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 #include <linux/of_fdt.h> 46 #include <linux/libfdt.h> 47 #include <linux/memremap.h> 48 49 #include <asm/pgalloc.h> 50 #include <asm/page.h> 51 #include <asm/prom.h> 52 #include <asm/rtas.h> 53 #include <asm/io.h> 54 #include <asm/mmu_context.h> 55 #include <asm/pgtable.h> 56 #include <asm/mmu.h> 57 #include <linux/uaccess.h> 58 #include <asm/smp.h> 59 #include <asm/machdep.h> 60 #include <asm/tlb.h> 61 #include <asm/eeh.h> 62 #include <asm/processor.h> 63 #include <asm/mmzone.h> 64 #include <asm/cputable.h> 65 #include <asm/sections.h> 66 #include <asm/iommu.h> 67 #include <asm/vdso.h> 68 69 #include "mmu_decl.h" 70 71 #ifdef CONFIG_PPC_BOOK3S_64 72 #if H_PGTABLE_RANGE > USER_VSID_RANGE 73 #warning Limited user VSID range means pagetable space is wasted 74 #endif 75 #endif /* CONFIG_PPC_BOOK3S_64 */ 76 77 phys_addr_t memstart_addr = ~0; 78 EXPORT_SYMBOL_GPL(memstart_addr); 79 phys_addr_t kernstart_addr; 80 EXPORT_SYMBOL_GPL(kernstart_addr); 81 82 #ifdef CONFIG_SPARSEMEM_VMEMMAP 83 /* 84 * Given an address within the vmemmap, determine the pfn of the page that 85 * represents the start of the section it is within. Note that we have to 86 * do this by hand as the proffered address may not be correctly aligned. 87 * Subtraction of non-aligned pointers produces undefined results. 88 */ 89 static unsigned long __meminit vmemmap_section_start(unsigned long page) 90 { 91 unsigned long offset = page - ((unsigned long)(vmemmap)); 92 93 /* Return the pfn of the start of the section. */ 94 return (offset / sizeof(struct page)) & PAGE_SECTION_MASK; 95 } 96 97 /* 98 * Check if this vmemmap page is already initialised. If any section 99 * which overlaps this vmemmap page is initialised then this page is 100 * initialised already. 101 */ 102 static int __meminit vmemmap_populated(unsigned long start, int page_size) 103 { 104 unsigned long end = start + page_size; 105 start = (unsigned long)(pfn_to_page(vmemmap_section_start(start))); 106 107 for (; start < end; start += (PAGES_PER_SECTION * sizeof(struct page))) 108 if (pfn_valid(page_to_pfn((struct page *)start))) 109 return 1; 110 111 return 0; 112 } 113 114 /* 115 * vmemmap virtual address space management does not have a traditonal page 116 * table to track which virtual struct pages are backed by physical mapping. 117 * The virtual to physical mappings are tracked in a simple linked list 118 * format. 'vmemmap_list' maintains the entire vmemmap physical mapping at 119 * all times where as the 'next' list maintains the available 120 * vmemmap_backing structures which have been deleted from the 121 * 'vmemmap_global' list during system runtime (memory hotplug remove 122 * operation). The freed 'vmemmap_backing' structures are reused later when 123 * new requests come in without allocating fresh memory. This pointer also 124 * tracks the allocated 'vmemmap_backing' structures as we allocate one 125 * full page memory at a time when we dont have any. 126 */ 127 struct vmemmap_backing *vmemmap_list; 128 static struct vmemmap_backing *next; 129 130 /* 131 * The same pointer 'next' tracks individual chunks inside the allocated 132 * full page during the boot time and again tracks the freeed nodes during 133 * runtime. It is racy but it does not happen as they are separated by the 134 * boot process. Will create problem if some how we have memory hotplug 135 * operation during boot !! 136 */ 137 static int num_left; 138 static int num_freed; 139 140 static __meminit struct vmemmap_backing * vmemmap_list_alloc(int node) 141 { 142 struct vmemmap_backing *vmem_back; 143 /* get from freed entries first */ 144 if (num_freed) { 145 num_freed--; 146 vmem_back = next; 147 next = next->list; 148 149 return vmem_back; 150 } 151 152 /* allocate a page when required and hand out chunks */ 153 if (!num_left) { 154 next = vmemmap_alloc_block(PAGE_SIZE, node); 155 if (unlikely(!next)) { 156 WARN_ON(1); 157 return NULL; 158 } 159 num_left = PAGE_SIZE / sizeof(struct vmemmap_backing); 160 } 161 162 num_left--; 163 164 return next++; 165 } 166 167 static __meminit void vmemmap_list_populate(unsigned long phys, 168 unsigned long start, 169 int node) 170 { 171 struct vmemmap_backing *vmem_back; 172 173 vmem_back = vmemmap_list_alloc(node); 174 if (unlikely(!vmem_back)) { 175 WARN_ON(1); 176 return; 177 } 178 179 vmem_back->phys = phys; 180 vmem_back->virt_addr = start; 181 vmem_back->list = vmemmap_list; 182 183 vmemmap_list = vmem_back; 184 } 185 186 int __meminit vmemmap_populate(unsigned long start, unsigned long end, int node, 187 struct vmem_altmap *altmap) 188 { 189 unsigned long page_size = 1 << mmu_psize_defs[mmu_vmemmap_psize].shift; 190 191 /* Align to the page size of the linear mapping. */ 192 start = _ALIGN_DOWN(start, page_size); 193 194 pr_debug("vmemmap_populate %lx..%lx, node %d\n", start, end, node); 195 196 for (; start < end; start += page_size) { 197 void *p; 198 int rc; 199 200 if (vmemmap_populated(start, page_size)) 201 continue; 202 203 if (altmap) 204 p = altmap_alloc_block_buf(page_size, altmap); 205 else 206 p = vmemmap_alloc_block_buf(page_size, node); 207 if (!p) 208 return -ENOMEM; 209 210 vmemmap_list_populate(__pa(p), start, node); 211 212 pr_debug(" * %016lx..%016lx allocated at %p\n", 213 start, start + page_size, p); 214 215 rc = vmemmap_create_mapping(start, page_size, __pa(p)); 216 if (rc < 0) { 217 pr_warn("%s: Unable to create vmemmap mapping: %d\n", 218 __func__, rc); 219 return -EFAULT; 220 } 221 } 222 223 return 0; 224 } 225 226 #ifdef CONFIG_MEMORY_HOTPLUG 227 static unsigned long vmemmap_list_free(unsigned long start) 228 { 229 struct vmemmap_backing *vmem_back, *vmem_back_prev; 230 231 vmem_back_prev = vmem_back = vmemmap_list; 232 233 /* look for it with prev pointer recorded */ 234 for (; vmem_back; vmem_back = vmem_back->list) { 235 if (vmem_back->virt_addr == start) 236 break; 237 vmem_back_prev = vmem_back; 238 } 239 240 if (unlikely(!vmem_back)) { 241 WARN_ON(1); 242 return 0; 243 } 244 245 /* remove it from vmemmap_list */ 246 if (vmem_back == vmemmap_list) /* remove head */ 247 vmemmap_list = vmem_back->list; 248 else 249 vmem_back_prev->list = vmem_back->list; 250 251 /* next point to this freed entry */ 252 vmem_back->list = next; 253 next = vmem_back; 254 num_freed++; 255 256 return vmem_back->phys; 257 } 258 259 void __ref vmemmap_free(unsigned long start, unsigned long end, 260 struct vmem_altmap *altmap) 261 { 262 unsigned long page_size = 1 << mmu_psize_defs[mmu_vmemmap_psize].shift; 263 unsigned long page_order = get_order(page_size); 264 265 start = _ALIGN_DOWN(start, page_size); 266 267 pr_debug("vmemmap_free %lx...%lx\n", start, end); 268 269 for (; start < end; start += page_size) { 270 unsigned long nr_pages, addr; 271 struct page *section_base; 272 struct page *page; 273 274 /* 275 * the section has already be marked as invalid, so 276 * vmemmap_populated() true means some other sections still 277 * in this page, so skip it. 278 */ 279 if (vmemmap_populated(start, page_size)) 280 continue; 281 282 addr = vmemmap_list_free(start); 283 if (!addr) 284 continue; 285 286 page = pfn_to_page(addr >> PAGE_SHIFT); 287 section_base = pfn_to_page(vmemmap_section_start(start)); 288 nr_pages = 1 << page_order; 289 290 if (altmap) { 291 vmem_altmap_free(altmap, nr_pages); 292 } else if (PageReserved(page)) { 293 /* allocated from bootmem */ 294 if (page_size < PAGE_SIZE) { 295 /* 296 * this shouldn't happen, but if it is 297 * the case, leave the memory there 298 */ 299 WARN_ON_ONCE(1); 300 } else { 301 while (nr_pages--) 302 free_reserved_page(page++); 303 } 304 } else { 305 free_pages((unsigned long)(__va(addr)), page_order); 306 } 307 308 vmemmap_remove_mapping(start, page_size); 309 } 310 } 311 #endif 312 void register_page_bootmem_memmap(unsigned long section_nr, 313 struct page *start_page, unsigned long size) 314 { 315 } 316 317 /* 318 * We do not have access to the sparsemem vmemmap, so we fallback to 319 * walking the list of sparsemem blocks which we already maintain for 320 * the sake of crashdump. In the long run, we might want to maintain 321 * a tree if performance of that linear walk becomes a problem. 322 * 323 * realmode_pfn_to_page functions can fail due to: 324 * 1) As real sparsemem blocks do not lay in RAM continously (they 325 * are in virtual address space which is not available in the real mode), 326 * the requested page struct can be split between blocks so get_page/put_page 327 * may fail. 328 * 2) When huge pages are used, the get_page/put_page API will fail 329 * in real mode as the linked addresses in the page struct are virtual 330 * too. 331 */ 332 struct page *realmode_pfn_to_page(unsigned long pfn) 333 { 334 struct vmemmap_backing *vmem_back; 335 struct page *page; 336 unsigned long page_size = 1 << mmu_psize_defs[mmu_vmemmap_psize].shift; 337 unsigned long pg_va = (unsigned long) pfn_to_page(pfn); 338 339 for (vmem_back = vmemmap_list; vmem_back; vmem_back = vmem_back->list) { 340 if (pg_va < vmem_back->virt_addr) 341 continue; 342 343 /* After vmemmap_list entry free is possible, need check all */ 344 if ((pg_va + sizeof(struct page)) <= 345 (vmem_back->virt_addr + page_size)) { 346 page = (struct page *) (vmem_back->phys + pg_va - 347 vmem_back->virt_addr); 348 return page; 349 } 350 } 351 352 /* Probably that page struct is split between real pages */ 353 return NULL; 354 } 355 EXPORT_SYMBOL_GPL(realmode_pfn_to_page); 356 357 #else 358 359 struct page *realmode_pfn_to_page(unsigned long pfn) 360 { 361 struct page *page = pfn_to_page(pfn); 362 return page; 363 } 364 EXPORT_SYMBOL_GPL(realmode_pfn_to_page); 365 366 #endif /* CONFIG_SPARSEMEM_VMEMMAP */ 367 368 #ifdef CONFIG_PPC_BOOK3S_64 369 static bool disable_radix = !IS_ENABLED(CONFIG_PPC_RADIX_MMU_DEFAULT); 370 371 static int __init parse_disable_radix(char *p) 372 { 373 bool val; 374 375 if (strlen(p) == 0) 376 val = true; 377 else if (kstrtobool(p, &val)) 378 return -EINVAL; 379 380 disable_radix = val; 381 382 return 0; 383 } 384 early_param("disable_radix", parse_disable_radix); 385 386 /* 387 * If we're running under a hypervisor, we need to check the contents of 388 * /chosen/ibm,architecture-vec-5 to see if the hypervisor is willing to do 389 * radix. If not, we clear the radix feature bit so we fall back to hash. 390 */ 391 static void __init early_check_vec5(void) 392 { 393 unsigned long root, chosen; 394 int size; 395 const u8 *vec5; 396 u8 mmu_supported; 397 398 root = of_get_flat_dt_root(); 399 chosen = of_get_flat_dt_subnode_by_name(root, "chosen"); 400 if (chosen == -FDT_ERR_NOTFOUND) { 401 cur_cpu_spec->mmu_features &= ~MMU_FTR_TYPE_RADIX; 402 return; 403 } 404 vec5 = of_get_flat_dt_prop(chosen, "ibm,architecture-vec-5", &size); 405 if (!vec5) { 406 cur_cpu_spec->mmu_features &= ~MMU_FTR_TYPE_RADIX; 407 return; 408 } 409 if (size <= OV5_INDX(OV5_MMU_SUPPORT)) { 410 cur_cpu_spec->mmu_features &= ~MMU_FTR_TYPE_RADIX; 411 return; 412 } 413 414 /* Check for supported configuration */ 415 mmu_supported = vec5[OV5_INDX(OV5_MMU_SUPPORT)] & 416 OV5_FEAT(OV5_MMU_SUPPORT); 417 if (mmu_supported == OV5_FEAT(OV5_MMU_RADIX)) { 418 /* Hypervisor only supports radix - check enabled && GTSE */ 419 if (!early_radix_enabled()) { 420 pr_warn("WARNING: Ignoring cmdline option disable_radix\n"); 421 } 422 if (!(vec5[OV5_INDX(OV5_RADIX_GTSE)] & 423 OV5_FEAT(OV5_RADIX_GTSE))) { 424 pr_warn("WARNING: Hypervisor doesn't support RADIX with GTSE\n"); 425 } 426 /* Do radix anyway - the hypervisor said we had to */ 427 cur_cpu_spec->mmu_features |= MMU_FTR_TYPE_RADIX; 428 } else if (mmu_supported == OV5_FEAT(OV5_MMU_HASH)) { 429 /* Hypervisor only supports hash - disable radix */ 430 cur_cpu_spec->mmu_features &= ~MMU_FTR_TYPE_RADIX; 431 } 432 } 433 434 void __init mmu_early_init_devtree(void) 435 { 436 /* Disable radix mode based on kernel command line. */ 437 if (disable_radix) 438 cur_cpu_spec->mmu_features &= ~MMU_FTR_TYPE_RADIX; 439 440 /* 441 * Check /chosen/ibm,architecture-vec-5 if running as a guest. 442 * When running bare-metal, we can use radix if we like 443 * even though the ibm,architecture-vec-5 property created by 444 * skiboot doesn't have the necessary bits set. 445 */ 446 if (!(mfmsr() & MSR_HV)) 447 early_check_vec5(); 448 449 if (early_radix_enabled()) 450 radix__early_init_devtree(); 451 else 452 hash__early_init_devtree(); 453 } 454 #endif /* CONFIG_PPC_BOOK3S_64 */ 455