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 #include <linux/memory.h> 44 45 #include <asm/pgalloc.h> 46 #include <asm/page.h> 47 #include <asm/prom.h> 48 #include <asm/rtas.h> 49 #include <asm/io.h> 50 #include <asm/mmu_context.h> 51 #include <asm/mmu.h> 52 #include <linux/uaccess.h> 53 #include <asm/smp.h> 54 #include <asm/machdep.h> 55 #include <asm/tlb.h> 56 #include <asm/eeh.h> 57 #include <asm/processor.h> 58 #include <asm/mmzone.h> 59 #include <asm/cputable.h> 60 #include <asm/sections.h> 61 #include <asm/iommu.h> 62 #include <asm/vdso.h> 63 #include <asm/hugetlb.h> 64 65 #include <mm/mmu_decl.h> 66 67 #ifdef CONFIG_SPARSEMEM_VMEMMAP 68 /* 69 * Given an address within the vmemmap, determine the page that 70 * represents the start of the subsection it is within. Note that we have to 71 * do this by hand as the proffered address may not be correctly aligned. 72 * Subtraction of non-aligned pointers produces undefined results. 73 */ 74 static struct page * __meminit vmemmap_subsection_start(unsigned long vmemmap_addr) 75 { 76 unsigned long start_pfn; 77 unsigned long offset = vmemmap_addr - ((unsigned long)(vmemmap)); 78 79 /* Return the pfn of the start of the section. */ 80 start_pfn = (offset / sizeof(struct page)) & PAGE_SUBSECTION_MASK; 81 return pfn_to_page(start_pfn); 82 } 83 84 /* 85 * Since memory is added in sub-section chunks, before creating a new vmemmap 86 * mapping, the kernel should check whether there is an existing memmap mapping 87 * covering the new subsection added. This is needed because kernel can map 88 * vmemmap area using 16MB pages which will cover a memory range of 16G. Such 89 * a range covers multiple subsections (2M) 90 * 91 * If any subsection in the 16G range mapped by vmemmap is valid we consider the 92 * vmemmap populated (There is a page table entry already present). We can't do 93 * a page table lookup here because with the hash translation we don't keep 94 * vmemmap details in linux page table. 95 */ 96 int __meminit vmemmap_populated(unsigned long vmemmap_addr, int vmemmap_map_size) 97 { 98 struct page *start; 99 unsigned long vmemmap_end = vmemmap_addr + vmemmap_map_size; 100 start = vmemmap_subsection_start(vmemmap_addr); 101 102 for (; (unsigned long)start < vmemmap_end; start += PAGES_PER_SUBSECTION) 103 /* 104 * pfn valid check here is intended to really check 105 * whether we have any subsection already initialized 106 * in this range. 107 */ 108 if (pfn_valid(page_to_pfn(start))) 109 return 1; 110 111 return 0; 112 } 113 114 /* 115 * vmemmap virtual address space management does not have a traditional 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 freed 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 int 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 pr_debug("vmemap list allocation failed\n"); 176 return -ENOMEM; 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 return 0; 185 } 186 187 bool altmap_cross_boundary(struct vmem_altmap *altmap, unsigned long start, 188 unsigned long page_size) 189 { 190 unsigned long nr_pfn = page_size / sizeof(struct page); 191 unsigned long start_pfn = page_to_pfn((struct page *)start); 192 193 if ((start_pfn + nr_pfn - 1) > altmap->end_pfn) 194 return true; 195 196 if (start_pfn < altmap->base_pfn) 197 return true; 198 199 return false; 200 } 201 202 static int __meminit __vmemmap_populate(unsigned long start, unsigned long end, int node, 203 struct vmem_altmap *altmap) 204 { 205 bool altmap_alloc; 206 unsigned long page_size = 1 << mmu_psize_defs[mmu_vmemmap_psize].shift; 207 208 /* Align to the page size of the linear mapping. */ 209 start = ALIGN_DOWN(start, page_size); 210 211 pr_debug("vmemmap_populate %lx..%lx, node %d\n", start, end, node); 212 213 for (; start < end; start += page_size) { 214 void *p = NULL; 215 int rc; 216 217 /* 218 * This vmemmap range is backing different subsections. If any 219 * of that subsection is marked valid, that means we already 220 * have initialized a page table covering this range and hence 221 * the vmemmap range is populated. 222 */ 223 if (vmemmap_populated(start, page_size)) 224 continue; 225 226 /* 227 * Allocate from the altmap first if we have one. This may 228 * fail due to alignment issues when using 16MB hugepages, so 229 * fall back to system memory if the altmap allocation fail. 230 */ 231 if (altmap && !altmap_cross_boundary(altmap, start, page_size)) { 232 p = vmemmap_alloc_block_buf(page_size, node, altmap); 233 if (!p) 234 pr_debug("altmap block allocation failed, falling back to system memory"); 235 else 236 altmap_alloc = true; 237 } 238 if (!p) { 239 p = vmemmap_alloc_block_buf(page_size, node, NULL); 240 altmap_alloc = false; 241 } 242 if (!p) 243 return -ENOMEM; 244 245 if (vmemmap_list_populate(__pa(p), start, node)) { 246 /* 247 * If we don't populate vmemap list, we don't have 248 * the ability to free the allocated vmemmap 249 * pages in section_deactivate. Hence free them 250 * here. 251 */ 252 int nr_pfns = page_size >> PAGE_SHIFT; 253 unsigned long page_order = get_order(page_size); 254 255 if (altmap_alloc) 256 vmem_altmap_free(altmap, nr_pfns); 257 else 258 free_pages((unsigned long)p, page_order); 259 return -ENOMEM; 260 } 261 262 pr_debug(" * %016lx..%016lx allocated at %p\n", 263 start, start + page_size, p); 264 265 rc = vmemmap_create_mapping(start, page_size, __pa(p)); 266 if (rc < 0) { 267 pr_warn("%s: Unable to create vmemmap mapping: %d\n", 268 __func__, rc); 269 return -EFAULT; 270 } 271 } 272 273 return 0; 274 } 275 276 int __meminit vmemmap_populate(unsigned long start, unsigned long end, int node, 277 struct vmem_altmap *altmap) 278 { 279 280 #ifdef CONFIG_PPC_BOOK3S_64 281 if (radix_enabled()) 282 return radix__vmemmap_populate(start, end, node, altmap); 283 #endif 284 285 return __vmemmap_populate(start, end, node, altmap); 286 } 287 288 #ifdef CONFIG_MEMORY_HOTPLUG 289 static unsigned long vmemmap_list_free(unsigned long start) 290 { 291 struct vmemmap_backing *vmem_back, *vmem_back_prev; 292 293 vmem_back_prev = vmem_back = vmemmap_list; 294 295 /* look for it with prev pointer recorded */ 296 for (; vmem_back; vmem_back = vmem_back->list) { 297 if (vmem_back->virt_addr == start) 298 break; 299 vmem_back_prev = vmem_back; 300 } 301 302 if (unlikely(!vmem_back)) 303 return 0; 304 305 /* remove it from vmemmap_list */ 306 if (vmem_back == vmemmap_list) /* remove head */ 307 vmemmap_list = vmem_back->list; 308 else 309 vmem_back_prev->list = vmem_back->list; 310 311 /* next point to this freed entry */ 312 vmem_back->list = next; 313 next = vmem_back; 314 num_freed++; 315 316 return vmem_back->phys; 317 } 318 319 static void __ref __vmemmap_free(unsigned long start, unsigned long end, 320 struct vmem_altmap *altmap) 321 { 322 unsigned long page_size = 1 << mmu_psize_defs[mmu_vmemmap_psize].shift; 323 unsigned long page_order = get_order(page_size); 324 unsigned long alt_start = ~0, alt_end = ~0; 325 unsigned long base_pfn; 326 327 start = ALIGN_DOWN(start, page_size); 328 if (altmap) { 329 alt_start = altmap->base_pfn; 330 alt_end = altmap->base_pfn + altmap->reserve + altmap->free; 331 } 332 333 pr_debug("vmemmap_free %lx...%lx\n", start, end); 334 335 for (; start < end; start += page_size) { 336 unsigned long nr_pages, addr; 337 struct page *page; 338 339 /* 340 * We have already marked the subsection we are trying to remove 341 * invalid. So if we want to remove the vmemmap range, we 342 * need to make sure there is no subsection marked valid 343 * in this range. 344 */ 345 if (vmemmap_populated(start, page_size)) 346 continue; 347 348 addr = vmemmap_list_free(start); 349 if (!addr) 350 continue; 351 352 page = pfn_to_page(addr >> PAGE_SHIFT); 353 nr_pages = 1 << page_order; 354 base_pfn = PHYS_PFN(addr); 355 356 if (base_pfn >= alt_start && base_pfn < alt_end) { 357 vmem_altmap_free(altmap, nr_pages); 358 } else if (PageReserved(page)) { 359 /* allocated from bootmem */ 360 if (page_size < PAGE_SIZE) { 361 /* 362 * this shouldn't happen, but if it is 363 * the case, leave the memory there 364 */ 365 WARN_ON_ONCE(1); 366 } else { 367 while (nr_pages--) 368 free_reserved_page(page++); 369 } 370 } else { 371 free_pages((unsigned long)(__va(addr)), page_order); 372 } 373 374 vmemmap_remove_mapping(start, page_size); 375 } 376 } 377 378 void __ref vmemmap_free(unsigned long start, unsigned long end, 379 struct vmem_altmap *altmap) 380 { 381 #ifdef CONFIG_PPC_BOOK3S_64 382 if (radix_enabled()) 383 return radix__vmemmap_free(start, end, altmap); 384 #endif 385 return __vmemmap_free(start, end, altmap); 386 } 387 388 #endif 389 390 #endif /* CONFIG_SPARSEMEM_VMEMMAP */ 391 392 #ifdef CONFIG_PPC_BOOK3S_64 393 unsigned int mmu_lpid_bits; 394 #ifdef CONFIG_KVM_BOOK3S_HV_POSSIBLE 395 EXPORT_SYMBOL_GPL(mmu_lpid_bits); 396 #endif 397 unsigned int mmu_pid_bits; 398 399 static bool disable_radix = !IS_ENABLED(CONFIG_PPC_RADIX_MMU_DEFAULT); 400 401 static int __init parse_disable_radix(char *p) 402 { 403 bool val; 404 405 if (!p) 406 val = true; 407 else if (kstrtobool(p, &val)) 408 return -EINVAL; 409 410 disable_radix = val; 411 412 return 0; 413 } 414 early_param("disable_radix", parse_disable_radix); 415 416 /* 417 * If we're running under a hypervisor, we need to check the contents of 418 * /chosen/ibm,architecture-vec-5 to see if the hypervisor is willing to do 419 * radix. If not, we clear the radix feature bit so we fall back to hash. 420 */ 421 static void __init early_check_vec5(void) 422 { 423 unsigned long root, chosen; 424 int size; 425 const u8 *vec5; 426 u8 mmu_supported; 427 428 root = of_get_flat_dt_root(); 429 chosen = of_get_flat_dt_subnode_by_name(root, "chosen"); 430 if (chosen == -FDT_ERR_NOTFOUND) { 431 cur_cpu_spec->mmu_features &= ~MMU_FTR_TYPE_RADIX; 432 return; 433 } 434 vec5 = of_get_flat_dt_prop(chosen, "ibm,architecture-vec-5", &size); 435 if (!vec5) { 436 cur_cpu_spec->mmu_features &= ~MMU_FTR_TYPE_RADIX; 437 return; 438 } 439 if (size <= OV5_INDX(OV5_MMU_SUPPORT)) { 440 cur_cpu_spec->mmu_features &= ~MMU_FTR_TYPE_RADIX; 441 return; 442 } 443 444 /* Check for supported configuration */ 445 mmu_supported = vec5[OV5_INDX(OV5_MMU_SUPPORT)] & 446 OV5_FEAT(OV5_MMU_SUPPORT); 447 if (mmu_supported == OV5_FEAT(OV5_MMU_RADIX)) { 448 /* Hypervisor only supports radix - check enabled && GTSE */ 449 if (!early_radix_enabled()) { 450 pr_warn("WARNING: Ignoring cmdline option disable_radix\n"); 451 } 452 if (!(vec5[OV5_INDX(OV5_RADIX_GTSE)] & 453 OV5_FEAT(OV5_RADIX_GTSE))) { 454 cur_cpu_spec->mmu_features &= ~MMU_FTR_GTSE; 455 } else 456 cur_cpu_spec->mmu_features |= MMU_FTR_GTSE; 457 /* Do radix anyway - the hypervisor said we had to */ 458 cur_cpu_spec->mmu_features |= MMU_FTR_TYPE_RADIX; 459 } else if (mmu_supported == OV5_FEAT(OV5_MMU_HASH)) { 460 /* Hypervisor only supports hash - disable radix */ 461 cur_cpu_spec->mmu_features &= ~MMU_FTR_TYPE_RADIX; 462 cur_cpu_spec->mmu_features &= ~MMU_FTR_GTSE; 463 } 464 } 465 466 static int __init dt_scan_mmu_pid_width(unsigned long node, 467 const char *uname, int depth, 468 void *data) 469 { 470 int size = 0; 471 const __be32 *prop; 472 const char *type = of_get_flat_dt_prop(node, "device_type", NULL); 473 474 /* We are scanning "cpu" nodes only */ 475 if (type == NULL || strcmp(type, "cpu") != 0) 476 return 0; 477 478 /* Find MMU LPID, PID register size */ 479 prop = of_get_flat_dt_prop(node, "ibm,mmu-lpid-bits", &size); 480 if (prop && size == 4) 481 mmu_lpid_bits = be32_to_cpup(prop); 482 483 prop = of_get_flat_dt_prop(node, "ibm,mmu-pid-bits", &size); 484 if (prop && size == 4) 485 mmu_pid_bits = be32_to_cpup(prop); 486 487 if (!mmu_pid_bits && !mmu_lpid_bits) 488 return 0; 489 490 return 1; 491 } 492 493 /* 494 * Outside hotplug the kernel uses this value to map the kernel direct map 495 * with radix. To be compatible with older kernels, let's keep this value 496 * as 16M which is also SECTION_SIZE with SPARSEMEM. We can ideally map 497 * things with 1GB size in the case where we don't support hotplug. 498 */ 499 #ifndef CONFIG_MEMORY_HOTPLUG 500 #define DEFAULT_MEMORY_BLOCK_SIZE SZ_16M 501 #else 502 #define DEFAULT_MEMORY_BLOCK_SIZE MIN_MEMORY_BLOCK_SIZE 503 #endif 504 505 static void update_memory_block_size(unsigned long *block_size, unsigned long mem_size) 506 { 507 unsigned long min_memory_block_size = DEFAULT_MEMORY_BLOCK_SIZE; 508 509 for (; *block_size > min_memory_block_size; *block_size >>= 2) { 510 if ((mem_size & *block_size) == 0) 511 break; 512 } 513 } 514 515 static int __init probe_memory_block_size(unsigned long node, const char *uname, int 516 depth, void *data) 517 { 518 const char *type; 519 unsigned long *block_size = (unsigned long *)data; 520 const __be32 *reg, *endp; 521 int l; 522 523 if (depth != 1) 524 return 0; 525 /* 526 * If we have dynamic-reconfiguration-memory node, use the 527 * lmb value. 528 */ 529 if (strcmp(uname, "ibm,dynamic-reconfiguration-memory") == 0) { 530 531 const __be32 *prop; 532 533 prop = of_get_flat_dt_prop(node, "ibm,lmb-size", &l); 534 535 if (!prop || l < dt_root_size_cells * sizeof(__be32)) 536 /* 537 * Nothing in the device tree 538 */ 539 *block_size = DEFAULT_MEMORY_BLOCK_SIZE; 540 else 541 *block_size = of_read_number(prop, dt_root_size_cells); 542 /* 543 * We have found the final value. Don't probe further. 544 */ 545 return 1; 546 } 547 /* 548 * Find all the device tree nodes of memory type and make sure 549 * the area can be mapped using the memory block size value 550 * we end up using. We start with 1G value and keep reducing 551 * it such that we can map the entire area using memory_block_size. 552 * This will be used on powernv and older pseries that don't 553 * have ibm,lmb-size node. 554 * For ex: with P5 we can end up with 555 * memory@0 -> 128MB 556 * memory@128M -> 64M 557 * This will end up using 64MB memory block size value. 558 */ 559 type = of_get_flat_dt_prop(node, "device_type", NULL); 560 if (type == NULL || strcmp(type, "memory") != 0) 561 return 0; 562 563 reg = of_get_flat_dt_prop(node, "linux,usable-memory", &l); 564 if (!reg) 565 reg = of_get_flat_dt_prop(node, "reg", &l); 566 if (!reg) 567 return 0; 568 569 endp = reg + (l / sizeof(__be32)); 570 while ((endp - reg) >= (dt_root_addr_cells + dt_root_size_cells)) { 571 const char *compatible; 572 u64 size; 573 574 dt_mem_next_cell(dt_root_addr_cells, ®); 575 size = dt_mem_next_cell(dt_root_size_cells, ®); 576 577 if (size) { 578 update_memory_block_size(block_size, size); 579 continue; 580 } 581 /* 582 * ibm,coherent-device-memory with linux,usable-memory = 0 583 * Force 256MiB block size. Work around for GPUs on P9 PowerNV 584 * linux,usable-memory == 0 implies driver managed memory and 585 * we can't use large memory block size due to hotplug/unplug 586 * limitations. 587 */ 588 compatible = of_get_flat_dt_prop(node, "compatible", NULL); 589 if (compatible && !strcmp(compatible, "ibm,coherent-device-memory")) { 590 if (*block_size > SZ_256M) 591 *block_size = SZ_256M; 592 /* 593 * We keep 256M as the upper limit with GPU present. 594 */ 595 return 0; 596 } 597 } 598 /* continue looking for other memory device types */ 599 return 0; 600 } 601 602 /* 603 * start with 1G memory block size. Early init will 604 * fix this with correct value. 605 */ 606 unsigned long memory_block_size __ro_after_init = 1UL << 30; 607 static void __init early_init_memory_block_size(void) 608 { 609 /* 610 * We need to do memory_block_size probe early so that 611 * radix__early_init_mmu() can use this as limit for 612 * mapping page size. 613 */ 614 of_scan_flat_dt(probe_memory_block_size, &memory_block_size); 615 } 616 617 void __init mmu_early_init_devtree(void) 618 { 619 bool hvmode = !!(mfmsr() & MSR_HV); 620 621 /* Disable radix mode based on kernel command line. */ 622 if (disable_radix) { 623 if (IS_ENABLED(CONFIG_PPC_64S_HASH_MMU)) 624 cur_cpu_spec->mmu_features &= ~MMU_FTR_TYPE_RADIX; 625 else 626 pr_warn("WARNING: Ignoring cmdline option disable_radix\n"); 627 } 628 629 of_scan_flat_dt(dt_scan_mmu_pid_width, NULL); 630 if (hvmode && !mmu_lpid_bits) { 631 if (early_cpu_has_feature(CPU_FTR_ARCH_207S)) 632 mmu_lpid_bits = 12; /* POWER8-10 */ 633 else 634 mmu_lpid_bits = 10; /* POWER7 */ 635 } 636 if (!mmu_pid_bits) { 637 if (early_cpu_has_feature(CPU_FTR_ARCH_300)) 638 mmu_pid_bits = 20; /* POWER9-10 */ 639 } 640 641 /* 642 * Check /chosen/ibm,architecture-vec-5 if running as a guest. 643 * When running bare-metal, we can use radix if we like 644 * even though the ibm,architecture-vec-5 property created by 645 * skiboot doesn't have the necessary bits set. 646 */ 647 if (!hvmode) 648 early_check_vec5(); 649 650 early_init_memory_block_size(); 651 652 if (early_radix_enabled()) { 653 radix__early_init_devtree(); 654 655 /* 656 * We have finalized the translation we are going to use by now. 657 * Radix mode is not limited by RMA / VRMA addressing. 658 * Hence don't limit memblock allocations. 659 */ 660 ppc64_rma_size = ULONG_MAX; 661 memblock_set_current_limit(MEMBLOCK_ALLOC_ANYWHERE); 662 } else 663 hash__early_init_devtree(); 664 665 if (IS_ENABLED(CONFIG_HUGETLB_PAGE_SIZE_VARIABLE)) 666 hugetlbpage_init_defaultsize(); 667 668 if (!(cur_cpu_spec->mmu_features & MMU_FTR_HPTE_TABLE) && 669 !(cur_cpu_spec->mmu_features & MMU_FTR_TYPE_RADIX)) 670 panic("kernel does not support any MMU type offered by platform"); 671 } 672 #endif /* CONFIG_PPC_BOOK3S_64 */ 673