1 // SPDX-License-Identifier: GPL-2.0 2 /* 3 * Initialize MMU support. 4 * 5 * Copyright (C) 1998-2003 Hewlett-Packard Co 6 * David Mosberger-Tang <davidm@hpl.hp.com> 7 */ 8 #include <linux/kernel.h> 9 #include <linux/init.h> 10 11 #include <linux/dma-map-ops.h> 12 #include <linux/dmar.h> 13 #include <linux/efi.h> 14 #include <linux/elf.h> 15 #include <linux/memblock.h> 16 #include <linux/mm.h> 17 #include <linux/sched/signal.h> 18 #include <linux/mmzone.h> 19 #include <linux/module.h> 20 #include <linux/personality.h> 21 #include <linux/reboot.h> 22 #include <linux/slab.h> 23 #include <linux/swap.h> 24 #include <linux/proc_fs.h> 25 #include <linux/bitops.h> 26 #include <linux/kexec.h> 27 #include <linux/swiotlb.h> 28 29 #include <asm/dma.h> 30 #include <asm/efi.h> 31 #include <asm/io.h> 32 #include <asm/numa.h> 33 #include <asm/patch.h> 34 #include <asm/pgalloc.h> 35 #include <asm/sal.h> 36 #include <asm/sections.h> 37 #include <asm/tlb.h> 38 #include <linux/uaccess.h> 39 #include <asm/unistd.h> 40 #include <asm/mca.h> 41 42 extern void ia64_tlb_init (void); 43 44 unsigned long MAX_DMA_ADDRESS = PAGE_OFFSET + 0x100000000UL; 45 46 struct page *zero_page_memmap_ptr; /* map entry for zero page */ 47 EXPORT_SYMBOL(zero_page_memmap_ptr); 48 49 void 50 __ia64_sync_icache_dcache (pte_t pte) 51 { 52 unsigned long addr; 53 struct folio *folio; 54 55 folio = page_folio(pte_page(pte)); 56 addr = (unsigned long)folio_address(folio); 57 58 if (test_bit(PG_arch_1, &folio->flags)) 59 return; /* i-cache is already coherent with d-cache */ 60 61 flush_icache_range(addr, addr + folio_size(folio)); 62 set_bit(PG_arch_1, &folio->flags); /* mark page as clean */ 63 } 64 65 /* 66 * Since DMA is i-cache coherent, any (complete) folios that were written via 67 * DMA can be marked as "clean" so that lazy_mmu_prot_update() doesn't have to 68 * flush them when they get mapped into an executable vm-area. 69 */ 70 void arch_dma_mark_clean(phys_addr_t paddr, size_t size) 71 { 72 unsigned long pfn = PHYS_PFN(paddr); 73 struct folio *folio = page_folio(pfn_to_page(pfn)); 74 ssize_t left = size; 75 size_t offset = offset_in_folio(folio, paddr); 76 77 if (offset) { 78 left -= folio_size(folio) - offset; 79 if (left <= 0) 80 return; 81 folio = folio_next(folio); 82 } 83 84 while (left >= (ssize_t)folio_size(folio)) { 85 left -= folio_size(folio); 86 set_bit(PG_arch_1, &pfn_to_page(pfn)->flags); 87 if (!left) 88 break; 89 folio = folio_next(folio); 90 } 91 } 92 93 inline void 94 ia64_set_rbs_bot (void) 95 { 96 unsigned long stack_size = rlimit_max(RLIMIT_STACK) & -16; 97 98 if (stack_size > MAX_USER_STACK_SIZE) 99 stack_size = MAX_USER_STACK_SIZE; 100 current->thread.rbs_bot = PAGE_ALIGN(current->mm->start_stack - stack_size); 101 } 102 103 /* 104 * This performs some platform-dependent address space initialization. 105 * On IA-64, we want to setup the VM area for the register backing 106 * store (which grows upwards) and install the gateway page which is 107 * used for signal trampolines, etc. 108 */ 109 void 110 ia64_init_addr_space (void) 111 { 112 struct vm_area_struct *vma; 113 114 ia64_set_rbs_bot(); 115 116 /* 117 * If we're out of memory and kmem_cache_alloc() returns NULL, we simply ignore 118 * the problem. When the process attempts to write to the register backing store 119 * for the first time, it will get a SEGFAULT in this case. 120 */ 121 vma = vm_area_alloc(current->mm); 122 if (vma) { 123 vma_set_anonymous(vma); 124 vma->vm_start = current->thread.rbs_bot & PAGE_MASK; 125 vma->vm_end = vma->vm_start + PAGE_SIZE; 126 vm_flags_init(vma, VM_DATA_DEFAULT_FLAGS|VM_GROWSUP|VM_ACCOUNT); 127 vma->vm_page_prot = vm_get_page_prot(vma->vm_flags); 128 mmap_write_lock(current->mm); 129 if (insert_vm_struct(current->mm, vma)) { 130 mmap_write_unlock(current->mm); 131 vm_area_free(vma); 132 return; 133 } 134 mmap_write_unlock(current->mm); 135 } 136 137 /* map NaT-page at address zero to speed up speculative dereferencing of NULL: */ 138 if (!(current->personality & MMAP_PAGE_ZERO)) { 139 vma = vm_area_alloc(current->mm); 140 if (vma) { 141 vma_set_anonymous(vma); 142 vma->vm_end = PAGE_SIZE; 143 vma->vm_page_prot = __pgprot(pgprot_val(PAGE_READONLY) | _PAGE_MA_NAT); 144 vm_flags_init(vma, VM_READ | VM_MAYREAD | VM_IO | 145 VM_DONTEXPAND | VM_DONTDUMP); 146 mmap_write_lock(current->mm); 147 if (insert_vm_struct(current->mm, vma)) { 148 mmap_write_unlock(current->mm); 149 vm_area_free(vma); 150 return; 151 } 152 mmap_write_unlock(current->mm); 153 } 154 } 155 } 156 157 void 158 free_initmem (void) 159 { 160 free_reserved_area(ia64_imva(__init_begin), ia64_imva(__init_end), 161 -1, "unused kernel"); 162 } 163 164 void __init 165 free_initrd_mem (unsigned long start, unsigned long end) 166 { 167 /* 168 * EFI uses 4KB pages while the kernel can use 4KB or bigger. 169 * Thus EFI and the kernel may have different page sizes. It is 170 * therefore possible to have the initrd share the same page as 171 * the end of the kernel (given current setup). 172 * 173 * To avoid freeing/using the wrong page (kernel sized) we: 174 * - align up the beginning of initrd 175 * - align down the end of initrd 176 * 177 * | | 178 * |=============| a000 179 * | | 180 * | | 181 * | | 9000 182 * |/////////////| 183 * |/////////////| 184 * |=============| 8000 185 * |///INITRD////| 186 * |/////////////| 187 * |/////////////| 7000 188 * | | 189 * |KKKKKKKKKKKKK| 190 * |=============| 6000 191 * |KKKKKKKKKKKKK| 192 * |KKKKKKKKKKKKK| 193 * K=kernel using 8KB pages 194 * 195 * In this example, we must free page 8000 ONLY. So we must align up 196 * initrd_start and keep initrd_end as is. 197 */ 198 start = PAGE_ALIGN(start); 199 end = end & PAGE_MASK; 200 201 if (start < end) 202 printk(KERN_INFO "Freeing initrd memory: %ldkB freed\n", (end - start) >> 10); 203 204 for (; start < end; start += PAGE_SIZE) { 205 if (!virt_addr_valid(start)) 206 continue; 207 free_reserved_page(virt_to_page(start)); 208 } 209 } 210 211 /* 212 * This installs a clean page in the kernel's page table. 213 */ 214 static struct page * __init 215 put_kernel_page (struct page *page, unsigned long address, pgprot_t pgprot) 216 { 217 pgd_t *pgd; 218 p4d_t *p4d; 219 pud_t *pud; 220 pmd_t *pmd; 221 pte_t *pte; 222 223 pgd = pgd_offset_k(address); /* note: this is NOT pgd_offset()! */ 224 225 { 226 p4d = p4d_alloc(&init_mm, pgd, address); 227 if (!p4d) 228 goto out; 229 pud = pud_alloc(&init_mm, p4d, address); 230 if (!pud) 231 goto out; 232 pmd = pmd_alloc(&init_mm, pud, address); 233 if (!pmd) 234 goto out; 235 pte = pte_alloc_kernel(pmd, address); 236 if (!pte) 237 goto out; 238 if (!pte_none(*pte)) 239 goto out; 240 set_pte(pte, mk_pte(page, pgprot)); 241 } 242 out: 243 /* no need for flush_tlb */ 244 return page; 245 } 246 247 static void __init 248 setup_gate (void) 249 { 250 struct page *page; 251 252 /* 253 * Map the gate page twice: once read-only to export the ELF 254 * headers etc. and once execute-only page to enable 255 * privilege-promotion via "epc": 256 */ 257 page = virt_to_page(ia64_imva(__start_gate_section)); 258 put_kernel_page(page, GATE_ADDR, PAGE_READONLY); 259 #ifdef HAVE_BUGGY_SEGREL 260 page = virt_to_page(ia64_imva(__start_gate_section + PAGE_SIZE)); 261 put_kernel_page(page, GATE_ADDR + PAGE_SIZE, PAGE_GATE); 262 #else 263 put_kernel_page(page, GATE_ADDR + PERCPU_PAGE_SIZE, PAGE_GATE); 264 /* Fill in the holes (if any) with read-only zero pages: */ 265 { 266 unsigned long addr; 267 268 for (addr = GATE_ADDR + PAGE_SIZE; 269 addr < GATE_ADDR + PERCPU_PAGE_SIZE; 270 addr += PAGE_SIZE) 271 { 272 put_kernel_page(ZERO_PAGE(0), addr, 273 PAGE_READONLY); 274 put_kernel_page(ZERO_PAGE(0), addr + PERCPU_PAGE_SIZE, 275 PAGE_READONLY); 276 } 277 } 278 #endif 279 ia64_patch_gate(); 280 } 281 282 static struct vm_area_struct gate_vma; 283 284 static int __init gate_vma_init(void) 285 { 286 vma_init(&gate_vma, NULL); 287 gate_vma.vm_start = FIXADDR_USER_START; 288 gate_vma.vm_end = FIXADDR_USER_END; 289 vm_flags_init(&gate_vma, VM_READ | VM_MAYREAD | VM_EXEC | VM_MAYEXEC); 290 gate_vma.vm_page_prot = __pgprot(__ACCESS_BITS | _PAGE_PL_3 | _PAGE_AR_RX); 291 292 return 0; 293 } 294 __initcall(gate_vma_init); 295 296 struct vm_area_struct *get_gate_vma(struct mm_struct *mm) 297 { 298 return &gate_vma; 299 } 300 301 int in_gate_area_no_mm(unsigned long addr) 302 { 303 if ((addr >= FIXADDR_USER_START) && (addr < FIXADDR_USER_END)) 304 return 1; 305 return 0; 306 } 307 308 int in_gate_area(struct mm_struct *mm, unsigned long addr) 309 { 310 return in_gate_area_no_mm(addr); 311 } 312 313 void ia64_mmu_init(void *my_cpu_data) 314 { 315 unsigned long pta, impl_va_bits; 316 extern void tlb_init(void); 317 318 #ifdef CONFIG_DISABLE_VHPT 319 # define VHPT_ENABLE_BIT 0 320 #else 321 # define VHPT_ENABLE_BIT 1 322 #endif 323 324 /* 325 * Check if the virtually mapped linear page table (VMLPT) overlaps with a mapped 326 * address space. The IA-64 architecture guarantees that at least 50 bits of 327 * virtual address space are implemented but if we pick a large enough page size 328 * (e.g., 64KB), the mapped address space is big enough that it will overlap with 329 * VMLPT. I assume that once we run on machines big enough to warrant 64KB pages, 330 * IMPL_VA_MSB will be significantly bigger, so this is unlikely to become a 331 * problem in practice. Alternatively, we could truncate the top of the mapped 332 * address space to not permit mappings that would overlap with the VMLPT. 333 * --davidm 00/12/06 334 */ 335 # define pte_bits 3 336 # define mapped_space_bits (3*(PAGE_SHIFT - pte_bits) + PAGE_SHIFT) 337 /* 338 * The virtual page table has to cover the entire implemented address space within 339 * a region even though not all of this space may be mappable. The reason for 340 * this is that the Access bit and Dirty bit fault handlers perform 341 * non-speculative accesses to the virtual page table, so the address range of the 342 * virtual page table itself needs to be covered by virtual page table. 343 */ 344 # define vmlpt_bits (impl_va_bits - PAGE_SHIFT + pte_bits) 345 # define POW2(n) (1ULL << (n)) 346 347 impl_va_bits = ffz(~(local_cpu_data->unimpl_va_mask | (7UL << 61))); 348 349 if (impl_va_bits < 51 || impl_va_bits > 61) 350 panic("CPU has bogus IMPL_VA_MSB value of %lu!\n", impl_va_bits - 1); 351 /* 352 * mapped_space_bits - PAGE_SHIFT is the total number of ptes we need, 353 * which must fit into "vmlpt_bits - pte_bits" slots. Second half of 354 * the test makes sure that our mapped space doesn't overlap the 355 * unimplemented hole in the middle of the region. 356 */ 357 if ((mapped_space_bits - PAGE_SHIFT > vmlpt_bits - pte_bits) || 358 (mapped_space_bits > impl_va_bits - 1)) 359 panic("Cannot build a big enough virtual-linear page table" 360 " to cover mapped address space.\n" 361 " Try using a smaller page size.\n"); 362 363 364 /* place the VMLPT at the end of each page-table mapped region: */ 365 pta = POW2(61) - POW2(vmlpt_bits); 366 367 /* 368 * Set the (virtually mapped linear) page table address. Bit 369 * 8 selects between the short and long format, bits 2-7 the 370 * size of the table, and bit 0 whether the VHPT walker is 371 * enabled. 372 */ 373 ia64_set_pta(pta | (0 << 8) | (vmlpt_bits << 2) | VHPT_ENABLE_BIT); 374 375 ia64_tlb_init(); 376 377 #ifdef CONFIG_HUGETLB_PAGE 378 ia64_set_rr(HPAGE_REGION_BASE, HPAGE_SHIFT << 2); 379 ia64_srlz_d(); 380 #endif 381 } 382 383 int __init register_active_ranges(u64 start, u64 len, int nid) 384 { 385 u64 end = start + len; 386 387 #ifdef CONFIG_KEXEC 388 if (start > crashk_res.start && start < crashk_res.end) 389 start = crashk_res.end; 390 if (end > crashk_res.start && end < crashk_res.end) 391 end = crashk_res.start; 392 #endif 393 394 if (start < end) 395 memblock_add_node(__pa(start), end - start, nid, MEMBLOCK_NONE); 396 return 0; 397 } 398 399 int 400 find_max_min_low_pfn (u64 start, u64 end, void *arg) 401 { 402 unsigned long pfn_start, pfn_end; 403 #ifdef CONFIG_FLATMEM 404 pfn_start = (PAGE_ALIGN(__pa(start))) >> PAGE_SHIFT; 405 pfn_end = (PAGE_ALIGN(__pa(end - 1))) >> PAGE_SHIFT; 406 #else 407 pfn_start = GRANULEROUNDDOWN(__pa(start)) >> PAGE_SHIFT; 408 pfn_end = GRANULEROUNDUP(__pa(end - 1)) >> PAGE_SHIFT; 409 #endif 410 min_low_pfn = min(min_low_pfn, pfn_start); 411 max_low_pfn = max(max_low_pfn, pfn_end); 412 return 0; 413 } 414 415 /* 416 * Boot command-line option "nolwsys" can be used to disable the use of any light-weight 417 * system call handler. When this option is in effect, all fsyscalls will end up bubbling 418 * down into the kernel and calling the normal (heavy-weight) syscall handler. This is 419 * useful for performance testing, but conceivably could also come in handy for debugging 420 * purposes. 421 */ 422 423 static int nolwsys __initdata; 424 425 static int __init 426 nolwsys_setup (char *s) 427 { 428 nolwsys = 1; 429 return 1; 430 } 431 432 __setup("nolwsys", nolwsys_setup); 433 434 void __init 435 mem_init (void) 436 { 437 int i; 438 439 BUG_ON(PTRS_PER_PGD * sizeof(pgd_t) != PAGE_SIZE); 440 BUG_ON(PTRS_PER_PMD * sizeof(pmd_t) != PAGE_SIZE); 441 BUG_ON(PTRS_PER_PTE * sizeof(pte_t) != PAGE_SIZE); 442 443 /* 444 * This needs to be called _after_ the command line has been parsed but 445 * _before_ any drivers that may need the PCI DMA interface are 446 * initialized or bootmem has been freed. 447 */ 448 do { 449 #ifdef CONFIG_INTEL_IOMMU 450 detect_intel_iommu(); 451 if (iommu_detected) 452 break; 453 #endif 454 swiotlb_init(true, SWIOTLB_VERBOSE); 455 } while (0); 456 457 #ifdef CONFIG_FLATMEM 458 BUG_ON(!mem_map); 459 #endif 460 461 set_max_mapnr(max_low_pfn); 462 high_memory = __va(max_low_pfn * PAGE_SIZE); 463 memblock_free_all(); 464 465 /* 466 * For fsyscall entrypoints with no light-weight handler, use the ordinary 467 * (heavy-weight) handler, but mark it by setting bit 0, so the fsyscall entry 468 * code can tell them apart. 469 */ 470 for (i = 0; i < NR_syscalls; ++i) { 471 extern unsigned long fsyscall_table[NR_syscalls]; 472 extern unsigned long sys_call_table[NR_syscalls]; 473 474 if (!fsyscall_table[i] || nolwsys) 475 fsyscall_table[i] = sys_call_table[i] | 1; 476 } 477 setup_gate(); 478 } 479 480 #ifdef CONFIG_MEMORY_HOTPLUG 481 int arch_add_memory(int nid, u64 start, u64 size, 482 struct mhp_params *params) 483 { 484 unsigned long start_pfn = start >> PAGE_SHIFT; 485 unsigned long nr_pages = size >> PAGE_SHIFT; 486 int ret; 487 488 if (WARN_ON_ONCE(params->pgprot.pgprot != PAGE_KERNEL.pgprot)) 489 return -EINVAL; 490 491 ret = __add_pages(nid, start_pfn, nr_pages, params); 492 if (ret) 493 printk("%s: Problem encountered in __add_pages() as ret=%d\n", 494 __func__, ret); 495 496 return ret; 497 } 498 499 void arch_remove_memory(u64 start, u64 size, struct vmem_altmap *altmap) 500 { 501 unsigned long start_pfn = start >> PAGE_SHIFT; 502 unsigned long nr_pages = size >> PAGE_SHIFT; 503 504 __remove_pages(start_pfn, nr_pages, altmap); 505 } 506 #endif 507 508 static const pgprot_t protection_map[16] = { 509 [VM_NONE] = PAGE_NONE, 510 [VM_READ] = PAGE_READONLY, 511 [VM_WRITE] = PAGE_READONLY, 512 [VM_WRITE | VM_READ] = PAGE_READONLY, 513 [VM_EXEC] = __pgprot(__ACCESS_BITS | _PAGE_PL_3 | 514 _PAGE_AR_X_RX), 515 [VM_EXEC | VM_READ] = __pgprot(__ACCESS_BITS | _PAGE_PL_3 | 516 _PAGE_AR_RX), 517 [VM_EXEC | VM_WRITE] = PAGE_COPY_EXEC, 518 [VM_EXEC | VM_WRITE | VM_READ] = PAGE_COPY_EXEC, 519 [VM_SHARED] = PAGE_NONE, 520 [VM_SHARED | VM_READ] = PAGE_READONLY, 521 [VM_SHARED | VM_WRITE] = PAGE_SHARED, 522 [VM_SHARED | VM_WRITE | VM_READ] = PAGE_SHARED, 523 [VM_SHARED | VM_EXEC] = __pgprot(__ACCESS_BITS | _PAGE_PL_3 | 524 _PAGE_AR_X_RX), 525 [VM_SHARED | VM_EXEC | VM_READ] = __pgprot(__ACCESS_BITS | _PAGE_PL_3 | 526 _PAGE_AR_RX), 527 [VM_SHARED | VM_EXEC | VM_WRITE] = __pgprot(__ACCESS_BITS | _PAGE_PL_3 | 528 _PAGE_AR_RWX), 529 [VM_SHARED | VM_EXEC | VM_WRITE | VM_READ] = __pgprot(__ACCESS_BITS | _PAGE_PL_3 | 530 _PAGE_AR_RWX) 531 }; 532 DECLARE_VM_GET_PAGE_PROT 533