asm/tlbflush.h> 19 #include <asm/sections.h> 20 #include <asm/set_memory.h> 21 22 static DEFINE_MUTEX(vmem_mutex); 23 24 static void __ref *vmem_alloc_pages(unsigned int order) 25 { 26 unsigned long size = PAGE_SIZE << order; 27 28 if (slab_is_available()) 29 return (void *)__get_free_pages(GFP_KERNEL, order); 30 return memblock_alloc(size, size); 31 } 32 33 static void vmem_free_pages(unsigned long addr, int order) 34 { 35 /* We don't expect boot memory to be removed ever. */ 36 if (!slab_is_available() || 37 WARN_ON_ONCE(PageReserved(virt_to_page(addr)))) 38 return; 39 free_pages(addr, order); 40 } 41 42 void *vmem_crst_alloc(unsigned long val) 43 { 44 unsigned long *table; 45 46 table = vmem_alloc_pages(CRST_ALLOC_ORDER); 47 if (table) 48 crst_table_init(table, val); 49 return table; 50 } 51 52 pte_t __ref *vmem_pte_alloc(void) 53 { 54 unsigned long size = PTRS_PER_PTE * sizeof(pte_t); 55 pte_t *pte; 56 57 if (slab_is_available()) 58 pte = (pte_t *) page_table_alloc(&init_mm); 59 else 60 pte = (pte_t *) memblock_alloc(size, size); 61 if (!pte) 62 return NULL; 63 memset64((u64 *)pte, _PAGE_INVALID, PTRS_PER_PTE); 64 return pte; 65 } 66 67 static void vmem_pte_free(unsigned long *table) 68 { 69 /* We don't expect boot memory to be removed ever. */ 70 if (!slab_is_available() || 71 WARN_ON_ONCE(PageReserved(virt_to_page(table)))) 72 return; 73 page_table_free(&init_mm, table); 74 } 75 76 #define PAGE_UNUSED 0xFD 77 78 /* 79 * The unused vmemmap range, which was not yet memset(PAGE_UNUSED) ranges 80 * from unused_sub_pmd_start to next PMD_SIZE boundary. 81 */ 82 static unsigned long unused_sub_pmd_start; 83 84 static void vmemmap_flush_unused_sub_pmd(void) 85 { 86 if (!unused_sub_pmd_start) 87 return; 88 memset((void *)unused_sub_pmd_start, PAGE_UNUSED, 89 ALIGN(unused_sub_pmd_start, PMD_SIZE) - unused_sub_pmd_start); 90 unused_sub_pmd_start = 0; 91 } 92 93 static void vmemmap_mark_sub_pmd_used(unsigned long start, unsigned long end) 94 { 95 /* 96 * As we expect to add in the same granularity as we remove, it's 97 * sufficient to mark only some piece used to block the memmap page from 98 * getting removed (just in case the memmap never gets initialized, 99 * e.g., because the memory block never gets onlined). 100 */ 101 memset((void *)start, 0, sizeof(struct page)); 102 } 103 104 static void vmemmap_use_sub_pmd(unsigned long start, unsigned long end) 105 { 106 /* 107 * We only optimize if the new used range directly follows the 108 * previously unused range (esp., when populating consecutive sections). 109 */ 110 if (unused_sub_pmd_start == start) { 111 unused_sub_pmd_start = end; 112 if (likely(IS_ALIGNED(unused_sub_pmd_start, PMD_SIZE))) 113 unused_sub_pmd_start = 0; 114 return; 115 } 116 vmemmap_flush_unused_sub_pmd(); 117 vmemmap_mark_sub_pmd_used(start, end); 118 } 119 120 static void vmemmap_use_new_sub_pmd(unsigned long start, unsigned long end) 121 { 122 unsigned long page = ALIGN_DOWN(start, PMD_SIZE); 123 124 vmemmap_flush_unused_sub_pmd(); 125 126 /* Could be our memmap page is filled with PAGE_UNUSED already ... */ 127 vmemmap_mark_sub_pmd_used(start, end); 128 129 /* Mark the unused parts of the new memmap page PAGE_UNUSED. */ 130 if (!IS_ALIGNED(start, PMD_SIZE)) 131 memset((void *)page, PAGE_UNUSED, start - page); 132 /* 133 * We want to avoid memset(PAGE_UNUSED) when populating the vmemmap of 134 * consecutive sections. Remember for the last added PMD the last 135 * unused range in the populated PMD. 136 */ 137 if (!IS_ALIGNED(end, PMD_SIZE)) 138 unused_sub_pmd_start = end; 139 } 140 141 /* Returns true if the PMD is completely unused and can be freed. */ 142 static bool vmemmap_unuse_sub_pmd(unsigned long start, unsigned long end) 143 { 144 unsigned long page = ALIGN_DOWN(start, PMD_SIZE); 145 146 vmemmap_flush_unused_sub_pmd(); 147 memset((void *)start, PAGE_UNUSED, end - start); 148 return !memchr_inv((void *)page, PAGE_UNUSED, PMD_SIZE); 149 } 150 151 /* __ref: we'll only call vmemmap_alloc_block() via vmemmap_populate() */ 152 static int __ref modify_pte_table(pmd_t *pmd, unsigned long addr, 153 unsigned long end, bool add, bool direct) 154 { 155 unsigned long prot, pages = 0; 156 int ret = -ENOMEM; 157 pte_t *pte; 158 159 prot = pgprot_val(PAGE_KERNEL); 160 if (!MACHINE_HAS_NX) 161 prot &= ~_PAGE_NOEXEC; 162 163 pte = pte_offset_kernel(pmd, addr); 164 for (; addr < end; addr += PAGE_SIZE, pte++) { 165 if (!add) { 166 if (pte_none(*pte)) 167 continue; 168 if (!direct) 169 vmem_free_pages((unsigned long) pfn_to_virt(pte_pfn(*pte)), 0); 170 pte_clear(&init_mm, addr, pte); 171 } else if (pte_none(*pte)) { 172 if (!direct) { 173 void *new_page = vmemmap_alloc_block(PAGE_SIZE, NUMA_NO_NODE); 174 175 if (!new_page) 176 goto out; 177 set_pte(pte, __pte(__pa(new_page) | prot)); 178 } else { 179 set_pte(pte, __pte(__pa(addr) | prot)); 180 } 181 } else { 182 continue; 183 } 184 pages++; 185 } 186 ret = 0; 187 out: 188 if (direct) 189 update_page_count(PG_DIRECT_MAP_4K, add ? pages : -pages); 190 return ret; 191 } 192 193 static void try_free_pte_table(pmd_t *pmd, unsigned long start) 194 { 195 pte_t *pte; 196 int i; 197 198 /* We can safely assume this is fully in 1:1 mapping & vmemmap area */ 199 pte = pte_offset_kernel(pmd, start); 200 for (i = 0; i < PTRS_PER_PTE; i++, pte++) { 201 if (!pte_none(*pte)) 202 return; 203 } 204 vmem_pte_free((unsigned long *) pmd_deref(*pmd)); 205 pmd_clear(pmd); 206 } 207 208 /* __ref: we'll only call vmemmap_alloc_block() via vmemmap_populate() */ 209 static int __ref modify_pmd_table(pud_t *pud, unsigned long addr, 210 unsigned long end, bool add, bool direct) 211 { 212 unsigned long next, prot, pages = 0; 213 int ret = -ENOMEM; 214 pmd_t *pmd; 215 pte_t *pte; 216 217 prot = pgprot_val(SEGMENT_KERNEL); 218 if (!MACHINE_HAS_NX) 219 prot &= ~_SEGMENT_ENTRY_NOEXEC; 220 221 pmd = pmd_offset(pud, addr); 222 for (; addr < end; addr = next, pmd++) { 223 next = pmd_addr_end(addr, end); 224 if (!add) { 225 if (pmd_none(*pmd)) 226 continue; 227 if (pmd_large(*pmd)) { 228 if (IS_ALIGNED(addr, PMD_SIZE) && 229 IS_ALIGNED(next, PMD_SIZE)) { 230 if (!direct) 231 vmem_free_pages(pmd_deref(*pmd), get_order(PMD_SIZE)); 232 pmd_clear(pmd); 233 pages++; 234 } else if (!direct && vmemmap_unuse_sub_pmd(addr, next)) { 235 vmem_free_pages(pmd_deref(*pmd), get_order(PMD_SIZE)); 236 pmd_clear(pmd); 237 } 238 continue; 239 } 240 } else if (pmd_none(*pmd)) { 241 if (IS_ALIGNED(addr, PMD_SIZE) && 242 IS_ALIGNED(next, PMD_SIZE) && 243 MACHINE_HAS_EDAT1 && direct && 244 !debug_pagealloc_enabled()) { 245 set_pmd(pmd, __pmd(__pa(addr) | prot)); 246 pages++; 247 continue; 248 } else if (!direct && MACHINE_HAS_EDAT1) { 249 void *new_page; 250 251 /* 252 * Use 1MB frames for vmemmap if available. We 253 * always use large frames even if they are only 254 * partially used. Otherwise we would have also 255 * page tables since vmemmap_populate gets 256 * called for each section separately. 257 */ 258 new_page = vmemmap_alloc_block(PMD_SIZE, NUMA_NO_NODE); 259 if (new_page) { 260 set_pmd(pmd, __pmd(__pa(new_page) | prot)); 261 if (!IS_ALIGNED(addr, PMD_SIZE) || 262 !IS_ALIGNED(next, PMD_SIZE)) { 263 vmemmap_use_new_sub_pmd(addr, next); 264 } 265 continue; 266 } 267 } 268 pte = vmem_pte_alloc(); 269 if (!pte) 270 goto out; 271 pmd_populate(&init_mm, pmd, pte); 272 } else if (pmd_large(*pmd)) { 273 if (!direct) 274 vmemmap_use_sub_pmd(addr, next); 275 continue; 276 } 277 ret = modify_pte_table(pmd, addr, next, add, direct); 278 if (ret) 279 goto out; 280 if (!add) 281 try_free_pte_table(pmd, addr & PMD_MASK); 282 } 283 ret = 0; 284 out: 285 if (direct) 286 update_page_count(PG_DIRECT_MAP_1M, add ? pages : -pages); 287 return ret; 288 } 289 290 static void try_free_pmd_table(pud_t *pud, unsigned long start) 291 { 292 const unsigned long end = start + PUD_SIZE; 293 pmd_t *pmd; 294 int i; 295 296 /* Don't mess with any tables not fully in 1:1 mapping & vmemmap area */ 297 if (end > VMALLOC_START) 298 return; 299 #ifdef CONFIG_KASAN 300 if (start < KASAN_SHADOW_END && KASAN_SHADOW_START > end) 301 return; 302 #endif 303 pmd = pmd_offset(pud, start); 304 for (i = 0; i < PTRS_PER_PMD; i++, pmd++) 305 if (!pmd_none(*pmd)) 306 return; 307 vmem_free_pages(pud_deref(*pud), CRST_ALLOC_ORDER); 308 pud_clear(pud); 309 } 310 311 static int modify_pud_table(p4d_t *p4d, unsigned long addr, unsigned long end, 312 bool add, bool direct) 313 { 314 unsigned long next, prot, pages = 0; 315 int ret = -ENOMEM; 316 pud_t *pud; 317 pmd_t *pmd; 318 319 prot = pgprot_val(REGION3_KERNEL); 320 if (!MACHINE_HAS_NX) 321 prot &= ~_REGION_ENTRY_NOEXEC; 322 pud = pud_offset(p4d, addr); 323 for (; addr < end; addr = next, pud++) { 324 next = pud_addr_end(addr, end); 325 if (!add) { 326 if (pud_none(*pud)) 327 continue; 328 if (pud_large(*pud)) { 329 if (IS_ALIGNED(addr, PUD_SIZE) && 330 IS_ALIGNED(next, PUD_SIZE)) { 331 pud_clear(pud); 332 pages++; 333 } 334 continue; 335 } 336 } else if (pud_none(*pud)) { 337 if (IS_ALIGNED(addr, PUD_SIZE) && 338 IS_ALIGNED(next, PUD_SIZE) && 339 MACHINE_HAS_EDAT2 && direct && 340 !debug_pagealloc_enabled()) { 341 set_pud(pud, __pud(__pa(addr) | prot)); 342 pages++; 343 continue; 344 } 345 pmd = vmem_crst_alloc(_SEGMENT_ENTRY_EMPTY); 346 if (!pmd) 347 goto out; 348 pud_populate(&init_mm, pud, pmd); 349 } else if (pud_large(*pud)) { 350 continue; 351 } 352 ret = modify_pmd_table(pud, addr, next, add, direct); 353 if (ret) 354 goto out; 355 if (!add) 356 try_free_pmd_table(pud, addr & PUD_MASK); 357 } 358 ret = 0; 359 out: 360 if (direct) 361 update_page_count(PG_DIRECT_MAP_2G, add ? pages : -pages); 362 return ret; 363 } 364 365 static void try_free_pud_table(p4d_t *p4d, unsigned long start) 366 { 367 const unsigned long end = start + P4D_SIZE; 368 pud_t *pud; 369 int i; 370 371 /* Don't mess with any tables not fully in 1:1 mapping & vmemmap area */ 372 if (end > VMALLOC_START) 373 return; 374 #ifdef CONFIG_KASAN 375 if (start < KASAN_SHADOW_END && KASAN_SHADOW_START > end) 376 return; 377 #endif 378 379 pud = pud_offset(p4d, start); 380 for (i = 0; i < PTRS_PER_PUD; i++, pud++) { 381 if (!pud_none(*pud)) 382 return; 383 } 384 vmem_free_pages(p4d_deref(*p4d), CRST_ALLOC_ORDER); 385 p4d_clear(p4d); 386 } 387 388 static int modify_p4d_table(pgd_t *pgd, unsigned long addr, unsigned long end, 389 bool add, bool direct) 390 { 391 unsigned long next; 392 int ret = -ENOMEM; 393 p4d_t *p4d; 394 pud_t *pud; 395 396 p4d = p4d_offset(pgd, addr); 397 for (; addr < end; addr = next, p4d++) { 398 next = p4d_addr_end(addr, end); 399 if (!add) { 400 if (p4d_none(*p4d)) 401 continue; 402 } else if (p4d_none(*p4d)) { 403 pud = vmem_crst_alloc(_REGION3_ENTRY_EMPTY); 404 if (!pud) 405 goto out; 406 p4d_populate(&init_mm, p4d, pud); 407 } 408 ret = modify_pud_table(p4d, addr, next, add, direct); 409 if (ret) 410 goto out; 411 if (!add) 412 try_free_pud_table(p4d, addr & P4D_MASK); 413 } 414 ret = 0; 415 out: 416 return ret; 417 } 418 419 static void try_free_p4d_table(pgd_t *pgd, unsigned long start) 420 { 421 const unsigned long end = start + PGDIR_SIZE; 422 p4d_t *p4d; 423 int i; 424 425 /* Don't mess with any tables not fully in 1:1 mapping & vmemmap area */ 426 if (end > VMALLOC_START) 427 return; 428 #ifdef CONFIG_KASAN 429 if (start < KASAN_SHADOW_END && KASAN_SHADOW_START > end) 430 return; 431 #endif 432 433 p4d = p4d_offset(pgd, start); 434 for (i = 0; i < PTRS_PER_P4D; i++, p4d++) { 435 if (!p4d_none(*p4d)) 436 return; 437 } 438 vmem_free_pages(pgd_deref(*pgd), CRST_ALLOC_ORDER); 439 pgd_clear(pgd); 440 } 441 442 static int modify_pagetable(unsigned long start, unsigned long end, bool add, 443 bool direct) 444 { 445 unsigned long addr, next; 446 int ret = -ENOMEM; 447 pgd_t *pgd; 448 p4d_t *p4d; 449 450 if (WARN_ON_ONCE(!PAGE_ALIGNED(start | end))) 451 return -EINVAL; 452 for (addr = start; addr < end; addr = next) { 453 next = pgd_addr_end(addr, end); 454 pgd = pgd_offset_k(addr); 455 456 if (!add) { 457 if (pgd_none(*pgd)) 458 continue; 459 } else if (pgd_none(*pgd)) { 460 p4d = vmem_crst_alloc(_REGION2_ENTRY_EMPTY); 461 if (!p4d) 462 goto out; 463 pgd_populate(&init_mm, pgd, p4d); 464 } 465 ret = modify_p4d_table(pgd, addr, next, add, direct); 466 if (ret) 467 goto out; 468 if (!add) 469 try_free_p4d_table(pgd, addr & PGDIR_MASK); 470 } 471 ret = 0; 472 out: 473 if (!add) 474 flush_tlb_kernel_range(start, end); 475 return ret; 476 } 477 478 static int add_pagetable(unsigned long start, unsigned long end, bool direct) 479 { 480 return modify_pagetable(start, end, true, direct); 481 } 482 483 static int remove_pagetable(unsigned long start, unsigned long end, bool direct) 484 { 485 return modify_pagetable(start, end, false, direct); 486 } 487 488 /* 489 * Add a physical memory range to the 1:1 mapping. 490 */ 491 static int vmem_add_range(unsigned long start, unsigned long size) 492 { 493 return add_pagetable(start, start + size, true); 494 } 495 496 /* 497 * Remove a physical memory range from the 1:1 mapping. 498 */ 499 static void vmem_remove_range(unsigned long start, unsigned long size) 500 { 501 remove_pagetable(start, start + size, true); 502 } 503 504 /* 505 * Add a backed mem_map array to the virtual mem_map array. 506 */ 507 int __meminit vmemmap_populate(unsigned long start, unsigned long end, int node, 508 struct vmem_altmap *altmap) 509 { 510 int ret; 511 512 mutex_lock(&vmem_mutex); 513 /* We don't care about the node, just use NUMA_NO_NODE on allocations */ 514 ret = add_pagetable(start, end, false); 515 if (ret) 516 remove_pagetable(start, end, false); 517 mutex_unlock(&vmem_mutex); 518 return ret; 519 } 520 521 void vmemmap_free(unsigned long start, unsigned long end, 522 struct vmem_altmap *altmap) 523 { 524 mutex_lock(&vmem_mutex); 525 remove_pagetable(start, end, false); 526 mutex_unlock(&vmem_mutex); 527 } 528 529 void vmem_remove_mapping(unsigned long start, unsigned long size) 530 { 531 mutex_lock(&vmem_mutex); 532 vmem_remove_range(start, size); 533 mutex_unlock(&vmem_mutex); 534 } 535 536 struct range arch_get_mappable_range(void) 537 { 538 struct range mhp_range; 539 540 mhp_range.start = 0; 541 mhp_range.end = VMEM_MAX_PHYS - 1; 542 return mhp_range; 543 } 544 545 int vmem_add_mapping(unsigned long start, unsigned long size) 546 { 547 struct range range = arch_get_mappable_range(); 548 int ret; 549 550 if (start < range.start || 551 start + size > range.end + 1 || 552 start + size < start) 553 return -ERANGE; 554 555 mutex_lock(&vmem_mutex); 556 ret = vmem_add_range(start, size); 557 if (ret) 558 vmem_remove_range(start, size); 559 mutex_unlock(&vmem_mutex); 560 return ret; 561 } 562 563 /* 564 * Allocate new or return existing page-table entry, but do not map it 565 * to any physical address. If missing, allocate segment- and region- 566 * table entries along. Meeting a large segment- or region-table entry 567 * while traversing is an error, since the function is expected to be 568 * called against virtual regions reserverd for 4KB mappings only. 569 */ 570 pte_t *vmem_get_alloc_pte(unsigned long addr, bool alloc) 571 { 572 pte_t *ptep = NULL; 573 pgd_t *pgd; 574 p4d_t *p4d; 575 pud_t *pud; 576 pmd_t *pmd; 577 pte_t *pte; 578 579 pgd = pgd_offset_k(addr); 580 if (pgd_none(*pgd)) { 581 if (!alloc) 582 goto out; 583 p4d = vmem_crst_alloc(_REGION2_ENTRY_EMPTY); 584 if (!p4d) 585 goto out; 586 pgd_populate(&init_mm, pgd, p4d); 587 } 588 p4d = p4d_offset(pgd, addr); 589 if (p4d_none(*p4d)) { 590 if (!alloc) 591 goto out; 592 pud = vmem_crst_alloc(_REGION3_ENTRY_EMPTY); 593 if (!pud) 594 goto out; 595 p4d_populate(&init_mm, p4d, pud); 596 } 597 pud = pud_offset(p4d, addr); 598 if (pud_none(*pud)) { 599 if (!alloc) 600 goto out; 601 pmd = vmem_crst_alloc(_SEGMENT_ENTRY_EMPTY); 602 if (!pmd) 603 goto out; 604 pud_populate(&init_mm, pud, pmd); 605 } else if (WARN_ON_ONCE(pud_large(*pud))) { 606 goto out; 607 } 608 pmd = pmd_offset(pud, addr); 609 if (pmd_none(*pmd)) { 610 if (!alloc) 611 goto out; 612 pte = vmem_pte_alloc(); 613 if (!pte) 614 goto out; 615 pmd_populate(&init_mm, pmd, pte); 616 } else if (WARN_ON_ONCE(pmd_large(*pmd))) { 617 goto out; 618 } 619 ptep = pte_offset_kernel(pmd, addr); 620 out: 621 return ptep; 622 } 623 624 int __vmem_map_4k_page(unsigned long addr, unsigned long phys, pgprot_t prot, bool alloc) 625 { 626 pte_t *ptep, pte; 627 628 if (!IS_ALIGNED(addr, PAGE_SIZE)) 629 return -EINVAL; 630 ptep = vmem_get_alloc_pte(addr, alloc); 631 if (!ptep) 632 return -ENOMEM; 633 __ptep_ipte(addr, ptep, 0, 0, IPTE_GLOBAL); 634 pte = mk_pte_phys(phys, prot); 635 set_pte(ptep, pte); 636 return 0; 637 } 638 639 int vmem_map_4k_page(unsigned long addr, unsigned long phys, pgprot_t prot) 640 { 641 int rc; 642 643 mutex_lock(&vmem_mutex); 644 rc = __vmem_map_4k_page(addr, phys, prot, true); 645 mutex_unlock(&vmem_mutex); 646 return rc; 647 } 648 649 void vmem_unmap_4k_page(unsigned long addr) 650 { 651 pte_t *ptep; 652 653 mutex_lock(&vmem_mutex); 654 ptep = virt_to_kpte(addr); 655 __ptep_ipte(addr, ptep, 0, 0, IPTE_GLOBAL); 656 pte_clear(&init_mm, addr, ptep); 657 mutex_unlock(&vmem_mutex); 658 } 659 660 /* 661 * map whole physical memory to virtual memory (identity mapping) 662 * we reserve enough space in the vmalloc area for vmemmap to hotplug 663 * additional memory segments. 664 */ 665 void __init vmem_map_init(void) 666 { 667 phys_addr_t base, end; 668 u64 i; 669 670 for_each_mem_range(i, &base, &end) 671 vmem_add_range(base, end - base); 672 __set_memory((unsigned long)_stext, 673 (unsigned long)(_etext - _stext) >> PAGE_SHIFT, 674 SET_MEMORY_RO | SET_MEMORY_X); 675 __set_memory((unsigned long)_etext, 676 (unsigned long)(__end_rodata - _etext) >> PAGE_SHIFT, 677 SET_MEMORY_RO); 678 __set_memory((unsigned long)_sinittext, 679 (unsigned long)(_einittext - _sinittext) >> PAGE_SHIFT, 680 SET_MEMORY_RO | SET_MEMORY_X); 681 __set_memory(__stext_amode31, (__etext_amode31 - __stext_amode31) >> PAGE_SHIFT, 682 SET_MEMORY_RO | SET_MEMORY_X); 683 684 /* lowcore requires 4k mapping for real addresses / prefixing */ 685 set_memory_4k(0, LC_PAGES); 686 687 /* lowcore must be executable for LPSWE */ 688 if (!static_key_enabled(&cpu_has_bear)) 689 set_memory_x(0, 1); 690 691 pr_info("Write protected kernel read-only data: %luk\n", 692 (unsigned long)(__end_rodata - _stext) >> 10); 693 } 694