1 /* 2 * PowerPC version 3 * Copyright (C) 1995-1996 Gary Thomas (gdt@linuxppc.org) 4 * 5 * Modifications by Paul Mackerras (PowerMac) (paulus@cs.anu.edu.au) 6 * and Cort Dougan (PReP) (cort@cs.nmt.edu) 7 * Copyright (C) 1996 Paul Mackerras 8 * 9 * Derived from "arch/i386/mm/init.c" 10 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds 11 * 12 * Dave Engebretsen <engebret@us.ibm.com> 13 * Rework for PPC64 port. 14 * 15 * This program is free software; you can redistribute it and/or 16 * modify it under the terms of the GNU General Public License 17 * as published by the Free Software Foundation; either version 18 * 2 of the License, or (at your option) any later version. 19 * 20 */ 21 22 #undef DEBUG 23 24 #include <linux/signal.h> 25 #include <linux/sched.h> 26 #include <linux/kernel.h> 27 #include <linux/errno.h> 28 #include <linux/string.h> 29 #include <linux/types.h> 30 #include <linux/mman.h> 31 #include <linux/mm.h> 32 #include <linux/swap.h> 33 #include <linux/stddef.h> 34 #include <linux/vmalloc.h> 35 #include <linux/init.h> 36 #include <linux/delay.h> 37 #include <linux/highmem.h> 38 #include <linux/idr.h> 39 #include <linux/nodemask.h> 40 #include <linux/module.h> 41 #include <linux/poison.h> 42 #include <linux/memblock.h> 43 #include <linux/hugetlb.h> 44 #include <linux/slab.h> 45 46 #include <asm/pgalloc.h> 47 #include <asm/page.h> 48 #include <asm/prom.h> 49 #include <asm/rtas.h> 50 #include <asm/io.h> 51 #include <asm/mmu_context.h> 52 #include <asm/pgtable.h> 53 #include <asm/mmu.h> 54 #include <asm/uaccess.h> 55 #include <asm/smp.h> 56 #include <asm/machdep.h> 57 #include <asm/tlb.h> 58 #include <asm/eeh.h> 59 #include <asm/processor.h> 60 #include <asm/mmzone.h> 61 #include <asm/cputable.h> 62 #include <asm/sections.h> 63 #include <asm/iommu.h> 64 #include <asm/vdso.h> 65 66 #include "mmu_decl.h" 67 68 #ifdef CONFIG_PPC_STD_MMU_64 69 #if PGTABLE_RANGE > USER_VSID_RANGE 70 #warning Limited user VSID range means pagetable space is wasted 71 #endif 72 73 #if (TASK_SIZE_USER64 < PGTABLE_RANGE) && (TASK_SIZE_USER64 < USER_VSID_RANGE) 74 #warning TASK_SIZE is smaller than it needs to be. 75 #endif 76 #endif /* CONFIG_PPC_STD_MMU_64 */ 77 78 phys_addr_t memstart_addr = ~0; 79 EXPORT_SYMBOL_GPL(memstart_addr); 80 phys_addr_t kernstart_addr; 81 EXPORT_SYMBOL_GPL(kernstart_addr); 82 83 static void pgd_ctor(void *addr) 84 { 85 memset(addr, 0, PGD_TABLE_SIZE); 86 } 87 88 static void pmd_ctor(void *addr) 89 { 90 #ifdef CONFIG_TRANSPARENT_HUGEPAGE 91 memset(addr, 0, PMD_TABLE_SIZE * 2); 92 #else 93 memset(addr, 0, PMD_TABLE_SIZE); 94 #endif 95 } 96 97 struct kmem_cache *pgtable_cache[MAX_PGTABLE_INDEX_SIZE]; 98 99 /* 100 * Create a kmem_cache() for pagetables. This is not used for PTE 101 * pages - they're linked to struct page, come from the normal free 102 * pages pool and have a different entry size (see real_pte_t) to 103 * everything else. Caches created by this function are used for all 104 * the higher level pagetables, and for hugepage pagetables. 105 */ 106 void pgtable_cache_add(unsigned shift, void (*ctor)(void *)) 107 { 108 char *name; 109 unsigned long table_size = sizeof(void *) << shift; 110 unsigned long align = table_size; 111 112 /* When batching pgtable pointers for RCU freeing, we store 113 * the index size in the low bits. Table alignment must be 114 * big enough to fit it. 115 * 116 * Likewise, hugeapge pagetable pointers contain a (different) 117 * shift value in the low bits. All tables must be aligned so 118 * as to leave enough 0 bits in the address to contain it. */ 119 unsigned long minalign = max(MAX_PGTABLE_INDEX_SIZE + 1, 120 HUGEPD_SHIFT_MASK + 1); 121 struct kmem_cache *new; 122 123 /* It would be nice if this was a BUILD_BUG_ON(), but at the 124 * moment, gcc doesn't seem to recognize is_power_of_2 as a 125 * constant expression, so so much for that. */ 126 BUG_ON(!is_power_of_2(minalign)); 127 BUG_ON((shift < 1) || (shift > MAX_PGTABLE_INDEX_SIZE)); 128 129 if (PGT_CACHE(shift)) 130 return; /* Already have a cache of this size */ 131 132 align = max_t(unsigned long, align, minalign); 133 name = kasprintf(GFP_KERNEL, "pgtable-2^%d", shift); 134 new = kmem_cache_create(name, table_size, align, 0, ctor); 135 pgtable_cache[shift - 1] = new; 136 pr_debug("Allocated pgtable cache for order %d\n", shift); 137 } 138 139 140 void pgtable_cache_init(void) 141 { 142 pgtable_cache_add(PGD_INDEX_SIZE, pgd_ctor); 143 pgtable_cache_add(PMD_CACHE_INDEX, pmd_ctor); 144 if (!PGT_CACHE(PGD_INDEX_SIZE) || !PGT_CACHE(PMD_CACHE_INDEX)) 145 panic("Couldn't allocate pgtable caches"); 146 /* In all current configs, when the PUD index exists it's the 147 * same size as either the pgd or pmd index. Verify that the 148 * initialization above has also created a PUD cache. This 149 * will need re-examiniation if we add new possibilities for 150 * the pagetable layout. */ 151 BUG_ON(PUD_INDEX_SIZE && !PGT_CACHE(PUD_INDEX_SIZE)); 152 } 153 154 #ifdef CONFIG_SPARSEMEM_VMEMMAP 155 /* 156 * Given an address within the vmemmap, determine the pfn of the page that 157 * represents the start of the section it is within. Note that we have to 158 * do this by hand as the proffered address may not be correctly aligned. 159 * Subtraction of non-aligned pointers produces undefined results. 160 */ 161 static unsigned long __meminit vmemmap_section_start(unsigned long page) 162 { 163 unsigned long offset = page - ((unsigned long)(vmemmap)); 164 165 /* Return the pfn of the start of the section. */ 166 return (offset / sizeof(struct page)) & PAGE_SECTION_MASK; 167 } 168 169 /* 170 * Check if this vmemmap page is already initialised. If any section 171 * which overlaps this vmemmap page is initialised then this page is 172 * initialised already. 173 */ 174 static int __meminit vmemmap_populated(unsigned long start, int page_size) 175 { 176 unsigned long end = start + page_size; 177 start = (unsigned long)(pfn_to_page(vmemmap_section_start(start))); 178 179 for (; start < end; start += (PAGES_PER_SECTION * sizeof(struct page))) 180 if (pfn_valid(page_to_pfn((struct page *)start))) 181 return 1; 182 183 return 0; 184 } 185 186 /* On hash-based CPUs, the vmemmap is bolted in the hash table. 187 * 188 * On Book3E CPUs, the vmemmap is currently mapped in the top half of 189 * the vmalloc space using normal page tables, though the size of 190 * pages encoded in the PTEs can be different 191 */ 192 193 #ifdef CONFIG_PPC_BOOK3E 194 static void __meminit vmemmap_create_mapping(unsigned long start, 195 unsigned long page_size, 196 unsigned long phys) 197 { 198 /* Create a PTE encoding without page size */ 199 unsigned long i, flags = _PAGE_PRESENT | _PAGE_ACCESSED | 200 _PAGE_KERNEL_RW; 201 202 /* PTEs only contain page size encodings up to 32M */ 203 BUG_ON(mmu_psize_defs[mmu_vmemmap_psize].enc > 0xf); 204 205 /* Encode the size in the PTE */ 206 flags |= mmu_psize_defs[mmu_vmemmap_psize].enc << 8; 207 208 /* For each PTE for that area, map things. Note that we don't 209 * increment phys because all PTEs are of the large size and 210 * thus must have the low bits clear 211 */ 212 for (i = 0; i < page_size; i += PAGE_SIZE) 213 BUG_ON(map_kernel_page(start + i, phys, flags)); 214 } 215 216 #ifdef CONFIG_MEMORY_HOTPLUG 217 static void vmemmap_remove_mapping(unsigned long start, 218 unsigned long page_size) 219 { 220 } 221 #endif 222 #else /* CONFIG_PPC_BOOK3E */ 223 static void __meminit vmemmap_create_mapping(unsigned long start, 224 unsigned long page_size, 225 unsigned long phys) 226 { 227 int mapped = htab_bolt_mapping(start, start + page_size, phys, 228 pgprot_val(PAGE_KERNEL), 229 mmu_vmemmap_psize, 230 mmu_kernel_ssize); 231 BUG_ON(mapped < 0); 232 } 233 234 #ifdef CONFIG_MEMORY_HOTPLUG 235 static void vmemmap_remove_mapping(unsigned long start, 236 unsigned long page_size) 237 { 238 int mapped = htab_remove_mapping(start, start + page_size, 239 mmu_vmemmap_psize, 240 mmu_kernel_ssize); 241 BUG_ON(mapped < 0); 242 } 243 #endif 244 245 #endif /* CONFIG_PPC_BOOK3E */ 246 247 struct vmemmap_backing *vmemmap_list; 248 static struct vmemmap_backing *next; 249 static int num_left; 250 static int num_freed; 251 252 static __meminit struct vmemmap_backing * vmemmap_list_alloc(int node) 253 { 254 struct vmemmap_backing *vmem_back; 255 /* get from freed entries first */ 256 if (num_freed) { 257 num_freed--; 258 vmem_back = next; 259 next = next->list; 260 261 return vmem_back; 262 } 263 264 /* allocate a page when required and hand out chunks */ 265 if (!num_left) { 266 next = vmemmap_alloc_block(PAGE_SIZE, node); 267 if (unlikely(!next)) { 268 WARN_ON(1); 269 return NULL; 270 } 271 num_left = PAGE_SIZE / sizeof(struct vmemmap_backing); 272 } 273 274 num_left--; 275 276 return next++; 277 } 278 279 static __meminit void vmemmap_list_populate(unsigned long phys, 280 unsigned long start, 281 int node) 282 { 283 struct vmemmap_backing *vmem_back; 284 285 vmem_back = vmemmap_list_alloc(node); 286 if (unlikely(!vmem_back)) { 287 WARN_ON(1); 288 return; 289 } 290 291 vmem_back->phys = phys; 292 vmem_back->virt_addr = start; 293 vmem_back->list = vmemmap_list; 294 295 vmemmap_list = vmem_back; 296 } 297 298 int __meminit vmemmap_populate(unsigned long start, unsigned long end, int node) 299 { 300 unsigned long page_size = 1 << mmu_psize_defs[mmu_vmemmap_psize].shift; 301 302 /* Align to the page size of the linear mapping. */ 303 start = _ALIGN_DOWN(start, page_size); 304 305 pr_debug("vmemmap_populate %lx..%lx, node %d\n", start, end, node); 306 307 for (; start < end; start += page_size) { 308 void *p; 309 310 if (vmemmap_populated(start, page_size)) 311 continue; 312 313 p = vmemmap_alloc_block(page_size, node); 314 if (!p) 315 return -ENOMEM; 316 317 vmemmap_list_populate(__pa(p), start, node); 318 319 pr_debug(" * %016lx..%016lx allocated at %p\n", 320 start, start + page_size, p); 321 322 vmemmap_create_mapping(start, page_size, __pa(p)); 323 } 324 325 return 0; 326 } 327 328 #ifdef CONFIG_MEMORY_HOTPLUG 329 static unsigned long vmemmap_list_free(unsigned long start) 330 { 331 struct vmemmap_backing *vmem_back, *vmem_back_prev; 332 333 vmem_back_prev = vmem_back = vmemmap_list; 334 335 /* look for it with prev pointer recorded */ 336 for (; vmem_back; vmem_back = vmem_back->list) { 337 if (vmem_back->virt_addr == start) 338 break; 339 vmem_back_prev = vmem_back; 340 } 341 342 if (unlikely(!vmem_back)) { 343 WARN_ON(1); 344 return 0; 345 } 346 347 /* remove it from vmemmap_list */ 348 if (vmem_back == vmemmap_list) /* remove head */ 349 vmemmap_list = vmem_back->list; 350 else 351 vmem_back_prev->list = vmem_back->list; 352 353 /* next point to this freed entry */ 354 vmem_back->list = next; 355 next = vmem_back; 356 num_freed++; 357 358 return vmem_back->phys; 359 } 360 361 void __ref vmemmap_free(unsigned long start, unsigned long end) 362 { 363 unsigned long page_size = 1 << mmu_psize_defs[mmu_vmemmap_psize].shift; 364 365 start = _ALIGN_DOWN(start, page_size); 366 367 pr_debug("vmemmap_free %lx...%lx\n", start, end); 368 369 for (; start < end; start += page_size) { 370 unsigned long addr; 371 372 /* 373 * the section has already be marked as invalid, so 374 * vmemmap_populated() true means some other sections still 375 * in this page, so skip it. 376 */ 377 if (vmemmap_populated(start, page_size)) 378 continue; 379 380 addr = vmemmap_list_free(start); 381 if (addr) { 382 struct page *page = pfn_to_page(addr >> PAGE_SHIFT); 383 384 if (PageReserved(page)) { 385 /* allocated from bootmem */ 386 if (page_size < PAGE_SIZE) { 387 /* 388 * this shouldn't happen, but if it is 389 * the case, leave the memory there 390 */ 391 WARN_ON_ONCE(1); 392 } else { 393 unsigned int nr_pages = 394 1 << get_order(page_size); 395 while (nr_pages--) 396 free_reserved_page(page++); 397 } 398 } else 399 free_pages((unsigned long)(__va(addr)), 400 get_order(page_size)); 401 402 vmemmap_remove_mapping(start, page_size); 403 } 404 } 405 } 406 #endif 407 void register_page_bootmem_memmap(unsigned long section_nr, 408 struct page *start_page, unsigned long size) 409 { 410 } 411 412 /* 413 * We do not have access to the sparsemem vmemmap, so we fallback to 414 * walking the list of sparsemem blocks which we already maintain for 415 * the sake of crashdump. In the long run, we might want to maintain 416 * a tree if performance of that linear walk becomes a problem. 417 * 418 * realmode_pfn_to_page functions can fail due to: 419 * 1) As real sparsemem blocks do not lay in RAM continously (they 420 * are in virtual address space which is not available in the real mode), 421 * the requested page struct can be split between blocks so get_page/put_page 422 * may fail. 423 * 2) When huge pages are used, the get_page/put_page API will fail 424 * in real mode as the linked addresses in the page struct are virtual 425 * too. 426 */ 427 struct page *realmode_pfn_to_page(unsigned long pfn) 428 { 429 struct vmemmap_backing *vmem_back; 430 struct page *page; 431 unsigned long page_size = 1 << mmu_psize_defs[mmu_vmemmap_psize].shift; 432 unsigned long pg_va = (unsigned long) pfn_to_page(pfn); 433 434 for (vmem_back = vmemmap_list; vmem_back; vmem_back = vmem_back->list) { 435 if (pg_va < vmem_back->virt_addr) 436 continue; 437 438 /* After vmemmap_list entry free is possible, need check all */ 439 if ((pg_va + sizeof(struct page)) <= 440 (vmem_back->virt_addr + page_size)) { 441 page = (struct page *) (vmem_back->phys + pg_va - 442 vmem_back->virt_addr); 443 return page; 444 } 445 } 446 447 /* Probably that page struct is split between real pages */ 448 return NULL; 449 } 450 EXPORT_SYMBOL_GPL(realmode_pfn_to_page); 451 452 #elif defined(CONFIG_FLATMEM) 453 454 struct page *realmode_pfn_to_page(unsigned long pfn) 455 { 456 struct page *page = pfn_to_page(pfn); 457 return page; 458 } 459 EXPORT_SYMBOL_GPL(realmode_pfn_to_page); 460 461 #endif /* CONFIG_SPARSEMEM_VMEMMAP/CONFIG_FLATMEM */ 462