Linux Audio

Check our new training course

Loading...
Note: File does not exist in v6.8.
   1/*
   2 * Page table handling routines for radix page table.
   3 *
   4 * Copyright 2015-2016, Aneesh Kumar K.V, IBM Corporation.
   5 *
   6 * This program is free software; you can redistribute it and/or
   7 * modify it under the terms of the GNU General Public License
   8 * as published by the Free Software Foundation; either version
   9 * 2 of the License, or (at your option) any later version.
  10 */
  11
  12#define pr_fmt(fmt) "radix-mmu: " fmt
  13
  14#include <linux/kernel.h>
  15#include <linux/sched/mm.h>
  16#include <linux/memblock.h>
  17#include <linux/of_fdt.h>
  18#include <linux/mm.h>
  19#include <linux/string_helpers.h>
  20#include <linux/stop_machine.h>
  21
  22#include <asm/pgtable.h>
  23#include <asm/pgalloc.h>
  24#include <asm/mmu_context.h>
  25#include <asm/dma.h>
  26#include <asm/machdep.h>
  27#include <asm/mmu.h>
  28#include <asm/firmware.h>
  29#include <asm/powernv.h>
  30#include <asm/sections.h>
  31#include <asm/trace.h>
  32
  33#include <trace/events/thp.h>
  34
  35unsigned int mmu_pid_bits;
  36unsigned int mmu_base_pid;
  37
  38static int native_register_process_table(unsigned long base, unsigned long pg_sz,
  39					 unsigned long table_size)
  40{
  41	unsigned long patb0, patb1;
  42
  43	patb0 = be64_to_cpu(partition_tb[0].patb0);
  44	patb1 = base | table_size | PATB_GR;
  45
  46	mmu_partition_table_set_entry(0, patb0, patb1);
  47
  48	return 0;
  49}
  50
  51static __ref void *early_alloc_pgtable(unsigned long size, int nid,
  52			unsigned long region_start, unsigned long region_end)
  53{
  54	unsigned long pa = 0;
  55	void *pt;
  56
  57	if (region_start || region_end) /* has region hint */
  58		pa = memblock_alloc_range(size, size, region_start, region_end,
  59						MEMBLOCK_NONE);
  60	else if (nid != -1) /* has node hint */
  61		pa = memblock_alloc_base_nid(size, size,
  62						MEMBLOCK_ALLOC_ANYWHERE,
  63						nid, MEMBLOCK_NONE);
  64
  65	if (!pa)
  66		pa = memblock_alloc_base(size, size, MEMBLOCK_ALLOC_ANYWHERE);
  67
  68	BUG_ON(!pa);
  69
  70	pt = __va(pa);
  71	memset(pt, 0, size);
  72
  73	return pt;
  74}
  75
  76static int early_map_kernel_page(unsigned long ea, unsigned long pa,
  77			  pgprot_t flags,
  78			  unsigned int map_page_size,
  79			  int nid,
  80			  unsigned long region_start, unsigned long region_end)
  81{
  82	unsigned long pfn = pa >> PAGE_SHIFT;
  83	pgd_t *pgdp;
  84	pud_t *pudp;
  85	pmd_t *pmdp;
  86	pte_t *ptep;
  87
  88	pgdp = pgd_offset_k(ea);
  89	if (pgd_none(*pgdp)) {
  90		pudp = early_alloc_pgtable(PUD_TABLE_SIZE, nid,
  91						region_start, region_end);
  92		pgd_populate(&init_mm, pgdp, pudp);
  93	}
  94	pudp = pud_offset(pgdp, ea);
  95	if (map_page_size == PUD_SIZE) {
  96		ptep = (pte_t *)pudp;
  97		goto set_the_pte;
  98	}
  99	if (pud_none(*pudp)) {
 100		pmdp = early_alloc_pgtable(PMD_TABLE_SIZE, nid,
 101						region_start, region_end);
 102		pud_populate(&init_mm, pudp, pmdp);
 103	}
 104	pmdp = pmd_offset(pudp, ea);
 105	if (map_page_size == PMD_SIZE) {
 106		ptep = pmdp_ptep(pmdp);
 107		goto set_the_pte;
 108	}
 109	if (!pmd_present(*pmdp)) {
 110		ptep = early_alloc_pgtable(PAGE_SIZE, nid,
 111						region_start, region_end);
 112		pmd_populate_kernel(&init_mm, pmdp, ptep);
 113	}
 114	ptep = pte_offset_kernel(pmdp, ea);
 115
 116set_the_pte:
 117	set_pte_at(&init_mm, ea, ptep, pfn_pte(pfn, flags));
 118	smp_wmb();
 119	return 0;
 120}
 121
 122/*
 123 * nid, region_start, and region_end are hints to try to place the page
 124 * table memory in the same node or region.
 125 */
 126static int __map_kernel_page(unsigned long ea, unsigned long pa,
 127			  pgprot_t flags,
 128			  unsigned int map_page_size,
 129			  int nid,
 130			  unsigned long region_start, unsigned long region_end)
 131{
 132	unsigned long pfn = pa >> PAGE_SHIFT;
 133	pgd_t *pgdp;
 134	pud_t *pudp;
 135	pmd_t *pmdp;
 136	pte_t *ptep;
 137	/*
 138	 * Make sure task size is correct as per the max adddr
 139	 */
 140	BUILD_BUG_ON(TASK_SIZE_USER64 > RADIX_PGTABLE_RANGE);
 141
 142	if (unlikely(!slab_is_available()))
 143		return early_map_kernel_page(ea, pa, flags, map_page_size,
 144						nid, region_start, region_end);
 145
 146	/*
 147	 * Should make page table allocation functions be able to take a
 148	 * node, so we can place kernel page tables on the right nodes after
 149	 * boot.
 150	 */
 151	pgdp = pgd_offset_k(ea);
 152	pudp = pud_alloc(&init_mm, pgdp, ea);
 153	if (!pudp)
 154		return -ENOMEM;
 155	if (map_page_size == PUD_SIZE) {
 156		ptep = (pte_t *)pudp;
 157		goto set_the_pte;
 158	}
 159	pmdp = pmd_alloc(&init_mm, pudp, ea);
 160	if (!pmdp)
 161		return -ENOMEM;
 162	if (map_page_size == PMD_SIZE) {
 163		ptep = pmdp_ptep(pmdp);
 164		goto set_the_pte;
 165	}
 166	ptep = pte_alloc_kernel(pmdp, ea);
 167	if (!ptep)
 168		return -ENOMEM;
 169
 170set_the_pte:
 171	set_pte_at(&init_mm, ea, ptep, pfn_pte(pfn, flags));
 172	smp_wmb();
 173	return 0;
 174}
 175
 176int radix__map_kernel_page(unsigned long ea, unsigned long pa,
 177			  pgprot_t flags,
 178			  unsigned int map_page_size)
 179{
 180	return __map_kernel_page(ea, pa, flags, map_page_size, -1, 0, 0);
 181}
 182
 183#ifdef CONFIG_STRICT_KERNEL_RWX
 184void radix__change_memory_range(unsigned long start, unsigned long end,
 185				unsigned long clear)
 186{
 187	unsigned long idx;
 188	pgd_t *pgdp;
 189	pud_t *pudp;
 190	pmd_t *pmdp;
 191	pte_t *ptep;
 192
 193	start = ALIGN_DOWN(start, PAGE_SIZE);
 194	end = PAGE_ALIGN(end); // aligns up
 195
 196	pr_debug("Changing flags on range %lx-%lx removing 0x%lx\n",
 197		 start, end, clear);
 198
 199	for (idx = start; idx < end; idx += PAGE_SIZE) {
 200		pgdp = pgd_offset_k(idx);
 201		pudp = pud_alloc(&init_mm, pgdp, idx);
 202		if (!pudp)
 203			continue;
 204		if (pud_huge(*pudp)) {
 205			ptep = (pte_t *)pudp;
 206			goto update_the_pte;
 207		}
 208		pmdp = pmd_alloc(&init_mm, pudp, idx);
 209		if (!pmdp)
 210			continue;
 211		if (pmd_huge(*pmdp)) {
 212			ptep = pmdp_ptep(pmdp);
 213			goto update_the_pte;
 214		}
 215		ptep = pte_alloc_kernel(pmdp, idx);
 216		if (!ptep)
 217			continue;
 218update_the_pte:
 219		radix__pte_update(&init_mm, idx, ptep, clear, 0, 0);
 220	}
 221
 222	radix__flush_tlb_kernel_range(start, end);
 223}
 224
 225void radix__mark_rodata_ro(void)
 226{
 227	unsigned long start, end;
 228
 229	/*
 230	 * mark_rodata_ro() will mark itself as !writable at some point.
 231	 * Due to DD1 workaround in radix__pte_update(), we'll end up with
 232	 * an invalid pte and the system will crash quite severly.
 233	 */
 234	if (cpu_has_feature(CPU_FTR_POWER9_DD1)) {
 235		pr_warn("Warning: Unable to mark rodata read only on P9 DD1\n");
 236		return;
 237	}
 238
 239	start = (unsigned long)_stext;
 240	end = (unsigned long)__init_begin;
 241
 242	radix__change_memory_range(start, end, _PAGE_WRITE);
 243}
 244
 245void radix__mark_initmem_nx(void)
 246{
 247	unsigned long start = (unsigned long)__init_begin;
 248	unsigned long end = (unsigned long)__init_end;
 249
 250	radix__change_memory_range(start, end, _PAGE_EXEC);
 251}
 252#endif /* CONFIG_STRICT_KERNEL_RWX */
 253
 254static inline void __meminit print_mapping(unsigned long start,
 255					   unsigned long end,
 256					   unsigned long size)
 257{
 258	char buf[10];
 259
 260	if (end <= start)
 261		return;
 262
 263	string_get_size(size, 1, STRING_UNITS_2, buf, sizeof(buf));
 264
 265	pr_info("Mapped 0x%016lx-0x%016lx with %s pages\n", start, end, buf);
 266}
 267
 268static int __meminit create_physical_mapping(unsigned long start,
 269					     unsigned long end,
 270					     int nid)
 271{
 272	unsigned long vaddr, addr, mapping_size = 0;
 273	pgprot_t prot;
 274	unsigned long max_mapping_size;
 275#ifdef CONFIG_STRICT_KERNEL_RWX
 276	int split_text_mapping = 1;
 277#else
 278	int split_text_mapping = 0;
 279#endif
 280
 281	start = _ALIGN_UP(start, PAGE_SIZE);
 282	for (addr = start; addr < end; addr += mapping_size) {
 283		unsigned long gap, previous_size;
 284		int rc;
 285
 286		gap = end - addr;
 287		previous_size = mapping_size;
 288		max_mapping_size = PUD_SIZE;
 289
 290retry:
 291		if (IS_ALIGNED(addr, PUD_SIZE) && gap >= PUD_SIZE &&
 292		    mmu_psize_defs[MMU_PAGE_1G].shift &&
 293		    PUD_SIZE <= max_mapping_size)
 294			mapping_size = PUD_SIZE;
 295		else if (IS_ALIGNED(addr, PMD_SIZE) && gap >= PMD_SIZE &&
 296			 mmu_psize_defs[MMU_PAGE_2M].shift)
 297			mapping_size = PMD_SIZE;
 298		else
 299			mapping_size = PAGE_SIZE;
 300
 301		if (split_text_mapping && (mapping_size == PUD_SIZE) &&
 302			(addr <= __pa_symbol(__init_begin)) &&
 303			(addr + mapping_size) >= __pa_symbol(_stext)) {
 304			max_mapping_size = PMD_SIZE;
 305			goto retry;
 306		}
 307
 308		if (split_text_mapping && (mapping_size == PMD_SIZE) &&
 309		    (addr <= __pa_symbol(__init_begin)) &&
 310		    (addr + mapping_size) >= __pa_symbol(_stext))
 311			mapping_size = PAGE_SIZE;
 312
 313		if (mapping_size != previous_size) {
 314			print_mapping(start, addr, previous_size);
 315			start = addr;
 316		}
 317
 318		vaddr = (unsigned long)__va(addr);
 319
 320		if (overlaps_kernel_text(vaddr, vaddr + mapping_size) ||
 321		    overlaps_interrupt_vector_text(vaddr, vaddr + mapping_size))
 322			prot = PAGE_KERNEL_X;
 323		else
 324			prot = PAGE_KERNEL;
 325
 326		rc = __map_kernel_page(vaddr, addr, prot, mapping_size, nid, start, end);
 327		if (rc)
 328			return rc;
 329	}
 330
 331	print_mapping(start, addr, mapping_size);
 332	return 0;
 333}
 334
 335void __init radix_init_pgtable(void)
 336{
 337	unsigned long rts_field;
 338	struct memblock_region *reg;
 339
 340	/* We don't support slb for radix */
 341	mmu_slb_size = 0;
 342	/*
 343	 * Create the linear mapping, using standard page size for now
 344	 */
 345	for_each_memblock(memory, reg) {
 346		/*
 347		 * The memblock allocator  is up at this point, so the
 348		 * page tables will be allocated within the range. No
 349		 * need or a node (which we don't have yet).
 350		 */
 351		WARN_ON(create_physical_mapping(reg->base,
 352						reg->base + reg->size,
 353						-1));
 354	}
 355
 356	/* Find out how many PID bits are supported */
 357	if (cpu_has_feature(CPU_FTR_HVMODE)) {
 358		if (!mmu_pid_bits)
 359			mmu_pid_bits = 20;
 360#ifdef CONFIG_KVM_BOOK3S_HV_POSSIBLE
 361		/*
 362		 * When KVM is possible, we only use the top half of the
 363		 * PID space to avoid collisions between host and guest PIDs
 364		 * which can cause problems due to prefetch when exiting the
 365		 * guest with AIL=3
 366		 */
 367		mmu_base_pid = 1 << (mmu_pid_bits - 1);
 368#else
 369		mmu_base_pid = 1;
 370#endif
 371	} else {
 372		/* The guest uses the bottom half of the PID space */
 373		if (!mmu_pid_bits)
 374			mmu_pid_bits = 19;
 375		mmu_base_pid = 1;
 376	}
 377
 378	/*
 379	 * Allocate Partition table and process table for the
 380	 * host.
 381	 */
 382	BUG_ON(PRTB_SIZE_SHIFT > 36);
 383	process_tb = early_alloc_pgtable(1UL << PRTB_SIZE_SHIFT, -1, 0, 0);
 384	/*
 385	 * Fill in the process table.
 386	 */
 387	rts_field = radix__get_tree_size();
 388	process_tb->prtb0 = cpu_to_be64(rts_field | __pa(init_mm.pgd) | RADIX_PGD_INDEX_SIZE);
 389	/*
 390	 * Fill in the partition table. We are suppose to use effective address
 391	 * of process table here. But our linear mapping also enable us to use
 392	 * physical address here.
 393	 */
 394	register_process_table(__pa(process_tb), 0, PRTB_SIZE_SHIFT - 12);
 395	pr_info("Process table %p and radix root for kernel: %p\n", process_tb, init_mm.pgd);
 396	asm volatile("ptesync" : : : "memory");
 397	asm volatile(PPC_TLBIE_5(%0,%1,2,1,1) : :
 398		     "r" (TLBIEL_INVAL_SET_LPID), "r" (0));
 399	asm volatile("eieio; tlbsync; ptesync" : : : "memory");
 400	trace_tlbie(0, 0, TLBIEL_INVAL_SET_LPID, 0, 2, 1, 1);
 401
 402	/*
 403	 * The init_mm context is given the first available (non-zero) PID,
 404	 * which is the "guard PID" and contains no page table. PIDR should
 405	 * never be set to zero because that duplicates the kernel address
 406	 * space at the 0x0... offset (quadrant 0)!
 407	 *
 408	 * An arbitrary PID that may later be allocated by the PID allocator
 409	 * for userspace processes must not be used either, because that
 410	 * would cause stale user mappings for that PID on CPUs outside of
 411	 * the TLB invalidation scheme (because it won't be in mm_cpumask).
 412	 *
 413	 * So permanently carve out one PID for the purpose of a guard PID.
 414	 */
 415	init_mm.context.id = mmu_base_pid;
 416	mmu_base_pid++;
 417}
 418
 419static void __init radix_init_partition_table(void)
 420{
 421	unsigned long rts_field, dw0;
 422
 423	mmu_partition_table_init();
 424	rts_field = radix__get_tree_size();
 425	dw0 = rts_field | __pa(init_mm.pgd) | RADIX_PGD_INDEX_SIZE | PATB_HR;
 426	mmu_partition_table_set_entry(0, dw0, 0);
 427
 428	pr_info("Initializing Radix MMU\n");
 429	pr_info("Partition table %p\n", partition_tb);
 430}
 431
 432void __init radix_init_native(void)
 433{
 434	register_process_table = native_register_process_table;
 435}
 436
 437static int __init get_idx_from_shift(unsigned int shift)
 438{
 439	int idx = -1;
 440
 441	switch (shift) {
 442	case 0xc:
 443		idx = MMU_PAGE_4K;
 444		break;
 445	case 0x10:
 446		idx = MMU_PAGE_64K;
 447		break;
 448	case 0x15:
 449		idx = MMU_PAGE_2M;
 450		break;
 451	case 0x1e:
 452		idx = MMU_PAGE_1G;
 453		break;
 454	}
 455	return idx;
 456}
 457
 458static int __init radix_dt_scan_page_sizes(unsigned long node,
 459					   const char *uname, int depth,
 460					   void *data)
 461{
 462	int size = 0;
 463	int shift, idx;
 464	unsigned int ap;
 465	const __be32 *prop;
 466	const char *type = of_get_flat_dt_prop(node, "device_type", NULL);
 467
 468	/* We are scanning "cpu" nodes only */
 469	if (type == NULL || strcmp(type, "cpu") != 0)
 470		return 0;
 471
 472	/* Find MMU PID size */
 473	prop = of_get_flat_dt_prop(node, "ibm,mmu-pid-bits", &size);
 474	if (prop && size == 4)
 475		mmu_pid_bits = be32_to_cpup(prop);
 476
 477	/* Grab page size encodings */
 478	prop = of_get_flat_dt_prop(node, "ibm,processor-radix-AP-encodings", &size);
 479	if (!prop)
 480		return 0;
 481
 482	pr_info("Page sizes from device-tree:\n");
 483	for (; size >= 4; size -= 4, ++prop) {
 484
 485		struct mmu_psize_def *def;
 486
 487		/* top 3 bit is AP encoding */
 488		shift = be32_to_cpu(prop[0]) & ~(0xe << 28);
 489		ap = be32_to_cpu(prop[0]) >> 29;
 490		pr_info("Page size shift = %d AP=0x%x\n", shift, ap);
 491
 492		idx = get_idx_from_shift(shift);
 493		if (idx < 0)
 494			continue;
 495
 496		def = &mmu_psize_defs[idx];
 497		def->shift = shift;
 498		def->ap  = ap;
 499	}
 500
 501	/* needed ? */
 502	cur_cpu_spec->mmu_features &= ~MMU_FTR_NO_SLBIE_B;
 503	return 1;
 504}
 505
 506void __init radix__early_init_devtree(void)
 507{
 508	int rc;
 509
 510	/*
 511	 * Try to find the available page sizes in the device-tree
 512	 */
 513	rc = of_scan_flat_dt(radix_dt_scan_page_sizes, NULL);
 514	if (rc != 0)  /* Found */
 515		goto found;
 516	/*
 517	 * let's assume we have page 4k and 64k support
 518	 */
 519	mmu_psize_defs[MMU_PAGE_4K].shift = 12;
 520	mmu_psize_defs[MMU_PAGE_4K].ap = 0x0;
 521
 522	mmu_psize_defs[MMU_PAGE_64K].shift = 16;
 523	mmu_psize_defs[MMU_PAGE_64K].ap = 0x5;
 524found:
 525#ifdef CONFIG_SPARSEMEM_VMEMMAP
 526	if (mmu_psize_defs[MMU_PAGE_2M].shift) {
 527		/*
 528		 * map vmemmap using 2M if available
 529		 */
 530		mmu_vmemmap_psize = MMU_PAGE_2M;
 531	}
 532#endif /* CONFIG_SPARSEMEM_VMEMMAP */
 533	return;
 534}
 535
 536static void update_hid_for_radix(void)
 537{
 538	unsigned long hid0;
 539	unsigned long rb = 3UL << PPC_BITLSHIFT(53); /* IS = 3 */
 540
 541	asm volatile("ptesync": : :"memory");
 542	/* prs = 0, ric = 2, rs = 0, r = 1 is = 3 */
 543	asm volatile(PPC_TLBIE_5(%0, %4, %3, %2, %1)
 544		     : : "r"(rb), "i"(1), "i"(0), "i"(2), "r"(0) : "memory");
 545	/* prs = 1, ric = 2, rs = 0, r = 1 is = 3 */
 546	asm volatile(PPC_TLBIE_5(%0, %4, %3, %2, %1)
 547		     : : "r"(rb), "i"(1), "i"(1), "i"(2), "r"(0) : "memory");
 548	asm volatile("eieio; tlbsync; ptesync; isync; slbia": : :"memory");
 549	trace_tlbie(0, 0, rb, 0, 2, 0, 1);
 550	trace_tlbie(0, 0, rb, 0, 2, 1, 1);
 551
 552	/*
 553	 * now switch the HID
 554	 */
 555	hid0  = mfspr(SPRN_HID0);
 556	hid0 |= HID0_POWER9_RADIX;
 557	mtspr(SPRN_HID0, hid0);
 558	asm volatile("isync": : :"memory");
 559
 560	/* Wait for it to happen */
 561	while (!(mfspr(SPRN_HID0) & HID0_POWER9_RADIX))
 562		cpu_relax();
 563}
 564
 565static void radix_init_amor(void)
 566{
 567	/*
 568	* In HV mode, we init AMOR (Authority Mask Override Register) so that
 569	* the hypervisor and guest can setup IAMR (Instruction Authority Mask
 570	* Register), enable key 0 and set it to 1.
 571	*
 572	* AMOR = 0b1100 .... 0000 (Mask for key 0 is 11)
 573	*/
 574	mtspr(SPRN_AMOR, (3ul << 62));
 575}
 576
 577static void radix_init_iamr(void)
 578{
 579	unsigned long iamr;
 580
 581	/*
 582	 * The IAMR should set to 0 on DD1.
 583	 */
 584	if (cpu_has_feature(CPU_FTR_POWER9_DD1))
 585		iamr = 0;
 586	else
 587		iamr = (1ul << 62);
 588
 589	/*
 590	 * Radix always uses key0 of the IAMR to determine if an access is
 591	 * allowed. We set bit 0 (IBM bit 1) of key0, to prevent instruction
 592	 * fetch.
 593	 */
 594	mtspr(SPRN_IAMR, iamr);
 595}
 596
 597void __init radix__early_init_mmu(void)
 598{
 599	unsigned long lpcr;
 600
 601#ifdef CONFIG_PPC_64K_PAGES
 602	/* PAGE_SIZE mappings */
 603	mmu_virtual_psize = MMU_PAGE_64K;
 604#else
 605	mmu_virtual_psize = MMU_PAGE_4K;
 606#endif
 607
 608#ifdef CONFIG_SPARSEMEM_VMEMMAP
 609	/* vmemmap mapping */
 610	mmu_vmemmap_psize = mmu_virtual_psize;
 611#endif
 612	/*
 613	 * initialize page table size
 614	 */
 615	__pte_index_size = RADIX_PTE_INDEX_SIZE;
 616	__pmd_index_size = RADIX_PMD_INDEX_SIZE;
 617	__pud_index_size = RADIX_PUD_INDEX_SIZE;
 618	__pgd_index_size = RADIX_PGD_INDEX_SIZE;
 619	__pud_cache_index = RADIX_PUD_INDEX_SIZE;
 620	__pmd_cache_index = RADIX_PMD_INDEX_SIZE;
 621	__pte_table_size = RADIX_PTE_TABLE_SIZE;
 622	__pmd_table_size = RADIX_PMD_TABLE_SIZE;
 623	__pud_table_size = RADIX_PUD_TABLE_SIZE;
 624	__pgd_table_size = RADIX_PGD_TABLE_SIZE;
 625
 626	__pmd_val_bits = RADIX_PMD_VAL_BITS;
 627	__pud_val_bits = RADIX_PUD_VAL_BITS;
 628	__pgd_val_bits = RADIX_PGD_VAL_BITS;
 629
 630	__kernel_virt_start = RADIX_KERN_VIRT_START;
 631	__kernel_virt_size = RADIX_KERN_VIRT_SIZE;
 632	__vmalloc_start = RADIX_VMALLOC_START;
 633	__vmalloc_end = RADIX_VMALLOC_END;
 634	__kernel_io_start = RADIX_KERN_IO_START;
 635	vmemmap = (struct page *)RADIX_VMEMMAP_BASE;
 636	ioremap_bot = IOREMAP_BASE;
 637
 638#ifdef CONFIG_PCI
 639	pci_io_base = ISA_IO_BASE;
 640#endif
 641	__pte_frag_nr = RADIX_PTE_FRAG_NR;
 642	__pte_frag_size_shift = RADIX_PTE_FRAG_SIZE_SHIFT;
 643
 644	if (!firmware_has_feature(FW_FEATURE_LPAR)) {
 645		radix_init_native();
 646		if (cpu_has_feature(CPU_FTR_POWER9_DD1))
 647			update_hid_for_radix();
 648		lpcr = mfspr(SPRN_LPCR);
 649		mtspr(SPRN_LPCR, lpcr | LPCR_UPRT | LPCR_HR);
 650		radix_init_partition_table();
 651		radix_init_amor();
 652	} else {
 653		radix_init_pseries();
 654	}
 655
 656	memblock_set_current_limit(MEMBLOCK_ALLOC_ANYWHERE);
 657
 658	radix_init_iamr();
 659	radix_init_pgtable();
 660	/* Switch to the guard PID before turning on MMU */
 661	radix__switch_mmu_context(NULL, &init_mm);
 662	if (cpu_has_feature(CPU_FTR_HVMODE))
 663		tlbiel_all();
 664}
 665
 666void radix__early_init_mmu_secondary(void)
 667{
 668	unsigned long lpcr;
 669	/*
 670	 * update partition table control register and UPRT
 671	 */
 672	if (!firmware_has_feature(FW_FEATURE_LPAR)) {
 673
 674		if (cpu_has_feature(CPU_FTR_POWER9_DD1))
 675			update_hid_for_radix();
 676
 677		lpcr = mfspr(SPRN_LPCR);
 678		mtspr(SPRN_LPCR, lpcr | LPCR_UPRT | LPCR_HR);
 679
 680		mtspr(SPRN_PTCR,
 681		      __pa(partition_tb) | (PATB_SIZE_SHIFT - 12));
 682		radix_init_amor();
 683	}
 684	radix_init_iamr();
 685
 686	radix__switch_mmu_context(NULL, &init_mm);
 687	if (cpu_has_feature(CPU_FTR_HVMODE))
 688		tlbiel_all();
 689}
 690
 691void radix__mmu_cleanup_all(void)
 692{
 693	unsigned long lpcr;
 694
 695	if (!firmware_has_feature(FW_FEATURE_LPAR)) {
 696		lpcr = mfspr(SPRN_LPCR);
 697		mtspr(SPRN_LPCR, lpcr & ~LPCR_UPRT);
 698		mtspr(SPRN_PTCR, 0);
 699		powernv_set_nmmu_ptcr(0);
 700		radix__flush_tlb_all();
 701	}
 702}
 703
 704void radix__setup_initial_memory_limit(phys_addr_t first_memblock_base,
 705				phys_addr_t first_memblock_size)
 706{
 707	/* We don't currently support the first MEMBLOCK not mapping 0
 708	 * physical on those processors
 709	 */
 710	BUG_ON(first_memblock_base != 0);
 711
 712	/*
 713	 * Radix mode is not limited by RMA / VRMA addressing.
 714	 */
 715	ppc64_rma_size = ULONG_MAX;
 716}
 717
 718#ifdef CONFIG_MEMORY_HOTPLUG
 719static void free_pte_table(pte_t *pte_start, pmd_t *pmd)
 720{
 721	pte_t *pte;
 722	int i;
 723
 724	for (i = 0; i < PTRS_PER_PTE; i++) {
 725		pte = pte_start + i;
 726		if (!pte_none(*pte))
 727			return;
 728	}
 729
 730	pte_free_kernel(&init_mm, pte_start);
 731	pmd_clear(pmd);
 732}
 733
 734static void free_pmd_table(pmd_t *pmd_start, pud_t *pud)
 735{
 736	pmd_t *pmd;
 737	int i;
 738
 739	for (i = 0; i < PTRS_PER_PMD; i++) {
 740		pmd = pmd_start + i;
 741		if (!pmd_none(*pmd))
 742			return;
 743	}
 744
 745	pmd_free(&init_mm, pmd_start);
 746	pud_clear(pud);
 747}
 748
 749struct change_mapping_params {
 750	pte_t *pte;
 751	unsigned long start;
 752	unsigned long end;
 753	unsigned long aligned_start;
 754	unsigned long aligned_end;
 755};
 756
 757static int __meminit stop_machine_change_mapping(void *data)
 758{
 759	struct change_mapping_params *params =
 760			(struct change_mapping_params *)data;
 761
 762	if (!data)
 763		return -1;
 764
 765	spin_unlock(&init_mm.page_table_lock);
 766	pte_clear(&init_mm, params->aligned_start, params->pte);
 767	create_physical_mapping(params->aligned_start, params->start, -1);
 768	create_physical_mapping(params->end, params->aligned_end, -1);
 769	spin_lock(&init_mm.page_table_lock);
 770	return 0;
 771}
 772
 773static void remove_pte_table(pte_t *pte_start, unsigned long addr,
 774			     unsigned long end)
 775{
 776	unsigned long next;
 777	pte_t *pte;
 778
 779	pte = pte_start + pte_index(addr);
 780	for (; addr < end; addr = next, pte++) {
 781		next = (addr + PAGE_SIZE) & PAGE_MASK;
 782		if (next > end)
 783			next = end;
 784
 785		if (!pte_present(*pte))
 786			continue;
 787
 788		if (!PAGE_ALIGNED(addr) || !PAGE_ALIGNED(next)) {
 789			/*
 790			 * The vmemmap_free() and remove_section_mapping()
 791			 * codepaths call us with aligned addresses.
 792			 */
 793			WARN_ONCE(1, "%s: unaligned range\n", __func__);
 794			continue;
 795		}
 796
 797		pte_clear(&init_mm, addr, pte);
 798	}
 799}
 800
 801/*
 802 * clear the pte and potentially split the mapping helper
 803 */
 804static void __meminit split_kernel_mapping(unsigned long addr, unsigned long end,
 805				unsigned long size, pte_t *pte)
 806{
 807	unsigned long mask = ~(size - 1);
 808	unsigned long aligned_start = addr & mask;
 809	unsigned long aligned_end = addr + size;
 810	struct change_mapping_params params;
 811	bool split_region = false;
 812
 813	if ((end - addr) < size) {
 814		/*
 815		 * We're going to clear the PTE, but not flushed
 816		 * the mapping, time to remap and flush. The
 817		 * effects if visible outside the processor or
 818		 * if we are running in code close to the
 819		 * mapping we cleared, we are in trouble.
 820		 */
 821		if (overlaps_kernel_text(aligned_start, addr) ||
 822			overlaps_kernel_text(end, aligned_end)) {
 823			/*
 824			 * Hack, just return, don't pte_clear
 825			 */
 826			WARN_ONCE(1, "Linear mapping %lx->%lx overlaps kernel "
 827				  "text, not splitting\n", addr, end);
 828			return;
 829		}
 830		split_region = true;
 831	}
 832
 833	if (split_region) {
 834		params.pte = pte;
 835		params.start = addr;
 836		params.end = end;
 837		params.aligned_start = addr & ~(size - 1);
 838		params.aligned_end = min_t(unsigned long, aligned_end,
 839				(unsigned long)__va(memblock_end_of_DRAM()));
 840		stop_machine(stop_machine_change_mapping, &params, NULL);
 841		return;
 842	}
 843
 844	pte_clear(&init_mm, addr, pte);
 845}
 846
 847static void remove_pmd_table(pmd_t *pmd_start, unsigned long addr,
 848			     unsigned long end)
 849{
 850	unsigned long next;
 851	pte_t *pte_base;
 852	pmd_t *pmd;
 853
 854	pmd = pmd_start + pmd_index(addr);
 855	for (; addr < end; addr = next, pmd++) {
 856		next = pmd_addr_end(addr, end);
 857
 858		if (!pmd_present(*pmd))
 859			continue;
 860
 861		if (pmd_huge(*pmd)) {
 862			split_kernel_mapping(addr, end, PMD_SIZE, (pte_t *)pmd);
 863			continue;
 864		}
 865
 866		pte_base = (pte_t *)pmd_page_vaddr(*pmd);
 867		remove_pte_table(pte_base, addr, next);
 868		free_pte_table(pte_base, pmd);
 869	}
 870}
 871
 872static void remove_pud_table(pud_t *pud_start, unsigned long addr,
 873			     unsigned long end)
 874{
 875	unsigned long next;
 876	pmd_t *pmd_base;
 877	pud_t *pud;
 878
 879	pud = pud_start + pud_index(addr);
 880	for (; addr < end; addr = next, pud++) {
 881		next = pud_addr_end(addr, end);
 882
 883		if (!pud_present(*pud))
 884			continue;
 885
 886		if (pud_huge(*pud)) {
 887			split_kernel_mapping(addr, end, PUD_SIZE, (pte_t *)pud);
 888			continue;
 889		}
 890
 891		pmd_base = (pmd_t *)pud_page_vaddr(*pud);
 892		remove_pmd_table(pmd_base, addr, next);
 893		free_pmd_table(pmd_base, pud);
 894	}
 895}
 896
 897static void __meminit remove_pagetable(unsigned long start, unsigned long end)
 898{
 899	unsigned long addr, next;
 900	pud_t *pud_base;
 901	pgd_t *pgd;
 902
 903	spin_lock(&init_mm.page_table_lock);
 904
 905	for (addr = start; addr < end; addr = next) {
 906		next = pgd_addr_end(addr, end);
 907
 908		pgd = pgd_offset_k(addr);
 909		if (!pgd_present(*pgd))
 910			continue;
 911
 912		if (pgd_huge(*pgd)) {
 913			split_kernel_mapping(addr, end, PGDIR_SIZE, (pte_t *)pgd);
 914			continue;
 915		}
 916
 917		pud_base = (pud_t *)pgd_page_vaddr(*pgd);
 918		remove_pud_table(pud_base, addr, next);
 919	}
 920
 921	spin_unlock(&init_mm.page_table_lock);
 922	radix__flush_tlb_kernel_range(start, end);
 923}
 924
 925int __meminit radix__create_section_mapping(unsigned long start, unsigned long end, int nid)
 926{
 927	return create_physical_mapping(start, end, nid);
 928}
 929
 930int __meminit radix__remove_section_mapping(unsigned long start, unsigned long end)
 931{
 932	remove_pagetable(start, end);
 933	return 0;
 934}
 935#endif /* CONFIG_MEMORY_HOTPLUG */
 936
 937#ifdef CONFIG_SPARSEMEM_VMEMMAP
 938static int __map_kernel_page_nid(unsigned long ea, unsigned long pa,
 939				 pgprot_t flags, unsigned int map_page_size,
 940				 int nid)
 941{
 942	return __map_kernel_page(ea, pa, flags, map_page_size, nid, 0, 0);
 943}
 944
 945int __meminit radix__vmemmap_create_mapping(unsigned long start,
 946				      unsigned long page_size,
 947				      unsigned long phys)
 948{
 949	/* Create a PTE encoding */
 950	unsigned long flags = _PAGE_PRESENT | _PAGE_ACCESSED | _PAGE_KERNEL_RW;
 951	int nid = early_pfn_to_nid(phys >> PAGE_SHIFT);
 952	int ret;
 953
 954	ret = __map_kernel_page_nid(start, phys, __pgprot(flags), page_size, nid);
 955	BUG_ON(ret);
 956
 957	return 0;
 958}
 959
 960#ifdef CONFIG_MEMORY_HOTPLUG
 961void __meminit radix__vmemmap_remove_mapping(unsigned long start, unsigned long page_size)
 962{
 963	remove_pagetable(start, start + page_size);
 964}
 965#endif
 966#endif
 967
 968#ifdef CONFIG_TRANSPARENT_HUGEPAGE
 969
 970unsigned long radix__pmd_hugepage_update(struct mm_struct *mm, unsigned long addr,
 971				  pmd_t *pmdp, unsigned long clr,
 972				  unsigned long set)
 973{
 974	unsigned long old;
 975
 976#ifdef CONFIG_DEBUG_VM
 977	WARN_ON(!radix__pmd_trans_huge(*pmdp) && !pmd_devmap(*pmdp));
 978	assert_spin_locked(&mm->page_table_lock);
 979#endif
 980
 981	old = radix__pte_update(mm, addr, (pte_t *)pmdp, clr, set, 1);
 982	trace_hugepage_update(addr, old, clr, set);
 983
 984	return old;
 985}
 986
 987pmd_t radix__pmdp_collapse_flush(struct vm_area_struct *vma, unsigned long address,
 988			pmd_t *pmdp)
 989
 990{
 991	pmd_t pmd;
 992
 993	VM_BUG_ON(address & ~HPAGE_PMD_MASK);
 994	VM_BUG_ON(radix__pmd_trans_huge(*pmdp));
 995	VM_BUG_ON(pmd_devmap(*pmdp));
 996	/*
 997	 * khugepaged calls this for normal pmd
 998	 */
 999	pmd = *pmdp;
1000	pmd_clear(pmdp);
1001
1002	/*FIXME!!  Verify whether we need this kick below */
1003	serialize_against_pte_lookup(vma->vm_mm);
1004
1005	radix__flush_tlb_collapsed_pmd(vma->vm_mm, address);
1006
1007	return pmd;
1008}
1009
1010/*
1011 * For us pgtable_t is pte_t *. Inorder to save the deposisted
1012 * page table, we consider the allocated page table as a list
1013 * head. On withdraw we need to make sure we zero out the used
1014 * list_head memory area.
1015 */
1016void radix__pgtable_trans_huge_deposit(struct mm_struct *mm, pmd_t *pmdp,
1017				 pgtable_t pgtable)
1018{
1019        struct list_head *lh = (struct list_head *) pgtable;
1020
1021        assert_spin_locked(pmd_lockptr(mm, pmdp));
1022
1023        /* FIFO */
1024        if (!pmd_huge_pte(mm, pmdp))
1025                INIT_LIST_HEAD(lh);
1026        else
1027                list_add(lh, (struct list_head *) pmd_huge_pte(mm, pmdp));
1028        pmd_huge_pte(mm, pmdp) = pgtable;
1029}
1030
1031pgtable_t radix__pgtable_trans_huge_withdraw(struct mm_struct *mm, pmd_t *pmdp)
1032{
1033        pte_t *ptep;
1034        pgtable_t pgtable;
1035        struct list_head *lh;
1036
1037        assert_spin_locked(pmd_lockptr(mm, pmdp));
1038
1039        /* FIFO */
1040        pgtable = pmd_huge_pte(mm, pmdp);
1041        lh = (struct list_head *) pgtable;
1042        if (list_empty(lh))
1043                pmd_huge_pte(mm, pmdp) = NULL;
1044        else {
1045                pmd_huge_pte(mm, pmdp) = (pgtable_t) lh->next;
1046                list_del(lh);
1047        }
1048        ptep = (pte_t *) pgtable;
1049        *ptep = __pte(0);
1050        ptep++;
1051        *ptep = __pte(0);
1052        return pgtable;
1053}
1054
1055
1056pmd_t radix__pmdp_huge_get_and_clear(struct mm_struct *mm,
1057			       unsigned long addr, pmd_t *pmdp)
1058{
1059	pmd_t old_pmd;
1060	unsigned long old;
1061
1062	old = radix__pmd_hugepage_update(mm, addr, pmdp, ~0UL, 0);
1063	old_pmd = __pmd(old);
1064	/*
1065	 * Serialize against find_current_mm_pte which does lock-less
1066	 * lookup in page tables with local interrupts disabled. For huge pages
1067	 * it casts pmd_t to pte_t. Since format of pte_t is different from
1068	 * pmd_t we want to prevent transit from pmd pointing to page table
1069	 * to pmd pointing to huge page (and back) while interrupts are disabled.
1070	 * We clear pmd to possibly replace it with page table pointer in
1071	 * different code paths. So make sure we wait for the parallel
1072	 * find_current_mm_pte to finish.
1073	 */
1074	serialize_against_pte_lookup(mm);
1075	return old_pmd;
1076}
1077
1078int radix__has_transparent_hugepage(void)
1079{
1080	/* For radix 2M at PMD level means thp */
1081	if (mmu_psize_defs[MMU_PAGE_2M].shift == PMD_SHIFT)
1082		return 1;
1083	return 0;
1084}
1085#endif /* CONFIG_TRANSPARENT_HUGEPAGE */