1 // SPDX-License-Identifier: GPL-2.0-or-later 2 /* 3 * PowerPC version 4 * Copyright (C) 1995-1996 Gary Thomas (gdt@linuxppc.org) 5 * 6 * Modifications by Paul Mackerras (PowerMac) (paulus@cs.anu.edu.au) 7 * and Cort Dougan (PReP) (cort@cs.nmt.edu) 8 * Copyright (C) 1996 Paul Mackerras 9 * 10 * Derived from "arch/i386/mm/init.c" 11 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds 12 * 13 * Dave Engebretsen <engebret@us.ibm.com> 14 * Rework for PPC64 port. 15 */ 16 17 #undef DEBUG 18 19 #include <linux/signal.h> 20 #include <linux/sched.h> 21 #include <linux/kernel.h> 22 #include <linux/errno.h> 23 #include <linux/string.h> 24 #include <linux/types.h> 25 #include <linux/mman.h> 26 #include <linux/mm.h> 27 #include <linux/swap.h> 28 #include <linux/stddef.h> 29 #include <linux/vmalloc.h> 30 #include <linux/init.h> 31 #include <linux/delay.h> 32 #include <linux/highmem.h> 33 #include <linux/idr.h> 34 #include <linux/nodemask.h> 35 #include <linux/module.h> 36 #include <linux/poison.h> 37 #include <linux/memblock.h> 38 #include <linux/hugetlb.h> 39 #include <linux/slab.h> 40 #include <linux/of_fdt.h> 41 #include <linux/libfdt.h> 42 #include <linux/memremap.h> 43 44 #include <asm/pgalloc.h> 45 #include <asm/page.h> 46 #include <asm/prom.h> 47 #include <asm/rtas.h> 48 #include <asm/io.h> 49 #include <asm/mmu_context.h> 50 #include <asm/mmu.h> 51 #include <linux/uaccess.h> 52 #include <asm/smp.h> 53 #include <asm/machdep.h> 54 #include <asm/tlb.h> 55 #include <asm/eeh.h> 56 #include <asm/processor.h> 57 #include <asm/mmzone.h> 58 #include <asm/cputable.h> 59 #include <asm/sections.h> 60 #include <asm/iommu.h> 61 #include <asm/vdso.h> 62 63 #include <mm/mmu_decl.h> 64 65 #ifdef CONFIG_SPARSEMEM_VMEMMAP 66 /* 67 * Given an address within the vmemmap, determine the page that 68 * represents the start of the subsection it is within. Note that we have to 69 * do this by hand as the proffered address may not be correctly aligned. 70 * Subtraction of non-aligned pointers produces undefined results. 71 */ 72 static struct page * __meminit vmemmap_subsection_start(unsigned long vmemmap_addr) 73 { 74 unsigned long start_pfn; 75 unsigned long offset = vmemmap_addr - ((unsigned long)(vmemmap)); 76 77 /* Return the pfn of the start of the section. */ 78 start_pfn = (offset / sizeof(struct page)) & PAGE_SUBSECTION_MASK; 79 return pfn_to_page(start_pfn); 80 } 81 82 /* 83 * Since memory is added in sub-section chunks, before creating a new vmemmap 84 * mapping, the kernel should check whether there is an existing memmap mapping 85 * covering the new subsection added. This is needed because kernel can map 86 * vmemmap area using 16MB pages which will cover a memory range of 16G. Such 87 * a range covers multiple subsections (2M) 88 * 89 * If any subsection in the 16G range mapped by vmemmap is valid we consider the 90 * vmemmap populated (There is a page table entry already present). We can't do 91 * a page table lookup here because with the hash translation we don't keep 92 * vmemmap details in linux page table. 93 */ 94 static int __meminit vmemmap_populated(unsigned long vmemmap_addr, int vmemmap_map_size) 95 { 96 struct page *start; 97 unsigned long vmemmap_end = vmemmap_addr + vmemmap_map_size; 98 start = vmemmap_subsection_start(vmemmap_addr); 99 100 for (; (unsigned long)start < vmemmap_end; start += PAGES_PER_SUBSECTION) 101 /* 102 * pfn valid check here is intended to really check 103 * whether we have any subsection already initialized 104 * in this range. 105 */ 106 if (pfn_valid(page_to_pfn(start))) 107 return 1; 108 109 return 0; 110 } 111 112 /* 113 * vmemmap virtual address space management does not have a traditonal page 114 * table to track which virtual struct pages are backed by physical mapping. 115 * The virtual to physical mappings are tracked in a simple linked list 116 * format. 'vmemmap_list' maintains the entire vmemmap physical mapping at 117 * all times where as the 'next' list maintains the available 118 * vmemmap_backing structures which have been deleted from the 119 * 'vmemmap_global' list during system runtime (memory hotplug remove 120 * operation). The freed 'vmemmap_backing' structures are reused later when 121 * new requests come in without allocating fresh memory. This pointer also 122 * tracks the allocated 'vmemmap_backing' structures as we allocate one 123 * full page memory at a time when we dont have any. 124 */ 125 struct vmemmap_backing *vmemmap_list; 126 static struct vmemmap_backing *next; 127 128 /* 129 * The same pointer 'next' tracks individual chunks inside the allocated 130 * full page during the boot time and again tracks the freeed nodes during 131 * runtime. It is racy but it does not happen as they are separated by the 132 * boot process. Will create problem if some how we have memory hotplug 133 * operation during boot !! 134 */ 135 static int num_left; 136 static int num_freed; 137 138 static __meminit struct vmemmap_backing * vmemmap_list_alloc(int node) 139 { 140 struct vmemmap_backing *vmem_back; 141 /* get from freed entries first */ 142 if (num_freed) { 143 num_freed--; 144 vmem_back = next; 145 next = next->list; 146 147 return vmem_back; 148 } 149 150 /* allocate a page when required and hand out chunks */ 151 if (!num_left) { 152 next = vmemmap_alloc_block(PAGE_SIZE, node); 153 if (unlikely(!next)) { 154 WARN_ON(1); 155 return NULL; 156 } 157 num_left = PAGE_SIZE / sizeof(struct vmemmap_backing); 158 } 159 160 num_left--; 161 162 return next++; 163 } 164 165 static __meminit int vmemmap_list_populate(unsigned long phys, 166 unsigned long start, 167 int node) 168 { 169 struct vmemmap_backing *vmem_back; 170 171 vmem_back = vmemmap_list_alloc(node); 172 if (unlikely(!vmem_back)) { 173 pr_debug("vmemap list allocation failed\n"); 174 return -ENOMEM; 175 } 176 177 vmem_back->phys = phys; 178 vmem_back->virt_addr = start; 179 vmem_back->list = vmemmap_list; 180 181 vmemmap_list = vmem_back; 182 return 0; 183 } 184 185 static bool altmap_cross_boundary(struct vmem_altmap *altmap, unsigned long start, 186 unsigned long page_size) 187 { 188 unsigned long nr_pfn = page_size / sizeof(struct page); 189 unsigned long start_pfn = page_to_pfn((struct page *)start); 190 191 if ((start_pfn + nr_pfn) > altmap->end_pfn) 192 return true; 193 194 if (start_pfn < altmap->base_pfn) 195 return true; 196 197 return false; 198 } 199 200 int __meminit vmemmap_populate(unsigned long start, unsigned long end, int node, 201 struct vmem_altmap *altmap) 202 { 203 bool altmap_alloc; 204 unsigned long page_size = 1 << mmu_psize_defs[mmu_vmemmap_psize].shift; 205 206 /* Align to the page size of the linear mapping. */ 207 start = ALIGN_DOWN(start, page_size); 208 209 pr_debug("vmemmap_populate %lx..%lx, node %d\n", start, end, node); 210 211 for (; start < end; start += page_size) { 212 void *p = NULL; 213 int rc; 214 215 /* 216 * This vmemmap range is backing different subsections. If any 217 * of that subsection is marked valid, that means we already 218 * have initialized a page table covering this range and hence 219 * the vmemmap range is populated. 220 */ 221 if (vmemmap_populated(start, page_size)) 222 continue; 223 224 /* 225 * Allocate from the altmap first if we have one. This may 226 * fail due to alignment issues when using 16MB hugepages, so 227 * fall back to system memory if the altmap allocation fail. 228 */ 229 if (altmap && !altmap_cross_boundary(altmap, start, page_size)) { 230 p = vmemmap_alloc_block_buf(page_size, node, altmap); 231 if (!p) 232 pr_debug("altmap block allocation failed, falling back to system memory"); 233 else 234 altmap_alloc = true; 235 } 236 if (!p) { 237 p = vmemmap_alloc_block_buf(page_size, node, NULL); 238 altmap_alloc = false; 239 } 240 if (!p) 241 return -ENOMEM; 242 243 if (vmemmap_list_populate(__pa(p), start, node)) { 244 /* 245 * If we don't populate vmemap list, we don't have 246 * the ability to free the allocated vmemmap 247 * pages in section_deactivate. Hence free them 248 * here. 249 */ 250 int nr_pfns = page_size >> PAGE_SHIFT; 251 unsigned long page_order = get_order(page_size); 252 253 if (altmap_alloc) 254 vmem_altmap_free(altmap, nr_pfns); 255 else 256 free_pages((unsigned long)p, page_order); 257 return -ENOMEM; 258 } 259 260 pr_debug(" * %016lx..%016lx allocated at %p\n", 261 start, start + page_size, p); 262 263 rc = vmemmap_create_mapping(start, page_size, __pa(p)); 264 if (rc < 0) { 265 pr_warn("%s: Unable to create vmemmap mapping: %d\n", 266 __func__, rc); 267 return -EFAULT; 268 } 269 } 270 271 return 0; 272 } 273 274 #ifdef CONFIG_MEMORY_HOTPLUG 275 static unsigned long vmemmap_list_free(unsigned long start) 276 { 277 struct vmemmap_backing *vmem_back, *vmem_back_prev; 278 279 vmem_back_prev = vmem_back = vmemmap_list; 280 281 /* look for it with prev pointer recorded */ 282 for (; vmem_back; vmem_back = vmem_back->list) { 283 if (vmem_back->virt_addr == start) 284 break; 285 vmem_back_prev = vmem_back; 286 } 287 288 if (unlikely(!vmem_back)) 289 return 0; 290 291 /* remove it from vmemmap_list */ 292 if (vmem_back == vmemmap_list) /* remove head */ 293 vmemmap_list = vmem_back->list; 294 else 295 vmem_back_prev->list = vmem_back->list; 296 297 /* next point to this freed entry */ 298 vmem_back->list = next; 299 next = vmem_back; 300 num_freed++; 301 302 return vmem_back->phys; 303 } 304 305 void __ref vmemmap_free(unsigned long start, unsigned long end, 306 struct vmem_altmap *altmap) 307 { 308 unsigned long page_size = 1 << mmu_psize_defs[mmu_vmemmap_psize].shift; 309 unsigned long page_order = get_order(page_size); 310 unsigned long alt_start = ~0, alt_end = ~0; 311 unsigned long base_pfn; 312 313 start = ALIGN_DOWN(start, page_size); 314 if (altmap) { 315 alt_start = altmap->base_pfn; 316 alt_end = altmap->base_pfn + altmap->reserve + 317 altmap->free + altmap->alloc + altmap->align; 318 } 319 320 pr_debug("vmemmap_free %lx...%lx\n", start, end); 321 322 for (; start < end; start += page_size) { 323 unsigned long nr_pages, addr; 324 struct page *page; 325 326 /* 327 * We have already marked the subsection we are trying to remove 328 * invalid. So if we want to remove the vmemmap range, we 329 * need to make sure there is no subsection marked valid 330 * in this range. 331 */ 332 if (vmemmap_populated(start, page_size)) 333 continue; 334 335 addr = vmemmap_list_free(start); 336 if (!addr) 337 continue; 338 339 page = pfn_to_page(addr >> PAGE_SHIFT); 340 nr_pages = 1 << page_order; 341 base_pfn = PHYS_PFN(addr); 342 343 if (base_pfn >= alt_start && base_pfn < alt_end) { 344 vmem_altmap_free(altmap, nr_pages); 345 } else if (PageReserved(page)) { 346 /* allocated from bootmem */ 347 if (page_size < PAGE_SIZE) { 348 /* 349 * this shouldn't happen, but if it is 350 * the case, leave the memory there 351 */ 352 WARN_ON_ONCE(1); 353 } else { 354 while (nr_pages--) 355 free_reserved_page(page++); 356 } 357 } else { 358 free_pages((unsigned long)(__va(addr)), page_order); 359 } 360 361 vmemmap_remove_mapping(start, page_size); 362 } 363 } 364 #endif 365 void register_page_bootmem_memmap(unsigned long section_nr, 366 struct page *start_page, unsigned long size) 367 { 368 } 369 370 #endif /* CONFIG_SPARSEMEM_VMEMMAP */ 371 372 #ifdef CONFIG_PPC_BOOK3S_64 373 unsigned int mmu_lpid_bits; 374 unsigned int mmu_pid_bits; 375 376 static bool disable_radix = !IS_ENABLED(CONFIG_PPC_RADIX_MMU_DEFAULT); 377 378 static int __init parse_disable_radix(char *p) 379 { 380 bool val; 381 382 if (!p) 383 val = true; 384 else if (kstrtobool(p, &val)) 385 return -EINVAL; 386 387 disable_radix = val; 388 389 return 0; 390 } 391 early_param("disable_radix", parse_disable_radix); 392 393 /* 394 * If we're running under a hypervisor, we need to check the contents of 395 * /chosen/ibm,architecture-vec-5 to see if the hypervisor is willing to do 396 * radix. If not, we clear the radix feature bit so we fall back to hash. 397 */ 398 static void __init early_check_vec5(void) 399 { 400 unsigned long root, chosen; 401 int size; 402 const u8 *vec5; 403 u8 mmu_supported; 404 405 root = of_get_flat_dt_root(); 406 chosen = of_get_flat_dt_subnode_by_name(root, "chosen"); 407 if (chosen == -FDT_ERR_NOTFOUND) { 408 cur_cpu_spec->mmu_features &= ~MMU_FTR_TYPE_RADIX; 409 return; 410 } 411 vec5 = of_get_flat_dt_prop(chosen, "ibm,architecture-vec-5", &size); 412 if (!vec5) { 413 cur_cpu_spec->mmu_features &= ~MMU_FTR_TYPE_RADIX; 414 return; 415 } 416 if (size <= OV5_INDX(OV5_MMU_SUPPORT)) { 417 cur_cpu_spec->mmu_features &= ~MMU_FTR_TYPE_RADIX; 418 return; 419 } 420 421 /* Check for supported configuration */ 422 mmu_supported = vec5[OV5_INDX(OV5_MMU_SUPPORT)] & 423 OV5_FEAT(OV5_MMU_SUPPORT); 424 if (mmu_supported == OV5_FEAT(OV5_MMU_RADIX)) { 425 /* Hypervisor only supports radix - check enabled && GTSE */ 426 if (!early_radix_enabled()) { 427 pr_warn("WARNING: Ignoring cmdline option disable_radix\n"); 428 } 429 if (!(vec5[OV5_INDX(OV5_RADIX_GTSE)] & 430 OV5_FEAT(OV5_RADIX_GTSE))) { 431 cur_cpu_spec->mmu_features &= ~MMU_FTR_GTSE; 432 } else 433 cur_cpu_spec->mmu_features |= MMU_FTR_GTSE; 434 /* Do radix anyway - the hypervisor said we had to */ 435 cur_cpu_spec->mmu_features |= MMU_FTR_TYPE_RADIX; 436 } else if (mmu_supported == OV5_FEAT(OV5_MMU_HASH)) { 437 /* Hypervisor only supports hash - disable radix */ 438 cur_cpu_spec->mmu_features &= ~MMU_FTR_TYPE_RADIX; 439 cur_cpu_spec->mmu_features &= ~MMU_FTR_GTSE; 440 } 441 } 442 443 static int __init dt_scan_mmu_pid_width(unsigned long node, 444 const char *uname, int depth, 445 void *data) 446 { 447 int size = 0; 448 const __be32 *prop; 449 const char *type = of_get_flat_dt_prop(node, "device_type", NULL); 450 451 /* We are scanning "cpu" nodes only */ 452 if (type == NULL || strcmp(type, "cpu") != 0) 453 return 0; 454 455 /* Find MMU LPID, PID register size */ 456 prop = of_get_flat_dt_prop(node, "ibm,mmu-lpid-bits", &size); 457 if (prop && size == 4) 458 mmu_lpid_bits = be32_to_cpup(prop); 459 460 prop = of_get_flat_dt_prop(node, "ibm,mmu-pid-bits", &size); 461 if (prop && size == 4) 462 mmu_pid_bits = be32_to_cpup(prop); 463 464 if (!mmu_pid_bits && !mmu_lpid_bits) 465 return 0; 466 467 return 1; 468 } 469 470 void __init mmu_early_init_devtree(void) 471 { 472 bool hvmode = !!(mfmsr() & MSR_HV); 473 474 /* Disable radix mode based on kernel command line. */ 475 if (disable_radix) { 476 if (IS_ENABLED(CONFIG_PPC_64S_HASH_MMU)) 477 cur_cpu_spec->mmu_features &= ~MMU_FTR_TYPE_RADIX; 478 else 479 pr_warn("WARNING: Ignoring cmdline option disable_radix\n"); 480 } 481 482 of_scan_flat_dt(dt_scan_mmu_pid_width, NULL); 483 if (hvmode && !mmu_lpid_bits) { 484 if (early_cpu_has_feature(CPU_FTR_ARCH_207S)) 485 mmu_lpid_bits = 12; /* POWER8-10 */ 486 else 487 mmu_lpid_bits = 10; /* POWER7 */ 488 } 489 if (!mmu_pid_bits) { 490 if (early_cpu_has_feature(CPU_FTR_ARCH_300)) 491 mmu_pid_bits = 20; /* POWER9-10 */ 492 } 493 494 /* 495 * Check /chosen/ibm,architecture-vec-5 if running as a guest. 496 * When running bare-metal, we can use radix if we like 497 * even though the ibm,architecture-vec-5 property created by 498 * skiboot doesn't have the necessary bits set. 499 */ 500 if (!hvmode) 501 early_check_vec5(); 502 503 if (early_radix_enabled()) { 504 radix__early_init_devtree(); 505 506 /* 507 * We have finalized the translation we are going to use by now. 508 * Radix mode is not limited by RMA / VRMA addressing. 509 * Hence don't limit memblock allocations. 510 */ 511 ppc64_rma_size = ULONG_MAX; 512 memblock_set_current_limit(MEMBLOCK_ALLOC_ANYWHERE); 513 } else 514 hash__early_init_devtree(); 515 516 if (!(cur_cpu_spec->mmu_features & MMU_FTR_HPTE_TABLE) && 517 !(cur_cpu_spec->mmu_features & MMU_FTR_TYPE_RADIX)) 518 panic("kernel does not support any MMU type offered by platform"); 519 } 520 #endif /* CONFIG_PPC_BOOK3S_64 */ 521