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