arch/powerpc/mm/book3s64/radix_pgtable.c

   1 // SPDX-License-Identifier: GPL-2.0-or-later
   2 /*
   3  * Page table handling routines for radix page table.
   4  *
   5  * Copyright 2015-2016, Aneesh Kumar K.V, IBM Corporation.
   6  */
   7
   8 #define pr_fmt(fmt) "radix-mmu: " fmt
   9
  10 #include <linux/io.h>
  11 #include <linux/kernel.h>
  12 #include <linux/sched/mm.h>
  13 #include <linux/memblock.h>
  14 #include <linux/of.h>
  15 #include <linux/of_fdt.h>
  16 #include <linux/mm.h>
  17 #include <linux/hugetlb.h>
  18 #include <linux/string_helpers.h>
  19 #include <linux/memory.h>
  20
  21 #include <asm/pgalloc.h>
  22 #include <asm/mmu_context.h>
  23 #include <asm/dma.h>
  24 #include <asm/machdep.h>
  25 #include <asm/mmu.h>
  26 #include <asm/firmware.h>
  27 #include <asm/powernv.h>
  28 #include <asm/sections.h>
  29 #include <asm/smp.h>
  30 #include <asm/trace.h>
  31 #include <asm/uaccess.h>
  32 #include <asm/ultravisor.h>
  33 #include <asm/set_memory.h>
  34
  35 #include <trace/events/thp.h>
  36
  37 #include <mm/mmu_decl.h>
  38
  39 unsigned int mmu_base_pid;
  40
  41 static __ref void *early_alloc_pgtable(unsigned long size, int nid,
  42                         unsigned long region_start, unsigned long region_end)
  43 {
  44         phys_addr_t min_addr = MEMBLOCK_LOW_LIMIT;
  45         phys_addr_t max_addr = MEMBLOCK_ALLOC_ANYWHERE;
  46         void *ptr;
  47
  48         if (region_start)
  49                 min_addr = region_start;
  50         if (region_end)
  51                 max_addr = region_end;
  52
  53         ptr = memblock_alloc_try_nid(size, size, min_addr, max_addr, nid);
  54
  55         if (!ptr)
  56                 panic("%s: Failed to allocate %lu bytes align=0x%lx nid=%d from=%pa max_addr=%pa\n",
  57                       __func__, size, size, nid, &min_addr, &max_addr);
  58
  59         return ptr;
  60 }
  61
  62 /*
  63  * When allocating pud or pmd pointers, we allocate a complete page
  64  * of PAGE_SIZE rather than PUD_TABLE_SIZE or PMD_TABLE_SIZE. This
  65  * is to ensure that the page obtained from the memblock allocator
  66  * can be completely used as page table page and can be freed
  67  * correctly when the page table entries are removed.
  68  */
  69 static int early_map_kernel_page(unsigned long ea, unsigned long pa,
  70                           pgprot_t flags,
  71                           unsigned int map_page_size,
  72                           int nid,
  73                           unsigned long region_start, unsigned long region_end)
  74 {
  75         unsigned long pfn = pa >> PAGE_SHIFT;
  76         pgd_t *pgdp;
  77         p4d_t *p4dp;
  78         pud_t *pudp;
  79         pmd_t *pmdp;
  80         pte_t *ptep;
  81
  82         pgdp = pgd_offset_k(ea);
  83         p4dp = p4d_offset(pgdp, ea);
  84         if (p4d_none(*p4dp)) {
  85                 pudp = early_alloc_pgtable(PAGE_SIZE, nid,
  86                                            region_start, region_end);
  87                 p4d_populate(&init_mm, p4dp, pudp);
  88         }
  89         pudp = pud_offset(p4dp, ea);
  90         if (map_page_size == PUD_SIZE) {
  91                 ptep = (pte_t *)pudp;
  92                 goto set_the_pte;
  93         }
  94         if (pud_none(*pudp)) {
  95                 pmdp = early_alloc_pgtable(PAGE_SIZE, nid, region_start,
  96                                            region_end);
  97                 pud_populate(&init_mm, pudp, pmdp);
  98         }
  99         pmdp = pmd_offset(pudp, ea);
 100         if (map_page_size == PMD_SIZE) {
 101                 ptep = pmdp_ptep(pmdp);
 102                 goto set_the_pte;
 103         }
 104         if (!pmd_present(*pmdp)) {
 105                 ptep = early_alloc_pgtable(PAGE_SIZE, nid,
 106                                                 region_start, region_end);
 107                 pmd_populate_kernel(&init_mm, pmdp, ptep);
 108         }
 109         ptep = pte_offset_kernel(pmdp, ea);
 110
 111 set_the_pte:
 112         set_pte_at(&init_mm, ea, ptep, pfn_pte(pfn, flags));
 113         asm volatile("ptesync": : :"memory");
 114         return 0;
 115 }
 116
 117 /*
 118  * nid, region_start, and region_end are hints to try to place the page
 119  * table memory in the same node or region.
 120  */
 121 static int __map_kernel_page(unsigned long ea, unsigned long pa,
 122                           pgprot_t flags,
 123                           unsigned int map_page_size,
 124                           int nid,
 125                           unsigned long region_start, unsigned long region_end)
 126 {
 127         unsigned long pfn = pa >> PAGE_SHIFT;
 128         pgd_t *pgdp;
 129         p4d_t *p4dp;
 130         pud_t *pudp;
 131         pmd_t *pmdp;
 132         pte_t *ptep;
 133         /*
 134          * Make sure task size is correct as per the max adddr
 135          */
 136         BUILD_BUG_ON(TASK_SIZE_USER64 > RADIX_PGTABLE_RANGE);
 137
 138 #ifdef CONFIG_PPC_64K_PAGES
 139         BUILD_BUG_ON(RADIX_KERN_MAP_SIZE != (1UL << MAX_EA_BITS_PER_CONTEXT));
 140 #endif
 141
 142         if (unlikely(!slab_is_available()))
 143                 return early_map_kernel_page(ea, pa, flags, map_page_size,
 144                                                 nid, region_start, region_end);
 145
 146         /*
 147          * Should make page table allocation functions be able to take a
 148          * node, so we can place kernel page tables on the right nodes after
 149          * boot.
 150          */
 151         pgdp = pgd_offset_k(ea);
 152         p4dp = p4d_offset(pgdp, ea);
 153         pudp = pud_alloc(&init_mm, p4dp, ea);
 154         if (!pudp)
 155                 return -ENOMEM;
 156         if (map_page_size == PUD_SIZE) {
 157                 ptep = (pte_t *)pudp;
 158                 goto set_the_pte;
 159         }
 160         pmdp = pmd_alloc(&init_mm, pudp, ea);
 161         if (!pmdp)
 162                 return -ENOMEM;
 163         if (map_page_size == PMD_SIZE) {
 164                 ptep = pmdp_ptep(pmdp);
 165                 goto set_the_pte;
 166         }
 167         ptep = pte_alloc_kernel(pmdp, ea);
 168         if (!ptep)
 169                 return -ENOMEM;
 170
 171 set_the_pte:
 172         set_pte_at(&init_mm, ea, ptep, pfn_pte(pfn, flags));
 173         asm volatile("ptesync": : :"memory");
 174         return 0;
 175 }
 176
 177 int radix__map_kernel_page(unsigned long ea, unsigned long pa,
 178                           pgprot_t flags,
 179                           unsigned int map_page_size)
 180 {
 181         return __map_kernel_page(ea, pa, flags, map_page_size, -1, 0, 0);
 182 }
 183
 184 #ifdef CONFIG_STRICT_KERNEL_RWX
 185 static void radix__change_memory_range(unsigned long start, unsigned long end,
 186                                        unsigned long clear)
 187 {
 188         unsigned long idx;
 189         pgd_t *pgdp;
 190         p4d_t *p4dp;
 191         pud_t *pudp;
 192         pmd_t *pmdp;
 193         pte_t *ptep;
 194
 195         start = ALIGN_DOWN(start, PAGE_SIZE);
 196         end = PAGE_ALIGN(end); // aligns up
 197
 198         pr_debug("Changing flags on range %lx-%lx removing 0x%lx\n",
 199                  start, end, clear);
 200
 201         for (idx = start; idx < end; idx += PAGE_SIZE) {
 202                 pgdp = pgd_offset_k(idx);
 203                 p4dp = p4d_offset(pgdp, idx);
 204                 pudp = pud_alloc(&init_mm, p4dp, idx);
 205                 if (!pudp)
 206                         continue;
 207                 if (pud_leaf(*pudp)) {
 208                         ptep = (pte_t *)pudp;
 209                         goto update_the_pte;
 210                 }
 211                 pmdp = pmd_alloc(&init_mm, pudp, idx);
 212                 if (!pmdp)
 213                         continue;
 214                 if (pmd_leaf(*pmdp)) {
 215                         ptep = pmdp_ptep(pmdp);
 216                         goto update_the_pte;
 217                 }
 218                 ptep = pte_alloc_kernel(pmdp, idx);
 219                 if (!ptep)
 220                         continue;
 221 update_the_pte:
 222                 radix__pte_update(&init_mm, idx, ptep, clear, 0, 0);
 223         }
 224
 225         radix__flush_tlb_kernel_range(start, end);
 226 }
 227
 228 void radix__mark_rodata_ro(void)
 229 {
 230         unsigned long start, end;
 231
 232         start = (unsigned long)_stext;
 233         end = (unsigned long)__end_rodata;
 234
 235         radix__change_memory_range(start, end, _PAGE_WRITE);
 236
 237         for (start = PAGE_OFFSET; start < (unsigned long)_stext; start += PAGE_SIZE) {
 238                 end = start + PAGE_SIZE;
 239                 if (overlaps_interrupt_vector_text(start, end))
 240                         radix__change_memory_range(start, end, _PAGE_WRITE);
 241                 else
 242                         break;
 243         }
 244 }
 245
 246 void radix__mark_initmem_nx(void)
 247 {
 248         unsigned long start = (unsigned long)__init_begin;
 249         unsigned long end = (unsigned long)__init_end;
 250
 251         radix__change_memory_range(start, end, _PAGE_EXEC);
 252 }
 253 #endif /* CONFIG_STRICT_KERNEL_RWX */
 254
 255 static inline void __meminit
 256 print_mapping(unsigned long start, unsigned long end, unsigned long size, bool exec)
 257 {
 258         char buf[10];
 259
 260         if (end <= start)
 261                 return;
 262
 263         string_get_size(size, 1, STRING_UNITS_2, buf, sizeof(buf));
 264
 265         pr_info("Mapped 0x%016lx-0x%016lx with %s pages%s\n", start, end, buf,
 266                 exec ? " (exec)" : "");
 267 }
 268
 269 static unsigned long next_boundary(unsigned long addr, unsigned long end)
 270 {
 271 #ifdef CONFIG_STRICT_KERNEL_RWX
 272         unsigned long stext_phys;
 273
 274         stext_phys = __pa_symbol(_stext);
 275
 276         // Relocatable kernel running at non-zero real address
 277         if (stext_phys != 0) {
 278                 // The end of interrupts code at zero is a rodata boundary
 279                 unsigned long end_intr = __pa_symbol(__end_interrupts) - stext_phys;
 280                 if (addr < end_intr)
 281                         return end_intr;
 282
 283                 // Start of relocated kernel text is a rodata boundary
 284                 if (addr < stext_phys)
 285                         return stext_phys;
 286         }
 287
 288         if (addr < __pa_symbol(__srwx_boundary))
 289                 return __pa_symbol(__srwx_boundary);
 290 #endif
 291         return end;
 292 }
 293
 294 static int __meminit create_physical_mapping(unsigned long start,
 295                                              unsigned long end,
 296                                              int nid, pgprot_t _prot)
 297 {
 298         unsigned long vaddr, addr, mapping_size = 0;
 299         bool prev_exec, exec = false;
 300         pgprot_t prot;
 301         int psize;
 302         unsigned long max_mapping_size = memory_block_size;
 303
 304         if (debug_pagealloc_enabled_or_kfence())
 305                 max_mapping_size = PAGE_SIZE;
 306
 307         start = ALIGN(start, PAGE_SIZE);
 308         end   = ALIGN_DOWN(end, PAGE_SIZE);
 309         for (addr = start; addr < end; addr += mapping_size) {
 310                 unsigned long gap, previous_size;
 311                 int rc;
 312
 313                 gap = next_boundary(addr, end) - addr;
 314                 if (gap > max_mapping_size)
 315                         gap = max_mapping_size;
 316                 previous_size = mapping_size;
 317                 prev_exec = exec;
 318
 319                 if (IS_ALIGNED(addr, PUD_SIZE) && gap >= PUD_SIZE &&
 320                     mmu_psize_defs[MMU_PAGE_1G].shift) {
 321                         mapping_size = PUD_SIZE;
 322                         psize = MMU_PAGE_1G;
 323                 } else if (IS_ALIGNED(addr, PMD_SIZE) && gap >= PMD_SIZE &&
 324                            mmu_psize_defs[MMU_PAGE_2M].shift) {
 325                         mapping_size = PMD_SIZE;
 326                         psize = MMU_PAGE_2M;
 327                 } else {
 328                         mapping_size = PAGE_SIZE;
 329                         psize = mmu_virtual_psize;
 330                 }
 331
 332                 vaddr = (unsigned long)__va(addr);
 333
 334                 if (overlaps_kernel_text(vaddr, vaddr + mapping_size) ||
 335                     overlaps_interrupt_vector_text(vaddr, vaddr + mapping_size)) {
 336                         prot = PAGE_KERNEL_X;
 337                         exec = true;
 338                 } else {
 339                         prot = _prot;
 340                         exec = false;
 341                 }
 342
 343                 if (mapping_size != previous_size || exec != prev_exec) {
 344                         print_mapping(start, addr, previous_size, prev_exec);
 345                         start = addr;
 346                 }
 347
 348                 rc = __map_kernel_page(vaddr, addr, prot, mapping_size, nid, start, end);
 349                 if (rc)
 350                         return rc;
 351
 352                 update_page_count(psize, 1);
 353         }
 354
 355         print_mapping(start, addr, mapping_size, exec);
 356         return 0;
 357 }
 358
 359 static void __init radix_init_pgtable(void)
 360 {
 361         unsigned long rts_field;
 362         phys_addr_t start, end;
 363         u64 i;
 364
 365         /* We don't support slb for radix */
 366         slb_set_size(0);
 367
 368         /*
 369          * Create the linear mapping
 370          */
 371         for_each_mem_range(i, &start, &end) {
 372                 /*
 373                  * The memblock allocator  is up at this point, so the
 374                  * page tables will be allocated within the range. No
 375                  * need or a node (which we don't have yet).
 376                  */
 377
 378                 if (end >= RADIX_VMALLOC_START) {
 379                         pr_warn("Outside the supported range\n");
 380                         continue;
 381                 }
 382
 383                 WARN_ON(create_physical_mapping(start, end,
 384                                                 -1, PAGE_KERNEL));
 385         }
 386
 387         if (!cpu_has_feature(CPU_FTR_HVMODE) &&
 388                         cpu_has_feature(CPU_FTR_P9_RADIX_PREFETCH_BUG)) {
 389                 /*
 390                  * Older versions of KVM on these machines prefer if the
 391                  * guest only uses the low 19 PID bits.
 392                  */
 393                 mmu_pid_bits = 19;
 394         }
 395         mmu_base_pid = 1;
 396
 397         /*
 398          * Allocate Partition table and process table for the
 399          * host.
 400          */
 401         BUG_ON(PRTB_SIZE_SHIFT > 36);
 402         process_tb = early_alloc_pgtable(1UL << PRTB_SIZE_SHIFT, -1, 0, 0);
 403         /*
 404          * Fill in the process table.
 405          */
 406         rts_field = radix__get_tree_size();
 407         process_tb->prtb0 = cpu_to_be64(rts_field | __pa(init_mm.pgd) | RADIX_PGD_INDEX_SIZE);
 408
 409         /*
 410          * The init_mm context is given the first available (non-zero) PID,
 411          * which is the "guard PID" and contains no page table. PIDR should
 412          * never be set to zero because that duplicates the kernel address
 413          * space at the 0x0... offset (quadrant 0)!
 414          *
 415          * An arbitrary PID that may later be allocated by the PID allocator
 416          * for userspace processes must not be used either, because that
 417          * would cause stale user mappings for that PID on CPUs outside of
 418          * the TLB invalidation scheme (because it won't be in mm_cpumask).
 419          *
 420          * So permanently carve out one PID for the purpose of a guard PID.
 421          */
 422         init_mm.context.id = mmu_base_pid;
 423         mmu_base_pid++;
 424 }
 425
 426 static void __init radix_init_partition_table(void)
 427 {
 428         unsigned long rts_field, dw0, dw1;
 429
 430         mmu_partition_table_init();
 431         rts_field = radix__get_tree_size();
 432         dw0 = rts_field | __pa(init_mm.pgd) | RADIX_PGD_INDEX_SIZE | PATB_HR;
 433         dw1 = __pa(process_tb) | (PRTB_SIZE_SHIFT - 12) | PATB_GR;
 434         mmu_partition_table_set_entry(0, dw0, dw1, false);
 435
 436         pr_info("Initializing Radix MMU\n");
 437 }
 438
 439 static int __init get_idx_from_shift(unsigned int shift)
 440 {
 441         int idx = -1;
 442
 443         switch (shift) {
 444         case 0xc:
 445                 idx = MMU_PAGE_4K;
 446                 break;
 447         case 0x10:
 448                 idx = MMU_PAGE_64K;
 449                 break;
 450         case 0x15:
 451                 idx = MMU_PAGE_2M;
 452                 break;
 453         case 0x1e:
 454                 idx = MMU_PAGE_1G;
 455                 break;
 456         }
 457         return idx;
 458 }
 459
 460 static int __init radix_dt_scan_page_sizes(unsigned long node,
 461                                            const char *uname, int depth,
 462                                            void *data)
 463 {
 464         int size = 0;
 465         int shift, idx;
 466         unsigned int ap;
 467         const __be32 *prop;
 468         const char *type = of_get_flat_dt_prop(node, "device_type", NULL);
 469
 470         /* We are scanning "cpu" nodes only */
 471         if (type == NULL || strcmp(type, "cpu") != 0)
 472                 return 0;
 473
 474         /* Grab page size encodings */
 475         prop = of_get_flat_dt_prop(node, "ibm,processor-radix-AP-encodings", &size);
 476         if (!prop)
 477                 return 0;
 478
 479         pr_info("Page sizes from device-tree:\n");
 480         for (; size >= 4; size -= 4, ++prop) {
 481
 482                 struct mmu_psize_def *def;
 483
 484                 /* top 3 bit is AP encoding */
 485                 shift = be32_to_cpu(prop[0]) & ~(0xe << 28);
 486                 ap = be32_to_cpu(prop[0]) >> 29;
 487                 pr_info("Page size shift = %d AP=0x%x\n", shift, ap);
 488
 489                 idx = get_idx_from_shift(shift);
 490                 if (idx < 0)
 491                         continue;
 492
 493                 def = &mmu_psize_defs[idx];
 494                 def->shift = shift;
 495                 def->ap  = ap;
 496                 def->h_rpt_pgsize = psize_to_rpti_pgsize(idx);
 497         }
 498
 499         /* needed ? */
 500         cur_cpu_spec->mmu_features &= ~MMU_FTR_NO_SLBIE_B;
 501         return 1;
 502 }
 503
 504 void __init radix__early_init_devtree(void)
 505 {
 506         int rc;
 507
 508         /*
 509          * Try to find the available page sizes in the device-tree
 510          */
 511         rc = of_scan_flat_dt(radix_dt_scan_page_sizes, NULL);
 512         if (!rc) {
 513                 /*
 514                  * No page size details found in device tree.
 515                  * Let's assume we have page 4k and 64k support
 516                  */
 517                 mmu_psize_defs[MMU_PAGE_4K].shift = 12;
 518                 mmu_psize_defs[MMU_PAGE_4K].ap = 0x0;
 519                 mmu_psize_defs[MMU_PAGE_4K].h_rpt_pgsize =
 520                         psize_to_rpti_pgsize(MMU_PAGE_4K);
 521
 522                 mmu_psize_defs[MMU_PAGE_64K].shift = 16;
 523                 mmu_psize_defs[MMU_PAGE_64K].ap = 0x5;
 524                 mmu_psize_defs[MMU_PAGE_64K].h_rpt_pgsize =
 525                         psize_to_rpti_pgsize(MMU_PAGE_64K);
 526         }
 527         return;
 528 }
 529
 530 void __init radix__early_init_mmu(void)
 531 {
 532         unsigned long lpcr;
 533
 534 #ifdef CONFIG_PPC_64S_HASH_MMU
 535 #ifdef CONFIG_PPC_64K_PAGES
 536         /* PAGE_SIZE mappings */
 537         mmu_virtual_psize = MMU_PAGE_64K;
 538 #else
 539         mmu_virtual_psize = MMU_PAGE_4K;
 540 #endif
 541 #endif
 542         /*
 543          * initialize page table size
 544          */
 545         __pte_index_size = RADIX_PTE_INDEX_SIZE;
 546         __pmd_index_size = RADIX_PMD_INDEX_SIZE;
 547         __pud_index_size = RADIX_PUD_INDEX_SIZE;
 548         __pgd_index_size = RADIX_PGD_INDEX_SIZE;
 549         __pud_cache_index = RADIX_PUD_INDEX_SIZE;
 550         __pte_table_size = RADIX_PTE_TABLE_SIZE;
 551         __pmd_table_size = RADIX_PMD_TABLE_SIZE;
 552         __pud_table_size = RADIX_PUD_TABLE_SIZE;
 553         __pgd_table_size = RADIX_PGD_TABLE_SIZE;
 554
 555         __pmd_val_bits = RADIX_PMD_VAL_BITS;
 556         __pud_val_bits = RADIX_PUD_VAL_BITS;
 557         __pgd_val_bits = RADIX_PGD_VAL_BITS;
 558
 559         __kernel_virt_start = RADIX_KERN_VIRT_START;
 560         __vmalloc_start = RADIX_VMALLOC_START;
 561         __vmalloc_end = RADIX_VMALLOC_END;
 562         __kernel_io_start = RADIX_KERN_IO_START;
 563         __kernel_io_end = RADIX_KERN_IO_END;
 564         vmemmap = (struct page *)RADIX_VMEMMAP_START;
 565         ioremap_bot = IOREMAP_BASE;
 566
 567 #ifdef CONFIG_PCI
 568         pci_io_base = ISA_IO_BASE;
 569 #endif
 570         __pte_frag_nr = RADIX_PTE_FRAG_NR;
 571         __pte_frag_size_shift = RADIX_PTE_FRAG_SIZE_SHIFT;
 572         __pmd_frag_nr = RADIX_PMD_FRAG_NR;
 573         __pmd_frag_size_shift = RADIX_PMD_FRAG_SIZE_SHIFT;
 574
 575         radix_init_pgtable();
 576
 577         if (!firmware_has_feature(FW_FEATURE_LPAR)) {
 578                 lpcr = mfspr(SPRN_LPCR);
 579                 mtspr(SPRN_LPCR, lpcr | LPCR_UPRT | LPCR_HR);
 580                 radix_init_partition_table();
 581         } else {
 582                 radix_init_pseries();
 583         }
 584
 585         memblock_set_current_limit(MEMBLOCK_ALLOC_ANYWHERE);
 586
 587         /* Switch to the guard PID before turning on MMU */
 588         radix__switch_mmu_context(NULL, &init_mm);
 589         tlbiel_all();
 590 }
 591
 592 void radix__early_init_mmu_secondary(void)
 593 {
 594         unsigned long lpcr;
 595         /*
 596          * update partition table control register and UPRT
 597          */
 598         if (!firmware_has_feature(FW_FEATURE_LPAR)) {
 599                 lpcr = mfspr(SPRN_LPCR);
 600                 mtspr(SPRN_LPCR, lpcr | LPCR_UPRT | LPCR_HR);
 601
 602                 set_ptcr_when_no_uv(__pa(partition_tb) |
 603                                     (PATB_SIZE_SHIFT - 12));
 604         }
 605
 606         radix__switch_mmu_context(NULL, &init_mm);
 607         tlbiel_all();
 608
 609         /* Make sure userspace can't change the AMR */
 610         mtspr(SPRN_UAMOR, 0);
 611 }
 612
 613 /* Called during kexec sequence with MMU off */
 614 notrace void radix__mmu_cleanup_all(void)
 615 {
 616         unsigned long lpcr;
 617
 618         if (!firmware_has_feature(FW_FEATURE_LPAR)) {
 619                 lpcr = mfspr(SPRN_LPCR);
 620                 mtspr(SPRN_LPCR, lpcr & ~LPCR_UPRT);
 621                 set_ptcr_when_no_uv(0);
 622                 powernv_set_nmmu_ptcr(0);
 623                 radix__flush_tlb_all();
 624         }
 625 }
 626
 627 #ifdef CONFIG_MEMORY_HOTPLUG
 628 static void free_pte_table(pte_t *pte_start, pmd_t *pmd)
 629 {
 630         pte_t *pte;
 631         int i;
 632
 633         for (i = 0; i < PTRS_PER_PTE; i++) {
 634                 pte = pte_start + i;
 635                 if (!pte_none(*pte))
 636                         return;
 637         }
 638
 639         pte_free_kernel(&init_mm, pte_start);
 640         pmd_clear(pmd);
 641 }
 642
 643 static void free_pmd_table(pmd_t *pmd_start, pud_t *pud)
 644 {
 645         pmd_t *pmd;
 646         int i;
 647
 648         for (i = 0; i < PTRS_PER_PMD; i++) {
 649                 pmd = pmd_start + i;
 650                 if (!pmd_none(*pmd))
 651                         return;
 652         }
 653
 654         pmd_free(&init_mm, pmd_start);
 655         pud_clear(pud);
 656 }
 657
 658 static void free_pud_table(pud_t *pud_start, p4d_t *p4d)
 659 {
 660         pud_t *pud;
 661         int i;
 662
 663         for (i = 0; i < PTRS_PER_PUD; i++) {
 664                 pud = pud_start + i;
 665                 if (!pud_none(*pud))
 666                         return;
 667         }
 668
 669         pud_free(&init_mm, pud_start);
 670         p4d_clear(p4d);
 671 }
 672
 673 #ifdef CONFIG_SPARSEMEM_VMEMMAP
 674 static bool __meminit vmemmap_pmd_is_unused(unsigned long addr, unsigned long end)
 675 {
 676         unsigned long start = ALIGN_DOWN(addr, PMD_SIZE);
 677
 678         return !vmemmap_populated(start, PMD_SIZE);
 679 }
 680
 681 static bool __meminit vmemmap_page_is_unused(unsigned long addr, unsigned long end)
 682 {
 683         unsigned long start = ALIGN_DOWN(addr, PAGE_SIZE);
 684
 685         return !vmemmap_populated(start, PAGE_SIZE);
 686
 687 }
 688 #endif
 689
 690 static void __meminit free_vmemmap_pages(struct page *page,
 691                                          struct vmem_altmap *altmap,
 692                                          int order)
 693 {
 694         unsigned int nr_pages = 1 << order;
 695
 696         if (altmap) {
 697                 unsigned long alt_start, alt_end;
 698                 unsigned long base_pfn = page_to_pfn(page);
 699
 700                 /*
 701                  * with 2M vmemmap mmaping we can have things setup
 702                  * such that even though atlmap is specified we never
 703                  * used altmap.
 704                  */
 705                 alt_start = altmap->base_pfn;
 706                 alt_end = altmap->base_pfn + altmap->reserve + altmap->free;
 707
 708                 if (base_pfn >= alt_start && base_pfn < alt_end) {
 709                         vmem_altmap_free(altmap, nr_pages);
 710                         return;
 711                 }
 712         }
 713
 714         if (PageReserved(page)) {
 715                 /* allocated from memblock */
 716                 while (nr_pages--)
 717                         free_reserved_page(page++);
 718         } else
 719                 free_pages((unsigned long)page_address(page), order);
 720 }
 721
 722 static void __meminit remove_pte_table(pte_t *pte_start, unsigned long addr,
 723                                        unsigned long end, bool direct,
 724                                        struct vmem_altmap *altmap)
 725 {
 726         unsigned long next, pages = 0;
 727         pte_t *pte;
 728
 729         pte = pte_start + pte_index(addr);
 730         for (; addr < end; addr = next, pte++) {
 731                 next = (addr + PAGE_SIZE) & PAGE_MASK;
 732                 if (next > end)
 733                         next = end;
 734
 735                 if (!pte_present(*pte))
 736                         continue;
 737
 738                 if (PAGE_ALIGNED(addr) && PAGE_ALIGNED(next)) {
 739                         if (!direct)
 740                                 free_vmemmap_pages(pte_page(*pte), altmap, 0);
 741                         pte_clear(&init_mm, addr, pte);
 742                         pages++;
 743                 }
 744 #ifdef CONFIG_SPARSEMEM_VMEMMAP
 745                 else if (!direct && vmemmap_page_is_unused(addr, next)) {
 746                         free_vmemmap_pages(pte_page(*pte), altmap, 0);
 747                         pte_clear(&init_mm, addr, pte);
 748                 }
 749 #endif
 750         }
 751         if (direct)
 752                 update_page_count(mmu_virtual_psize, -pages);
 753 }
 754
 755 static void __meminit remove_pmd_table(pmd_t *pmd_start, unsigned long addr,
 756                                        unsigned long end, bool direct,
 757                                        struct vmem_altmap *altmap)
 758 {
 759         unsigned long next, pages = 0;
 760         pte_t *pte_base;
 761         pmd_t *pmd;
 762
 763         pmd = pmd_start + pmd_index(addr);
 764         for (; addr < end; addr = next, pmd++) {
 765                 next = pmd_addr_end(addr, end);
 766
 767                 if (!pmd_present(*pmd))
 768                         continue;
 769
 770                 if (pmd_leaf(*pmd)) {
 771                         if (IS_ALIGNED(addr, PMD_SIZE) &&
 772                             IS_ALIGNED(next, PMD_SIZE)) {
 773                                 if (!direct)
 774                                         free_vmemmap_pages(pmd_page(*pmd), altmap, get_order(PMD_SIZE));
 775                                 pte_clear(&init_mm, addr, (pte_t *)pmd);
 776                                 pages++;
 777                         }
 778 #ifdef CONFIG_SPARSEMEM_VMEMMAP
 779                         else if (!direct && vmemmap_pmd_is_unused(addr, next)) {
 780                                 free_vmemmap_pages(pmd_page(*pmd), altmap, get_order(PMD_SIZE));
 781                                 pte_clear(&init_mm, addr, (pte_t *)pmd);
 782                         }
 783 #endif
 784                         continue;
 785                 }
 786
 787                 pte_base = (pte_t *)pmd_page_vaddr(*pmd);
 788                 remove_pte_table(pte_base, addr, next, direct, altmap);
 789                 free_pte_table(pte_base, pmd);
 790         }
 791         if (direct)
 792                 update_page_count(MMU_PAGE_2M, -pages);
 793 }
 794
 795 static void __meminit remove_pud_table(pud_t *pud_start, unsigned long addr,
 796                                        unsigned long end, bool direct,
 797                                        struct vmem_altmap *altmap)
 798 {
 799         unsigned long next, pages = 0;
 800         pmd_t *pmd_base;
 801         pud_t *pud;
 802
 803         pud = pud_start + pud_index(addr);
 804         for (; addr < end; addr = next, pud++) {
 805                 next = pud_addr_end(addr, end);
 806
 807                 if (!pud_present(*pud))
 808                         continue;
 809
 810                 if (pud_leaf(*pud)) {
 811                         if (!IS_ALIGNED(addr, PUD_SIZE) ||
 812                             !IS_ALIGNED(next, PUD_SIZE)) {
 813                                 WARN_ONCE(1, "%s: unaligned range\n", __func__);
 814                                 continue;
 815                         }
 816                         pte_clear(&init_mm, addr, (pte_t *)pud);
 817                         pages++;
 818                         continue;
 819                 }
 820
 821                 pmd_base = pud_pgtable(*pud);
 822                 remove_pmd_table(pmd_base, addr, next, direct, altmap);
 823                 free_pmd_table(pmd_base, pud);
 824         }
 825         if (direct)
 826                 update_page_count(MMU_PAGE_1G, -pages);
 827 }
 828
 829 static void __meminit
 830 remove_pagetable(unsigned long start, unsigned long end, bool direct,
 831                  struct vmem_altmap *altmap)
 832 {
 833         unsigned long addr, next;
 834         pud_t *pud_base;
 835         pgd_t *pgd;
 836         p4d_t *p4d;
 837
 838         spin_lock(&init_mm.page_table_lock);
 839
 840         for (addr = start; addr < end; addr = next) {
 841                 next = pgd_addr_end(addr, end);
 842
 843                 pgd = pgd_offset_k(addr);
 844                 p4d = p4d_offset(pgd, addr);
 845                 if (!p4d_present(*p4d))
 846                         continue;
 847
 848                 if (p4d_leaf(*p4d)) {
 849                         if (!IS_ALIGNED(addr, P4D_SIZE) ||
 850                             !IS_ALIGNED(next, P4D_SIZE)) {
 851                                 WARN_ONCE(1, "%s: unaligned range\n", __func__);
 852                                 continue;
 853                         }
 854
 855                         pte_clear(&init_mm, addr, (pte_t *)pgd);
 856                         continue;
 857                 }
 858
 859                 pud_base = p4d_pgtable(*p4d);
 860                 remove_pud_table(pud_base, addr, next, direct, altmap);
 861                 free_pud_table(pud_base, p4d);
 862         }
 863
 864         spin_unlock(&init_mm.page_table_lock);
 865         radix__flush_tlb_kernel_range(start, end);
 866 }
 867
 868 int __meminit radix__create_section_mapping(unsigned long start,
 869                                             unsigned long end, int nid,
 870                                             pgprot_t prot)
 871 {
 872         if (end >= RADIX_VMALLOC_START) {
 873                 pr_warn("Outside the supported range\n");
 874                 return -1;
 875         }
 876
 877         return create_physical_mapping(__pa(start), __pa(end),
 878                                        nid, prot);
 879 }
 880
 881 int __meminit radix__remove_section_mapping(unsigned long start, unsigned long end)
 882 {
 883         remove_pagetable(start, end, true, NULL);
 884         return 0;
 885 }
 886 #endif /* CONFIG_MEMORY_HOTPLUG */
 887
 888 #ifdef CONFIG_SPARSEMEM_VMEMMAP
 889 static int __map_kernel_page_nid(unsigned long ea, unsigned long pa,
 890                                  pgprot_t flags, unsigned int map_page_size,
 891                                  int nid)
 892 {
 893         return __map_kernel_page(ea, pa, flags, map_page_size, nid, 0, 0);
 894 }
 895
 896 int __meminit radix__vmemmap_create_mapping(unsigned long start,
 897                                       unsigned long page_size,
 898                                       unsigned long phys)
 899 {
 900         /* Create a PTE encoding */
 901         int nid = early_pfn_to_nid(phys >> PAGE_SHIFT);
 902         int ret;
 903
 904         if ((start + page_size) >= RADIX_VMEMMAP_END) {
 905                 pr_warn("Outside the supported range\n");
 906                 return -1;
 907         }
 908
 909         ret = __map_kernel_page_nid(start, phys, PAGE_KERNEL, page_size, nid);
 910         BUG_ON(ret);
 911
 912         return 0;
 913 }
 914
 915
 916 bool vmemmap_can_optimize(struct vmem_altmap *altmap, struct dev_pagemap *pgmap)
 917 {
 918         if (radix_enabled())
 919                 return __vmemmap_can_optimize(altmap, pgmap);
 920
 921         return false;
 922 }
 923
 924 int __meminit vmemmap_check_pmd(pmd_t *pmdp, int node,
 925                                 unsigned long addr, unsigned long next)
 926 {
 927         int large = pmd_leaf(*pmdp);
 928
 929         if (large)
 930                 vmemmap_verify(pmdp_ptep(pmdp), node, addr, next);
 931
 932         return large;
 933 }
 934
 935 void __meminit vmemmap_set_pmd(pmd_t *pmdp, void *p, int node,
 936                                unsigned long addr, unsigned long next)
 937 {
 938         pte_t entry;
 939         pte_t *ptep = pmdp_ptep(pmdp);
 940
 941         VM_BUG_ON(!IS_ALIGNED(addr, PMD_SIZE));
 942         entry = pfn_pte(__pa(p) >> PAGE_SHIFT, PAGE_KERNEL);
 943         set_pte_at(&init_mm, addr, ptep, entry);
 944         asm volatile("ptesync": : :"memory");
 945
 946         vmemmap_verify(ptep, node, addr, next);
 947 }
 948
 949 static pte_t * __meminit radix__vmemmap_pte_populate(pmd_t *pmdp, unsigned long addr,
 950                                                      int node,
 951                                                      struct vmem_altmap *altmap,
 952                                                      struct page *reuse)
 953 {
 954         pte_t *pte = pte_offset_kernel(pmdp, addr);
 955
 956         if (pte_none(*pte)) {
 957                 pte_t entry;
 958                 void *p;
 959
 960                 if (!reuse) {
 961                         /*
 962                          * make sure we don't create altmap mappings
 963                          * covering things outside the device.
 964                          */
 965                         if (altmap && altmap_cross_boundary(altmap, addr, PAGE_SIZE))
 966                                 altmap = NULL;
 967
 968                         p = vmemmap_alloc_block_buf(PAGE_SIZE, node, altmap);
 969                         if (!p && altmap)
 970                                 p = vmemmap_alloc_block_buf(PAGE_SIZE, node, NULL);
 971                         if (!p)
 972                                 return NULL;
 973                         pr_debug("PAGE_SIZE vmemmap mapping\n");
 974                 } else {
 975                         /*
 976                          * When a PTE/PMD entry is freed from the init_mm
 977                          * there's a free_pages() call to this page allocated
 978                          * above. Thus this get_page() is paired with the
 979                          * put_page_testzero() on the freeing path.
 980                          * This can only called by certain ZONE_DEVICE path,
 981                          * and through vmemmap_populate_compound_pages() when
 982                          * slab is available.
 983                          */
 984                         get_page(reuse);
 985                         p = page_to_virt(reuse);
 986                         pr_debug("Tail page reuse vmemmap mapping\n");
 987                 }
 988
 989                 VM_BUG_ON(!PAGE_ALIGNED(addr));
 990                 entry = pfn_pte(__pa(p) >> PAGE_SHIFT, PAGE_KERNEL);
 991                 set_pte_at(&init_mm, addr, pte, entry);
 992                 asm volatile("ptesync": : :"memory");
 993         }
 994         return pte;
 995 }
 996
 997 static inline pud_t *vmemmap_pud_alloc(p4d_t *p4dp, int node,
 998                                        unsigned long address)
 999 {
1000         pud_t *pud;
1001
1002         /* All early vmemmap mapping to keep simple do it at PAGE_SIZE */
1003         if (unlikely(p4d_none(*p4dp))) {
1004                 if (unlikely(!slab_is_available())) {
1005                         pud = early_alloc_pgtable(PAGE_SIZE, node, 0, 0);
1006                         p4d_populate(&init_mm, p4dp, pud);
1007                         /* go to the pud_offset */
1008                 } else
1009                         return pud_alloc(&init_mm, p4dp, address);
1010         }
1011         return pud_offset(p4dp, address);
1012 }
1013
1014 static inline pmd_t *vmemmap_pmd_alloc(pud_t *pudp, int node,
1015                                        unsigned long address)
1016 {
1017         pmd_t *pmd;
1018
1019         /* All early vmemmap mapping to keep simple do it at PAGE_SIZE */
1020         if (unlikely(pud_none(*pudp))) {
1021                 if (unlikely(!slab_is_available())) {
1022                         pmd = early_alloc_pgtable(PAGE_SIZE, node, 0, 0);
1023                         pud_populate(&init_mm, pudp, pmd);
1024                 } else
1025                         return pmd_alloc(&init_mm, pudp, address);
1026         }
1027         return pmd_offset(pudp, address);
1028 }
1029
1030 static inline pte_t *vmemmap_pte_alloc(pmd_t *pmdp, int node,
1031                                        unsigned long address)
1032 {
1033         pte_t *pte;
1034
1035         /* All early vmemmap mapping to keep simple do it at PAGE_SIZE */
1036         if (unlikely(pmd_none(*pmdp))) {
1037                 if (unlikely(!slab_is_available())) {
1038                         pte = early_alloc_pgtable(PAGE_SIZE, node, 0, 0);
1039                         pmd_populate(&init_mm, pmdp, pte);
1040                 } else
1041                         return pte_alloc_kernel(pmdp, address);
1042         }
1043         return pte_offset_kernel(pmdp, address);
1044 }
1045
1046
1047
1048 int __meminit radix__vmemmap_populate(unsigned long start, unsigned long end, int node,
1049                                       struct vmem_altmap *altmap)
1050 {
1051         unsigned long addr;
1052         unsigned long next;
1053         pgd_t *pgd;
1054         p4d_t *p4d;
1055         pud_t *pud;
1056         pmd_t *pmd;
1057         pte_t *pte;
1058
1059         for (addr = start; addr < end; addr = next) {
1060                 next = pmd_addr_end(addr, end);
1061
1062                 pgd = pgd_offset_k(addr);
1063                 p4d = p4d_offset(pgd, addr);
1064                 pud = vmemmap_pud_alloc(p4d, node, addr);
1065                 if (!pud)
1066                         return -ENOMEM;
1067                 pmd = vmemmap_pmd_alloc(pud, node, addr);
1068                 if (!pmd)
1069                         return -ENOMEM;
1070
1071                 if (pmd_none(READ_ONCE(*pmd))) {
1072                         void *p;
1073
1074                         /*
1075                          * keep it simple by checking addr PMD_SIZE alignment
1076                          * and verifying the device boundary condition.
1077                          * For us to use a pmd mapping, both addr and pfn should
1078                          * be aligned. We skip if addr is not aligned and for
1079                          * pfn we hope we have extra area in the altmap that
1080                          * can help to find an aligned block. This can result
1081                          * in altmap block allocation failures, in which case
1082                          * we fallback to RAM for vmemmap allocation.
1083                          */
1084                         if (altmap && (!IS_ALIGNED(addr, PMD_SIZE) ||
1085                                        altmap_cross_boundary(altmap, addr, PMD_SIZE))) {
1086                                 /*
1087                                  * make sure we don't create altmap mappings
1088                                  * covering things outside the device.
1089                                  */
1090                                 goto base_mapping;
1091                         }
1092
1093                         p = vmemmap_alloc_block_buf(PMD_SIZE, node, altmap);
1094                         if (p) {
1095                                 vmemmap_set_pmd(pmd, p, node, addr, next);
1096                                 pr_debug("PMD_SIZE vmemmap mapping\n");
1097                                 continue;
1098                         } else if (altmap) {
1099                                 /*
1100                                  * A vmemmap block allocation can fail due to
1101                                  * alignment requirements and we trying to align
1102                                  * things aggressively there by running out of
1103                                  * space. Try base mapping on failure.
1104                                  */
1105                                 goto base_mapping;
1106                         }
1107                 } else if (vmemmap_check_pmd(pmd, node, addr, next)) {
1108                         /*
1109                          * If a huge mapping exist due to early call to
1110                          * vmemmap_populate, let's try to use that.
1111                          */
1112                         continue;
1113                 }
1114 base_mapping:
1115                 /*
1116                  * Not able allocate higher order memory to back memmap
1117                  * or we found a pointer to pte page. Allocate base page
1118                  * size vmemmap
1119                  */
1120                 pte = vmemmap_pte_alloc(pmd, node, addr);
1121                 if (!pte)
1122                         return -ENOMEM;
1123
1124                 pte = radix__vmemmap_pte_populate(pmd, addr, node, altmap, NULL);
1125                 if (!pte)
1126                         return -ENOMEM;
1127
1128                 vmemmap_verify(pte, node, addr, addr + PAGE_SIZE);
1129                 next = addr + PAGE_SIZE;
1130         }
1131         return 0;
1132 }
1133
1134 static pte_t * __meminit radix__vmemmap_populate_address(unsigned long addr, int node,
1135                                                          struct vmem_altmap *altmap,
1136                                                          struct page *reuse)
1137 {
1138         pgd_t *pgd;
1139         p4d_t *p4d;
1140         pud_t *pud;
1141         pmd_t *pmd;
1142         pte_t *pte;
1143
1144         pgd = pgd_offset_k(addr);
1145         p4d = p4d_offset(pgd, addr);
1146         pud = vmemmap_pud_alloc(p4d, node, addr);
1147         if (!pud)
1148                 return NULL;
1149         pmd = vmemmap_pmd_alloc(pud, node, addr);
1150         if (!pmd)
1151                 return NULL;
1152         if (pmd_leaf(*pmd))
1153                 /*
1154                  * The second page is mapped as a hugepage due to a nearby request.
1155                  * Force our mapping to page size without deduplication
1156                  */
1157                 return NULL;
1158         pte = vmemmap_pte_alloc(pmd, node, addr);
1159         if (!pte)
1160                 return NULL;
1161         radix__vmemmap_pte_populate(pmd, addr, node, NULL, NULL);
1162         vmemmap_verify(pte, node, addr, addr + PAGE_SIZE);
1163
1164         return pte;
1165 }
1166
1167 static pte_t * __meminit vmemmap_compound_tail_page(unsigned long addr,
1168                                                     unsigned long pfn_offset, int node)
1169 {
1170         pgd_t *pgd;
1171         p4d_t *p4d;
1172         pud_t *pud;
1173         pmd_t *pmd;
1174         pte_t *pte;
1175         unsigned long map_addr;
1176
1177         /* the second vmemmap page which we use for duplication */
1178         map_addr = addr - pfn_offset * sizeof(struct page) + PAGE_SIZE;
1179         pgd = pgd_offset_k(map_addr);
1180         p4d = p4d_offset(pgd, map_addr);
1181         pud = vmemmap_pud_alloc(p4d, node, map_addr);
1182         if (!pud)
1183                 return NULL;
1184         pmd = vmemmap_pmd_alloc(pud, node, map_addr);
1185         if (!pmd)
1186                 return NULL;
1187         if (pmd_leaf(*pmd))
1188                 /*
1189                  * The second page is mapped as a hugepage due to a nearby request.
1190                  * Force our mapping to page size without deduplication
1191                  */
1192                 return NULL;
1193         pte = vmemmap_pte_alloc(pmd, node, map_addr);
1194         if (!pte)
1195                 return NULL;
1196         /*
1197          * Check if there exist a mapping to the left
1198          */
1199         if (pte_none(*pte)) {
1200                 /*
1201                  * Populate the head page vmemmap page.
1202                  * It can fall in different pmd, hence
1203                  * vmemmap_populate_address()
1204                  */
1205                 pte = radix__vmemmap_populate_address(map_addr - PAGE_SIZE, node, NULL, NULL);
1206                 if (!pte)
1207                         return NULL;
1208                 /*
1209                  * Populate the tail pages vmemmap page
1210                  */
1211                 pte = radix__vmemmap_pte_populate(pmd, map_addr, node, NULL, NULL);
1212                 if (!pte)
1213                         return NULL;
1214                 vmemmap_verify(pte, node, map_addr, map_addr + PAGE_SIZE);
1215                 return pte;
1216         }
1217         return pte;
1218 }
1219
1220 int __meminit vmemmap_populate_compound_pages(unsigned long start_pfn,
1221                                               unsigned long start,
1222                                               unsigned long end, int node,
1223                                               struct dev_pagemap *pgmap)
1224 {
1225         /*
1226          * we want to map things as base page size mapping so that
1227          * we can save space in vmemmap. We could have huge mapping
1228          * covering out both edges.
1229          */
1230         unsigned long addr;
1231         unsigned long addr_pfn = start_pfn;
1232         unsigned long next;
1233         pgd_t *pgd;
1234         p4d_t *p4d;
1235         pud_t *pud;
1236         pmd_t *pmd;
1237         pte_t *pte;
1238
1239         for (addr = start; addr < end; addr = next) {
1240
1241                 pgd = pgd_offset_k(addr);
1242                 p4d = p4d_offset(pgd, addr);
1243                 pud = vmemmap_pud_alloc(p4d, node, addr);
1244                 if (!pud)
1245                         return -ENOMEM;
1246                 pmd = vmemmap_pmd_alloc(pud, node, addr);
1247                 if (!pmd)
1248                         return -ENOMEM;
1249
1250                 if (pmd_leaf(READ_ONCE(*pmd))) {
1251                         /* existing huge mapping. Skip the range */
1252                         addr_pfn += (PMD_SIZE >> PAGE_SHIFT);
1253                         next = pmd_addr_end(addr, end);
1254                         continue;
1255                 }
1256                 pte = vmemmap_pte_alloc(pmd, node, addr);
1257                 if (!pte)
1258                         return -ENOMEM;
1259                 if (!pte_none(*pte)) {
1260                         /*
1261                          * This could be because we already have a compound
1262                          * page whose VMEMMAP_RESERVE_NR pages were mapped and
1263                          * this request fall in those pages.
1264                          */
1265                         addr_pfn += 1;
1266                         next = addr + PAGE_SIZE;
1267                         continue;
1268                 } else {
1269                         unsigned long nr_pages = pgmap_vmemmap_nr(pgmap);
1270                         unsigned long pfn_offset = addr_pfn - ALIGN_DOWN(addr_pfn, nr_pages);
1271                         pte_t *tail_page_pte;
1272
1273                         /*
1274                          * if the address is aligned to huge page size it is the
1275                          * head mapping.
1276                          */
1277                         if (pfn_offset == 0) {
1278                                 /* Populate the head page vmemmap page */
1279                                 pte = radix__vmemmap_pte_populate(pmd, addr, node, NULL, NULL);
1280                                 if (!pte)
1281                                         return -ENOMEM;
1282                                 vmemmap_verify(pte, node, addr, addr + PAGE_SIZE);
1283
1284                                 /*
1285                                  * Populate the tail pages vmemmap page
1286                                  * It can fall in different pmd, hence
1287                                  * vmemmap_populate_address()
1288                                  */
1289                                 pte = radix__vmemmap_populate_address(addr + PAGE_SIZE, node, NULL, NULL);
1290                                 if (!pte)
1291                                         return -ENOMEM;
1292
1293                                 addr_pfn += 2;
1294                                 next = addr + 2 * PAGE_SIZE;
1295                                 continue;
1296                         }
1297                         /*
1298                          * get the 2nd mapping details
1299                          * Also create it if that doesn't exist
1300                          */
1301                         tail_page_pte = vmemmap_compound_tail_page(addr, pfn_offset, node);
1302                         if (!tail_page_pte) {
1303
1304                                 pte = radix__vmemmap_pte_populate(pmd, addr, node, NULL, NULL);
1305                                 if (!pte)
1306                                         return -ENOMEM;
1307                                 vmemmap_verify(pte, node, addr, addr + PAGE_SIZE);
1308
1309                                 addr_pfn += 1;
1310                                 next = addr + PAGE_SIZE;
1311                                 continue;
1312                         }
1313
1314                         pte = radix__vmemmap_pte_populate(pmd, addr, node, NULL, pte_page(*tail_page_pte));
1315                         if (!pte)
1316                                 return -ENOMEM;
1317                         vmemmap_verify(pte, node, addr, addr + PAGE_SIZE);
1318
1319                         addr_pfn += 1;
1320                         next = addr + PAGE_SIZE;
1321                         continue;
1322                 }
1323         }
1324         return 0;
1325 }
1326
1327
1328 #ifdef CONFIG_MEMORY_HOTPLUG
1329 void __meminit radix__vmemmap_remove_mapping(unsigned long start, unsigned long page_size)
1330 {
1331         remove_pagetable(start, start + page_size, true, NULL);
1332 }
1333
1334 void __ref radix__vmemmap_free(unsigned long start, unsigned long end,
1335                                struct vmem_altmap *altmap)
1336 {
1337         remove_pagetable(start, end, false, altmap);
1338 }
1339 #endif
1340 #endif
1341
1342 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1343
1344 unsigned long radix__pmd_hugepage_update(struct mm_struct *mm, unsigned long addr,
1345                                   pmd_t *pmdp, unsigned long clr,
1346                                   unsigned long set)
1347 {
1348         unsigned long old;
1349
1350 #ifdef CONFIG_DEBUG_VM
1351         WARN_ON(!radix__pmd_trans_huge(*pmdp) && !pmd_devmap(*pmdp));
1352         assert_spin_locked(pmd_lockptr(mm, pmdp));
1353 #endif
1354
1355         old = radix__pte_update(mm, addr, pmdp_ptep(pmdp), clr, set, 1);
1356         trace_hugepage_update_pmd(addr, old, clr, set);
1357
1358         return old;
1359 }
1360
1361 unsigned long radix__pud_hugepage_update(struct mm_struct *mm, unsigned long addr,
1362                                          pud_t *pudp, unsigned long clr,
1363                                          unsigned long set)
1364 {
1365         unsigned long old;
1366
1367 #ifdef CONFIG_DEBUG_VM
1368         WARN_ON(!pud_devmap(*pudp));
1369         assert_spin_locked(pud_lockptr(mm, pudp));
1370 #endif
1371
1372         old = radix__pte_update(mm, addr, pudp_ptep(pudp), clr, set, 1);
1373         trace_hugepage_update_pud(addr, old, clr, set);
1374
1375         return old;
1376 }
1377
1378 pmd_t radix__pmdp_collapse_flush(struct vm_area_struct *vma, unsigned long address,
1379                         pmd_t *pmdp)
1380
1381 {
1382         pmd_t pmd;
1383
1384         VM_BUG_ON(address & ~HPAGE_PMD_MASK);
1385         VM_BUG_ON(radix__pmd_trans_huge(*pmdp));
1386         VM_BUG_ON(pmd_devmap(*pmdp));
1387         /*
1388          * khugepaged calls this for normal pmd
1389          */
1390         pmd = *pmdp;
1391         pmd_clear(pmdp);
1392
1393         radix__flush_tlb_collapsed_pmd(vma->vm_mm, address);
1394
1395         return pmd;
1396 }
1397
1398 /*
1399  * For us pgtable_t is pte_t *. Inorder to save the deposisted
1400  * page table, we consider the allocated page table as a list
1401  * head. On withdraw we need to make sure we zero out the used
1402  * list_head memory area.
1403  */
1404 void radix__pgtable_trans_huge_deposit(struct mm_struct *mm, pmd_t *pmdp,
1405                                  pgtable_t pgtable)
1406 {
1407         struct list_head *lh = (struct list_head *) pgtable;
1408
1409         assert_spin_locked(pmd_lockptr(mm, pmdp));
1410
1411         /* FIFO */
1412         if (!pmd_huge_pte(mm, pmdp))
1413                 INIT_LIST_HEAD(lh);
1414         else
1415                 list_add(lh, (struct list_head *) pmd_huge_pte(mm, pmdp));
1416         pmd_huge_pte(mm, pmdp) = pgtable;
1417 }
1418
1419 pgtable_t radix__pgtable_trans_huge_withdraw(struct mm_struct *mm, pmd_t *pmdp)
1420 {
1421         pte_t *ptep;
1422         pgtable_t pgtable;
1423         struct list_head *lh;
1424
1425         assert_spin_locked(pmd_lockptr(mm, pmdp));
1426
1427         /* FIFO */
1428         pgtable = pmd_huge_pte(mm, pmdp);
1429         lh = (struct list_head *) pgtable;
1430         if (list_empty(lh))
1431                 pmd_huge_pte(mm, pmdp) = NULL;
1432         else {
1433                 pmd_huge_pte(mm, pmdp) = (pgtable_t) lh->next;
1434                 list_del(lh);
1435         }
1436         ptep = (pte_t *) pgtable;
1437         *ptep = __pte(0);
1438         ptep++;
1439         *ptep = __pte(0);
1440         return pgtable;
1441 }
1442
1443 pmd_t radix__pmdp_huge_get_and_clear(struct mm_struct *mm,
1444                                      unsigned long addr, pmd_t *pmdp)
1445 {
1446         pmd_t old_pmd;
1447         unsigned long old;
1448
1449         old = radix__pmd_hugepage_update(mm, addr, pmdp, ~0UL, 0);
1450         old_pmd = __pmd(old);
1451         return old_pmd;
1452 }
1453
1454 pud_t radix__pudp_huge_get_and_clear(struct mm_struct *mm,
1455                                      unsigned long addr, pud_t *pudp)
1456 {
1457         pud_t old_pud;
1458         unsigned long old;
1459
1460         old = radix__pud_hugepage_update(mm, addr, pudp, ~0UL, 0);
1461         old_pud = __pud(old);
1462         return old_pud;
1463 }
1464
1465 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
1466
1467 void radix__ptep_set_access_flags(struct vm_area_struct *vma, pte_t *ptep,
1468                                   pte_t entry, unsigned long address, int psize)
1469 {
1470         struct mm_struct *mm = vma->vm_mm;
1471         unsigned long set = pte_val(entry) & (_PAGE_DIRTY | _PAGE_SOFT_DIRTY |
1472                                               _PAGE_ACCESSED | _PAGE_RW | _PAGE_EXEC);
1473
1474         unsigned long change = pte_val(entry) ^ pte_val(*ptep);
1475         /*
1476          * On POWER9, the NMMU is not able to relax PTE access permissions
1477          * for a translation with a TLB. The PTE must be invalidated, TLB
1478          * flushed before the new PTE is installed.
1479          *
1480          * This only needs to be done for radix, because hash translation does
1481          * flush when updating the linux pte (and we don't support NMMU
1482          * accelerators on HPT on POWER9 anyway XXX: do we?).
1483          *
1484          * POWER10 (and P9P) NMMU does behave as per ISA.
1485          */
1486         if (!cpu_has_feature(CPU_FTR_ARCH_31) && (change & _PAGE_RW) &&
1487             atomic_read(&mm->context.copros) > 0) {
1488                 unsigned long old_pte, new_pte;
1489
1490                 old_pte = __radix_pte_update(ptep, _PAGE_PRESENT, _PAGE_INVALID);
1491                 new_pte = old_pte | set;
1492                 radix__flush_tlb_page_psize(mm, address, psize);
1493                 __radix_pte_update(ptep, _PAGE_INVALID, new_pte);
1494         } else {
1495                 __radix_pte_update(ptep, 0, set);
1496                 /*
1497                  * Book3S does not require a TLB flush when relaxing access
1498                  * restrictions when the address space (modulo the POWER9 nest
1499                  * MMU issue above) because the MMU will reload the PTE after
1500                  * taking an access fault, as defined by the architecture. See
1501                  * "Setting a Reference or Change Bit or Upgrading Access
1502                  *  Authority (PTE Subject to Atomic Hardware Updates)" in
1503                  *  Power ISA Version 3.1B.
1504                  */
1505         }
1506         /* See ptesync comment in radix__set_pte_at */
1507 }
1508
1509 void radix__ptep_modify_prot_commit(struct vm_area_struct *vma,
1510                                     unsigned long addr, pte_t *ptep,
1511                                     pte_t old_pte, pte_t pte)
1512 {
1513         struct mm_struct *mm = vma->vm_mm;
1514
1515         /*
1516          * POWER9 NMMU must flush the TLB after clearing the PTE before
1517          * installing a PTE with more relaxed access permissions, see
1518          * radix__ptep_set_access_flags.
1519          */
1520         if (!cpu_has_feature(CPU_FTR_ARCH_31) &&
1521             is_pte_rw_upgrade(pte_val(old_pte), pte_val(pte)) &&
1522             (atomic_read(&mm->context.copros) > 0))
1523                 radix__flush_tlb_page(vma, addr);
1524
1525         set_pte_at(mm, addr, ptep, pte);
1526 }
1527
1528 int pud_set_huge(pud_t *pud, phys_addr_t addr, pgprot_t prot)
1529 {
1530         pte_t *ptep = (pte_t *)pud;
1531         pte_t new_pud = pfn_pte(__phys_to_pfn(addr), prot);
1532
1533         if (!radix_enabled())
1534                 return 0;
1535
1536         set_pte_at(&init_mm, 0 /* radix unused */, ptep, new_pud);
1537
1538         return 1;
1539 }
1540
1541 int pud_clear_huge(pud_t *pud)
1542 {
1543         if (pud_leaf(*pud)) {
1544                 pud_clear(pud);
1545                 return 1;
1546         }
1547
1548         return 0;
1549 }
1550
1551 int pud_free_pmd_page(pud_t *pud, unsigned long addr)
1552 {
1553         pmd_t *pmd;
1554         int i;
1555
1556         pmd = pud_pgtable(*pud);
1557         pud_clear(pud);
1558
1559         flush_tlb_kernel_range(addr, addr + PUD_SIZE);
1560
1561         for (i = 0; i < PTRS_PER_PMD; i++) {
1562                 if (!pmd_none(pmd[i])) {
1563                         pte_t *pte;
1564                         pte = (pte_t *)pmd_page_vaddr(pmd[i]);
1565
1566                         pte_free_kernel(&init_mm, pte);
1567                 }
1568         }
1569
1570         pmd_free(&init_mm, pmd);
1571
1572         return 1;
1573 }
1574
1575 int pmd_set_huge(pmd_t *pmd, phys_addr_t addr, pgprot_t prot)
1576 {
1577         pte_t *ptep = (pte_t *)pmd;
1578         pte_t new_pmd = pfn_pte(__phys_to_pfn(addr), prot);
1579
1580         if (!radix_enabled())
1581                 return 0;
1582
1583         set_pte_at(&init_mm, 0 /* radix unused */, ptep, new_pmd);
1584
1585         return 1;
1586 }
1587
1588 int pmd_clear_huge(pmd_t *pmd)
1589 {
1590         if (pmd_leaf(*pmd)) {
1591                 pmd_clear(pmd);
1592                 return 1;
1593         }
1594
1595         return 0;
1596 }
1597
1598 int pmd_free_pte_page(pmd_t *pmd, unsigned long addr)
1599 {
1600         pte_t *pte;
1601
1602         pte = (pte_t *)pmd_page_vaddr(*pmd);
1603         pmd_clear(pmd);
1604
1605         flush_tlb_kernel_range(addr, addr + PMD_SIZE);
1606
1607         pte_free_kernel(&init_mm, pte);
1608
1609         return 1;
1610 }