1// SPDX-License-Identifier: GPL-2.0-only
   2/*
   3 * Copyright 2002 Andi Kleen, SuSE Labs.
   4 * Thanks to Ben LaHaise for precious feedback.
   5 */
   6#include <linux/highmem.h>
   7#include <linux/memblock.h>
   8#include <linux/sched.h>
   9#include <linux/mm.h>
  10#include <linux/interrupt.h>
  11#include <linux/seq_file.h>
 
  12#include <linux/debugfs.h>
  13#include <linux/pfn.h>
  14#include <linux/percpu.h>
  15#include <linux/gfp.h>
  16#include <linux/pci.h>
  17#include <linux/vmalloc.h>
  18#include <linux/libnvdimm.h>
  19#include <linux/vmstat.h>
  20#include <linux/kernel.h>
  21#include <linux/cc_platform.h>
  22#include <linux/set_memory.h>
  23#include <linux/memregion.h>
  24
  25#include <asm/e820/api.h>
  26#include <asm/processor.h>
  27#include <asm/tlbflush.h>
  28#include <asm/sections.h>
  29#include <asm/setup.h>
  30#include <linux/uaccess.h>
  31#include <asm/pgalloc.h>
  32#include <asm/proto.h>
  33#include <asm/memtype.h>
  34#include <asm/hyperv-tlfs.h>
  35#include <asm/mshyperv.h>
  36
  37#include "../mm_internal.h"
  38
  39/*
  40 * The current flushing context - we pass it instead of 5 arguments:
  41 */
  42struct cpa_data {
  43	unsigned long	*vaddr;
  44	pgd_t		*pgd;
  45	pgprot_t	mask_set;
  46	pgprot_t	mask_clr;
  47	unsigned long	numpages;
  48	unsigned long	curpage;
  49	unsigned long	pfn;
  50	unsigned int	flags;
  51	unsigned int	force_split		: 1,
  52			force_static_prot	: 1,
  53			force_flush_all		: 1;
  54	struct page	**pages;
  55};
  56
  57enum cpa_warn {
  58	CPA_CONFLICT,
  59	CPA_PROTECT,
  60	CPA_DETECT,
  61};
  62
  63static const int cpa_warn_level = CPA_PROTECT;
  64
  65/*
  66 * Serialize cpa() (for !DEBUG_PAGEALLOC which uses large identity mappings)
  67 * using cpa_lock. So that we don't allow any other cpu, with stale large tlb
  68 * entries change the page attribute in parallel to some other cpu
  69 * splitting a large page entry along with changing the attribute.
  70 */
  71static DEFINE_SPINLOCK(cpa_lock);
  72
  73#define CPA_FLUSHTLB 1
  74#define CPA_ARRAY 2
  75#define CPA_PAGES_ARRAY 4
  76#define CPA_NO_CHECK_ALIAS 8 /* Do not search for aliases */
  77
  78static inline pgprot_t cachemode2pgprot(enum page_cache_mode pcm)
  79{
  80	return __pgprot(cachemode2protval(pcm));
  81}
  82
  83#ifdef CONFIG_PROC_FS
  84static unsigned long direct_pages_count[PG_LEVEL_NUM];
  85
  86void update_page_count(int level, unsigned long pages)
  87{
  88	/* Protect against CPA */
  89	spin_lock(&pgd_lock);
  90	direct_pages_count[level] += pages;
  91	spin_unlock(&pgd_lock);
  92}
  93
  94static void split_page_count(int level)
  95{
  96	if (direct_pages_count[level] == 0)
  97		return;
  98
  99	direct_pages_count[level]--;
 100	if (system_state == SYSTEM_RUNNING) {
 101		if (level == PG_LEVEL_2M)
 102			count_vm_event(DIRECT_MAP_LEVEL2_SPLIT);
 103		else if (level == PG_LEVEL_1G)
 104			count_vm_event(DIRECT_MAP_LEVEL3_SPLIT);
 105	}
 106	direct_pages_count[level - 1] += PTRS_PER_PTE;
 107}
 108
 109void arch_report_meminfo(struct seq_file *m)
 110{
 111	seq_printf(m, "DirectMap4k:    %8lu kB\n",
 112			direct_pages_count[PG_LEVEL_4K] << 2);
 113#if defined(CONFIG_X86_64) || defined(CONFIG_X86_PAE)
 114	seq_printf(m, "DirectMap2M:    %8lu kB\n",
 115			direct_pages_count[PG_LEVEL_2M] << 11);
 116#else
 117	seq_printf(m, "DirectMap4M:    %8lu kB\n",
 118			direct_pages_count[PG_LEVEL_2M] << 12);
 119#endif
 120	if (direct_gbpages)
 121		seq_printf(m, "DirectMap1G:    %8lu kB\n",
 122			direct_pages_count[PG_LEVEL_1G] << 20);
 123}
 124#else
 125static inline void split_page_count(int level) { }
 126#endif
 127
 128#ifdef CONFIG_X86_CPA_STATISTICS
 129
 130static unsigned long cpa_1g_checked;
 131static unsigned long cpa_1g_sameprot;
 132static unsigned long cpa_1g_preserved;
 133static unsigned long cpa_2m_checked;
 134static unsigned long cpa_2m_sameprot;
 135static unsigned long cpa_2m_preserved;
 136static unsigned long cpa_4k_install;
 137
 138static inline void cpa_inc_1g_checked(void)
 139{
 140	cpa_1g_checked++;
 141}
 142
 143static inline void cpa_inc_2m_checked(void)
 144{
 145	cpa_2m_checked++;
 146}
 147
 148static inline void cpa_inc_4k_install(void)
 149{
 150	data_race(cpa_4k_install++);
 151}
 152
 153static inline void cpa_inc_lp_sameprot(int level)
 154{
 155	if (level == PG_LEVEL_1G)
 156		cpa_1g_sameprot++;
 157	else
 158		cpa_2m_sameprot++;
 159}
 160
 161static inline void cpa_inc_lp_preserved(int level)
 162{
 163	if (level == PG_LEVEL_1G)
 164		cpa_1g_preserved++;
 165	else
 166		cpa_2m_preserved++;
 167}
 168
 169static int cpastats_show(struct seq_file *m, void *p)
 170{
 171	seq_printf(m, "1G pages checked:     %16lu\n", cpa_1g_checked);
 172	seq_printf(m, "1G pages sameprot:    %16lu\n", cpa_1g_sameprot);
 173	seq_printf(m, "1G pages preserved:   %16lu\n", cpa_1g_preserved);
 174	seq_printf(m, "2M pages checked:     %16lu\n", cpa_2m_checked);
 175	seq_printf(m, "2M pages sameprot:    %16lu\n", cpa_2m_sameprot);
 176	seq_printf(m, "2M pages preserved:   %16lu\n", cpa_2m_preserved);
 177	seq_printf(m, "4K pages set-checked: %16lu\n", cpa_4k_install);
 178	return 0;
 179}
 180
 181static int cpastats_open(struct inode *inode, struct file *file)
 182{
 183	return single_open(file, cpastats_show, NULL);
 184}
 185
 186static const struct file_operations cpastats_fops = {
 187	.open		= cpastats_open,
 188	.read		= seq_read,
 189	.llseek		= seq_lseek,
 190	.release	= single_release,
 191};
 192
 193static int __init cpa_stats_init(void)
 194{
 195	debugfs_create_file("cpa_stats", S_IRUSR, arch_debugfs_dir, NULL,
 196			    &cpastats_fops);
 197	return 0;
 198}
 199late_initcall(cpa_stats_init);
 200#else
 201static inline void cpa_inc_1g_checked(void) { }
 202static inline void cpa_inc_2m_checked(void) { }
 203static inline void cpa_inc_4k_install(void) { }
 204static inline void cpa_inc_lp_sameprot(int level) { }
 205static inline void cpa_inc_lp_preserved(int level) { }
 206#endif
 207
 208
 209static inline int
 210within(unsigned long addr, unsigned long start, unsigned long end)
 211{
 212	return addr >= start && addr < end;
 213}
 214
 215static inline int
 216within_inclusive(unsigned long addr, unsigned long start, unsigned long end)
 217{
 218	return addr >= start && addr <= end;
 219}
 220
 221#ifdef CONFIG_X86_64
 222
 223/*
 224 * The kernel image is mapped into two places in the virtual address space
 225 * (addresses without KASLR, of course):
 226 *
 227 * 1. The kernel direct map (0xffff880000000000)
 228 * 2. The "high kernel map" (0xffffffff81000000)
 229 *
 230 * We actually execute out of #2. If we get the address of a kernel symbol, it
 231 * points to #2, but almost all physical-to-virtual translations point to #1.
 232 *
 233 * This is so that we can have both a directmap of all physical memory *and*
 234 * take full advantage of the the limited (s32) immediate addressing range (2G)
 235 * of x86_64.
 236 *
 237 * See Documentation/x86/x86_64/mm.rst for more detail.
 238 */
 239
 240static inline unsigned long highmap_start_pfn(void)
 241{
 242	return __pa_symbol(_text) >> PAGE_SHIFT;
 243}
 244
 245static inline unsigned long highmap_end_pfn(void)
 246{
 247	/* Do not reference physical address outside the kernel. */
 248	return __pa_symbol(roundup(_brk_end, PMD_SIZE) - 1) >> PAGE_SHIFT;
 249}
 250
 251static bool __cpa_pfn_in_highmap(unsigned long pfn)
 252{
 253	/*
 254	 * Kernel text has an alias mapping at a high address, known
 255	 * here as "highmap".
 256	 */
 257	return within_inclusive(pfn, highmap_start_pfn(), highmap_end_pfn());
 258}
 259
 260#else
 261
 262static bool __cpa_pfn_in_highmap(unsigned long pfn)
 263{
 264	/* There is no highmap on 32-bit */
 265	return false;
 266}
 267
 268#endif
 269
 270/*
 271 * See set_mce_nospec().
 272 *
 273 * Machine check recovery code needs to change cache mode of poisoned pages to
 274 * UC to avoid speculative access logging another error. But passing the
 275 * address of the 1:1 mapping to set_memory_uc() is a fine way to encourage a
 276 * speculative access. So we cheat and flip the top bit of the address. This
 277 * works fine for the code that updates the page tables. But at the end of the
 278 * process we need to flush the TLB and cache and the non-canonical address
 279 * causes a #GP fault when used by the INVLPG and CLFLUSH instructions.
 280 *
 281 * But in the common case we already have a canonical address. This code
 282 * will fix the top bit if needed and is a no-op otherwise.
 283 */
 284static inline unsigned long fix_addr(unsigned long addr)
 285{
 286#ifdef CONFIG_X86_64
 287	return (long)(addr << 1) >> 1;
 288#else
 289	return addr;
 290#endif
 291}
 292
 293static unsigned long __cpa_addr(struct cpa_data *cpa, unsigned long idx)
 294{
 295	if (cpa->flags & CPA_PAGES_ARRAY) {
 296		struct page *page = cpa->pages[idx];
 297
 298		if (unlikely(PageHighMem(page)))
 299			return 0;
 300
 301		return (unsigned long)page_address(page);
 302	}
 303
 304	if (cpa->flags & CPA_ARRAY)
 305		return cpa->vaddr[idx];
 306
 307	return *cpa->vaddr + idx * PAGE_SIZE;
 308}
 309
 310/*
 311 * Flushing functions
 312 */
 313
 314static void clflush_cache_range_opt(void *vaddr, unsigned int size)
 315{
 316	const unsigned long clflush_size = boot_cpu_data.x86_clflush_size;
 317	void *p = (void *)((unsigned long)vaddr & ~(clflush_size - 1));
 318	void *vend = vaddr + size;
 319
 320	if (p >= vend)
 321		return;
 322
 323	for (; p < vend; p += clflush_size)
 324		clflushopt(p);
 325}
 326
 327/**
 328 * clflush_cache_range - flush a cache range with clflush
 329 * @vaddr:	virtual start address
 330 * @size:	number of bytes to flush
 331 *
 332 * CLFLUSHOPT is an unordered instruction which needs fencing with MFENCE or
 333 * SFENCE to avoid ordering issues.
 334 */
 335void clflush_cache_range(void *vaddr, unsigned int size)
 336{
 337	mb();
 338	clflush_cache_range_opt(vaddr, size);
 339	mb();
 340}
 341EXPORT_SYMBOL_GPL(clflush_cache_range);
 342
 343#ifdef CONFIG_ARCH_HAS_PMEM_API
 344void arch_invalidate_pmem(void *addr, size_t size)
 345{
 346	clflush_cache_range(addr, size);
 347}
 348EXPORT_SYMBOL_GPL(arch_invalidate_pmem);
 349#endif
 350
 351#ifdef CONFIG_ARCH_HAS_CPU_CACHE_INVALIDATE_MEMREGION
 352bool cpu_cache_has_invalidate_memregion(void)
 353{
 354	return !cpu_feature_enabled(X86_FEATURE_HYPERVISOR);
 355}
 356EXPORT_SYMBOL_NS_GPL(cpu_cache_has_invalidate_memregion, DEVMEM);
 357
 358int cpu_cache_invalidate_memregion(int res_desc)
 359{
 360	if (WARN_ON_ONCE(!cpu_cache_has_invalidate_memregion()))
 361		return -ENXIO;
 362	wbinvd_on_all_cpus();
 363	return 0;
 364}
 365EXPORT_SYMBOL_NS_GPL(cpu_cache_invalidate_memregion, DEVMEM);
 366#endif
 367
 368static void __cpa_flush_all(void *arg)
 369{
 370	unsigned long cache = (unsigned long)arg;
 371
 372	/*
 373	 * Flush all to work around Errata in early athlons regarding
 374	 * large page flushing.
 375	 */
 376	__flush_tlb_all();
 377
 378	if (cache && boot_cpu_data.x86 >= 4)
 379		wbinvd();
 380}
 381
 382static void cpa_flush_all(unsigned long cache)
 383{
 384	BUG_ON(irqs_disabled() && !early_boot_irqs_disabled);
 385
 386	on_each_cpu(__cpa_flush_all, (void *) cache, 1);
 387}
 388
 389static void __cpa_flush_tlb(void *data)
 390{
 391	struct cpa_data *cpa = data;
 392	unsigned int i;
 393
 394	for (i = 0; i < cpa->numpages; i++)
 395		flush_tlb_one_kernel(fix_addr(__cpa_addr(cpa, i)));
 396}
 397
 398static void cpa_flush(struct cpa_data *data, int cache)
 399{
 400	struct cpa_data *cpa = data;
 401	unsigned int i;
 402
 403	BUG_ON(irqs_disabled() && !early_boot_irqs_disabled);
 404
 405	if (cache && !static_cpu_has(X86_FEATURE_CLFLUSH)) {
 406		cpa_flush_all(cache);
 407		return;
 408	}
 409
 410	if (cpa->force_flush_all || cpa->numpages > tlb_single_page_flush_ceiling)
 411		flush_tlb_all();
 412	else
 413		on_each_cpu(__cpa_flush_tlb, cpa, 1);
 414
 415	if (!cache)
 416		return;
 417
 418	mb();
 419	for (i = 0; i < cpa->numpages; i++) {
 420		unsigned long addr = __cpa_addr(cpa, i);
 421		unsigned int level;
 422
 423		pte_t *pte = lookup_address(addr, &level);
 424
 425		/*
 426		 * Only flush present addresses:
 427		 */
 428		if (pte && (pte_val(*pte) & _PAGE_PRESENT))
 429			clflush_cache_range_opt((void *)fix_addr(addr), PAGE_SIZE);
 430	}
 431	mb();
 432}
 433
 434static bool overlaps(unsigned long r1_start, unsigned long r1_end,
 435		     unsigned long r2_start, unsigned long r2_end)
 436{
 437	return (r1_start <= r2_end && r1_end >= r2_start) ||
 438		(r2_start <= r1_end && r2_end >= r1_start);
 439}
 440
 441#ifdef CONFIG_PCI_BIOS
 442/*
 443 * The BIOS area between 640k and 1Mb needs to be executable for PCI BIOS
 444 * based config access (CONFIG_PCI_GOBIOS) support.
 445 */
 446#define BIOS_PFN	PFN_DOWN(BIOS_BEGIN)
 447#define BIOS_PFN_END	PFN_DOWN(BIOS_END - 1)
 448
 449static pgprotval_t protect_pci_bios(unsigned long spfn, unsigned long epfn)
 450{
 451	if (pcibios_enabled && overlaps(spfn, epfn, BIOS_PFN, BIOS_PFN_END))
 452		return _PAGE_NX;
 453	return 0;
 454}
 455#else
 456static pgprotval_t protect_pci_bios(unsigned long spfn, unsigned long epfn)
 457{
 458	return 0;
 459}
 460#endif
 461
 462/*
 463 * The .rodata section needs to be read-only. Using the pfn catches all
 464 * aliases.  This also includes __ro_after_init, so do not enforce until
 465 * kernel_set_to_readonly is true.
 466 */
 467static pgprotval_t protect_rodata(unsigned long spfn, unsigned long epfn)
 468{
 469	unsigned long epfn_ro, spfn_ro = PFN_DOWN(__pa_symbol(__start_rodata));
 470
 471	/*
 472	 * Note: __end_rodata is at page aligned and not inclusive, so
 473	 * subtract 1 to get the last enforced PFN in the rodata area.
 474	 */
 475	epfn_ro = PFN_DOWN(__pa_symbol(__end_rodata)) - 1;
 476
 477	if (kernel_set_to_readonly && overlaps(spfn, epfn, spfn_ro, epfn_ro))
 478		return _PAGE_RW;
 479	return 0;
 480}
 481
 482/*
 483 * Protect kernel text against becoming non executable by forbidding
 484 * _PAGE_NX.  This protects only the high kernel mapping (_text -> _etext)
 485 * out of which the kernel actually executes.  Do not protect the low
 486 * mapping.
 487 *
 488 * This does not cover __inittext since that is gone after boot.
 489 */
 490static pgprotval_t protect_kernel_text(unsigned long start, unsigned long end)
 491{
 492	unsigned long t_end = (unsigned long)_etext - 1;
 493	unsigned long t_start = (unsigned long)_text;
 494
 495	if (overlaps(start, end, t_start, t_end))
 496		return _PAGE_NX;
 497	return 0;
 498}
 499
 500#if defined(CONFIG_X86_64)
 501/*
 502 * Once the kernel maps the text as RO (kernel_set_to_readonly is set),
 503 * kernel text mappings for the large page aligned text, rodata sections
 504 * will be always read-only. For the kernel identity mappings covering the
 505 * holes caused by this alignment can be anything that user asks.
 506 *
 507 * This will preserve the large page mappings for kernel text/data at no
 508 * extra cost.
 509 */
 510static pgprotval_t protect_kernel_text_ro(unsigned long start,
 511					  unsigned long end)
 512{
 513	unsigned long t_end = (unsigned long)__end_rodata_hpage_align - 1;
 514	unsigned long t_start = (unsigned long)_text;
 515	unsigned int level;
 516
 517	if (!kernel_set_to_readonly || !overlaps(start, end, t_start, t_end))
 518		return 0;
 519	/*
 520	 * Don't enforce the !RW mapping for the kernel text mapping, if
 521	 * the current mapping is already using small page mapping.  No
 522	 * need to work hard to preserve large page mappings in this case.
 523	 *
 524	 * This also fixes the Linux Xen paravirt guest boot failure caused
 525	 * by unexpected read-only mappings for kernel identity
 526	 * mappings. In this paravirt guest case, the kernel text mapping
 527	 * and the kernel identity mapping share the same page-table pages,
 528	 * so the protections for kernel text and identity mappings have to
 529	 * be the same.
 530	 */
 531	if (lookup_address(start, &level) && (level != PG_LEVEL_4K))
 532		return _PAGE_RW;
 533	return 0;
 534}
 535#else
 536static pgprotval_t protect_kernel_text_ro(unsigned long start,
 537					  unsigned long end)
 538{
 539	return 0;
 540}
 541#endif
 542
 543static inline bool conflicts(pgprot_t prot, pgprotval_t val)
 544{
 545	return (pgprot_val(prot) & ~val) != pgprot_val(prot);
 546}
 547
 548static inline void check_conflict(int warnlvl, pgprot_t prot, pgprotval_t val,
 549				  unsigned long start, unsigned long end,
 550				  unsigned long pfn, const char *txt)
 551{
 552	static const char *lvltxt[] = {
 553		[CPA_CONFLICT]	= "conflict",
 554		[CPA_PROTECT]	= "protect",
 555		[CPA_DETECT]	= "detect",
 556	};
 557
 558	if (warnlvl > cpa_warn_level || !conflicts(prot, val))
 559		return;
 560
 561	pr_warn("CPA %8s %10s: 0x%016lx - 0x%016lx PFN %lx req %016llx prevent %016llx\n",
 562		lvltxt[warnlvl], txt, start, end, pfn, (unsigned long long)pgprot_val(prot),
 563		(unsigned long long)val);
 564}
 565
 566/*
 567 * Certain areas of memory on x86 require very specific protection flags,
 568 * for example the BIOS area or kernel text. Callers don't always get this
 569 * right (again, ioremap() on BIOS memory is not uncommon) so this function
 570 * checks and fixes these known static required protection bits.
 571 */
 572static inline pgprot_t static_protections(pgprot_t prot, unsigned long start,
 573					  unsigned long pfn, unsigned long npg,
 574					  unsigned long lpsize, int warnlvl)
 575{
 576	pgprotval_t forbidden, res;
 577	unsigned long end;
 578
 579	/*
 580	 * There is no point in checking RW/NX conflicts when the requested
 581	 * mapping is setting the page !PRESENT.
 582	 */
 583	if (!(pgprot_val(prot) & _PAGE_PRESENT))
 584		return prot;
 585
 586	/* Operate on the virtual address */
 587	end = start + npg * PAGE_SIZE - 1;
 588
 589	res = protect_kernel_text(start, end);
 590	check_conflict(warnlvl, prot, res, start, end, pfn, "Text NX");
 591	forbidden = res;
 592
 593	/*
 594	 * Special case to preserve a large page. If the change spawns the
 595	 * full large page mapping then there is no point to split it
 596	 * up. Happens with ftrace and is going to be removed once ftrace
 597	 * switched to text_poke().
 598	 */
 599	if (lpsize != (npg * PAGE_SIZE) || (start & (lpsize - 1))) {
 600		res = protect_kernel_text_ro(start, end);
 601		check_conflict(warnlvl, prot, res, start, end, pfn, "Text RO");
 602		forbidden |= res;
 603	}
 604
 605	/* Check the PFN directly */
 606	res = protect_pci_bios(pfn, pfn + npg - 1);
 607	check_conflict(warnlvl, prot, res, start, end, pfn, "PCIBIOS NX");
 608	forbidden |= res;
 609
 610	res = protect_rodata(pfn, pfn + npg - 1);
 611	check_conflict(warnlvl, prot, res, start, end, pfn, "Rodata RO");
 612	forbidden |= res;
 613
 614	return __pgprot(pgprot_val(prot) & ~forbidden);
 615}
 616
 617/*
 618 * Validate strict W^X semantics.
 619 */
 620static inline pgprot_t verify_rwx(pgprot_t old, pgprot_t new, unsigned long start,
 621				  unsigned long pfn, unsigned long npg)
 
 622{
 623	unsigned long end;
 624
 625	/*
 626	 * 32-bit has some unfixable W+X issues, like EFI code
 627	 * and writeable data being in the same page.  Disable
 628	 * detection and enforcement there.
 629	 */
 630	if (IS_ENABLED(CONFIG_X86_32))
 631		return new;
 632
 633	/* Only verify when NX is supported: */
 634	if (!(__supported_pte_mask & _PAGE_NX))
 635		return new;
 636
 637	if (!((pgprot_val(old) ^ pgprot_val(new)) & (_PAGE_RW | _PAGE_NX)))
 638		return new;
 639
 640	if ((pgprot_val(new) & (_PAGE_RW | _PAGE_NX)) != _PAGE_RW)
 641		return new;
 642
 
 
 
 
 643	end = start + npg * PAGE_SIZE - 1;
 644	WARN_ONCE(1, "CPA detected W^X violation: %016llx -> %016llx range: 0x%016lx - 0x%016lx PFN %lx\n",
 645		  (unsigned long long)pgprot_val(old),
 646		  (unsigned long long)pgprot_val(new),
 647		  start, end, pfn);
 648
 649	/*
 650	 * For now, allow all permission change attempts by returning the
 651	 * attempted permissions.  This can 'return old' to actively
 652	 * refuse the permission change at a later time.
 653	 */
 654	return new;
 655}
 656
 657/*
 658 * Lookup the page table entry for a virtual address in a specific pgd.
 659 * Return a pointer to the entry and the level of the mapping.
 
 660 */
 661pte_t *lookup_address_in_pgd(pgd_t *pgd, unsigned long address,
 662			     unsigned int *level)
 663{
 664	p4d_t *p4d;
 665	pud_t *pud;
 666	pmd_t *pmd;
 667
 668	*level = PG_LEVEL_NONE;
 
 
 669
 670	if (pgd_none(*pgd))
 671		return NULL;
 672
 
 
 
 673	p4d = p4d_offset(pgd, address);
 674	if (p4d_none(*p4d))
 675		return NULL;
 676
 677	*level = PG_LEVEL_512G;
 678	if (p4d_large(*p4d) || !p4d_present(*p4d))
 679		return (pte_t *)p4d;
 680
 
 
 
 681	pud = pud_offset(p4d, address);
 682	if (pud_none(*pud))
 683		return NULL;
 684
 685	*level = PG_LEVEL_1G;
 686	if (pud_large(*pud) || !pud_present(*pud))
 687		return (pte_t *)pud;
 688
 
 
 
 689	pmd = pmd_offset(pud, address);
 690	if (pmd_none(*pmd))
 691		return NULL;
 692
 693	*level = PG_LEVEL_2M;
 694	if (pmd_large(*pmd) || !pmd_present(*pmd))
 695		return (pte_t *)pmd;
 696
 
 
 
 697	*level = PG_LEVEL_4K;
 698
 699	return pte_offset_kernel(pmd, address);
 700}
 701
 702/*
 
 
 
 
 
 
 
 
 
 
 
 
 703 * Lookup the page table entry for a virtual address. Return a pointer
 704 * to the entry and the level of the mapping.
 705 *
 706 * Note: We return pud and pmd either when the entry is marked large
 707 * or when the present bit is not set. Otherwise we would return a
 708 * pointer to a nonexisting mapping.
 709 */
 710pte_t *lookup_address(unsigned long address, unsigned int *level)
 711{
 712	return lookup_address_in_pgd(pgd_offset_k(address), address, level);
 713}
 714EXPORT_SYMBOL_GPL(lookup_address);
 715
 716static pte_t *_lookup_address_cpa(struct cpa_data *cpa, unsigned long address,
 717				  unsigned int *level)
 718{
 719	if (cpa->pgd)
 720		return lookup_address_in_pgd(cpa->pgd + pgd_index(address),
 721					       address, level);
 722
 723	return lookup_address(address, level);
 
 
 
 
 
 724}
 725
 726/*
 727 * Lookup the PMD entry for a virtual address. Return a pointer to the entry
 728 * or NULL if not present.
 729 */
 730pmd_t *lookup_pmd_address(unsigned long address)
 731{
 732	pgd_t *pgd;
 733	p4d_t *p4d;
 734	pud_t *pud;
 735
 736	pgd = pgd_offset_k(address);
 737	if (pgd_none(*pgd))
 738		return NULL;
 739
 740	p4d = p4d_offset(pgd, address);
 741	if (p4d_none(*p4d) || p4d_large(*p4d) || !p4d_present(*p4d))
 742		return NULL;
 743
 744	pud = pud_offset(p4d, address);
 745	if (pud_none(*pud) || pud_large(*pud) || !pud_present(*pud))
 746		return NULL;
 747
 748	return pmd_offset(pud, address);
 749}
 750
 751/*
 752 * This is necessary because __pa() does not work on some
 753 * kinds of memory, like vmalloc() or the alloc_remap()
 754 * areas on 32-bit NUMA systems.  The percpu areas can
 755 * end up in this kind of memory, for instance.
 756 *
 757 * This could be optimized, but it is only intended to be
 758 * used at initialization time, and keeping it
 759 * unoptimized should increase the testing coverage for
 760 * the more obscure platforms.
 
 
 
 
 761 */
 762phys_addr_t slow_virt_to_phys(void *__virt_addr)
 763{
 764	unsigned long virt_addr = (unsigned long)__virt_addr;
 765	phys_addr_t phys_addr;
 766	unsigned long offset;
 767	enum pg_level level;
 768	pte_t *pte;
 769
 770	pte = lookup_address(virt_addr, &level);
 771	BUG_ON(!pte);
 772
 773	/*
 774	 * pXX_pfn() returns unsigned long, which must be cast to phys_addr_t
 775	 * before being left-shifted PAGE_SHIFT bits -- this trick is to
 776	 * make 32-PAE kernel work correctly.
 777	 */
 778	switch (level) {
 779	case PG_LEVEL_1G:
 780		phys_addr = (phys_addr_t)pud_pfn(*(pud_t *)pte) << PAGE_SHIFT;
 781		offset = virt_addr & ~PUD_MASK;
 782		break;
 783	case PG_LEVEL_2M:
 784		phys_addr = (phys_addr_t)pmd_pfn(*(pmd_t *)pte) << PAGE_SHIFT;
 785		offset = virt_addr & ~PMD_MASK;
 786		break;
 787	default:
 788		phys_addr = (phys_addr_t)pte_pfn(*pte) << PAGE_SHIFT;
 789		offset = virt_addr & ~PAGE_MASK;
 790	}
 791
 792	return (phys_addr_t)(phys_addr | offset);
 793}
 794EXPORT_SYMBOL_GPL(slow_virt_to_phys);
 795
 796/*
 797 * Set the new pmd in all the pgds we know about:
 798 */
 799static void __set_pmd_pte(pte_t *kpte, unsigned long address, pte_t pte)
 800{
 801	/* change init_mm */
 802	set_pte_atomic(kpte, pte);
 803#ifdef CONFIG_X86_32
 804	if (!SHARED_KERNEL_PMD) {
 805		struct page *page;
 806
 807		list_for_each_entry(page, &pgd_list, lru) {
 808			pgd_t *pgd;
 809			p4d_t *p4d;
 810			pud_t *pud;
 811			pmd_t *pmd;
 812
 813			pgd = (pgd_t *)page_address(page) + pgd_index(address);
 814			p4d = p4d_offset(pgd, address);
 815			pud = pud_offset(p4d, address);
 816			pmd = pmd_offset(pud, address);
 817			set_pte_atomic((pte_t *)pmd, pte);
 818		}
 819	}
 820#endif
 821}
 822
 823static pgprot_t pgprot_clear_protnone_bits(pgprot_t prot)
 824{
 825	/*
 826	 * _PAGE_GLOBAL means "global page" for present PTEs.
 827	 * But, it is also used to indicate _PAGE_PROTNONE
 828	 * for non-present PTEs.
 829	 *
 830	 * This ensures that a _PAGE_GLOBAL PTE going from
 831	 * present to non-present is not confused as
 832	 * _PAGE_PROTNONE.
 833	 */
 834	if (!(pgprot_val(prot) & _PAGE_PRESENT))
 835		pgprot_val(prot) &= ~_PAGE_GLOBAL;
 836
 837	return prot;
 838}
 839
 840static int __should_split_large_page(pte_t *kpte, unsigned long address,
 841				     struct cpa_data *cpa)
 842{
 843	unsigned long numpages, pmask, psize, lpaddr, pfn, old_pfn;
 844	pgprot_t old_prot, new_prot, req_prot, chk_prot;
 845	pte_t new_pte, *tmp;
 846	enum pg_level level;
 
 847
 848	/*
 849	 * Check for races, another CPU might have split this page
 850	 * up already:
 851	 */
 852	tmp = _lookup_address_cpa(cpa, address, &level);
 853	if (tmp != kpte)
 854		return 1;
 855
 856	switch (level) {
 857	case PG_LEVEL_2M:
 858		old_prot = pmd_pgprot(*(pmd_t *)kpte);
 859		old_pfn = pmd_pfn(*(pmd_t *)kpte);
 860		cpa_inc_2m_checked();
 861		break;
 862	case PG_LEVEL_1G:
 863		old_prot = pud_pgprot(*(pud_t *)kpte);
 864		old_pfn = pud_pfn(*(pud_t *)kpte);
 865		cpa_inc_1g_checked();
 866		break;
 867	default:
 868		return -EINVAL;
 869	}
 870
 871	psize = page_level_size(level);
 872	pmask = page_level_mask(level);
 873
 874	/*
 875	 * Calculate the number of pages, which fit into this large
 876	 * page starting at address:
 877	 */
 878	lpaddr = (address + psize) & pmask;
 879	numpages = (lpaddr - address) >> PAGE_SHIFT;
 880	if (numpages < cpa->numpages)
 881		cpa->numpages = numpages;
 882
 883	/*
 884	 * We are safe now. Check whether the new pgprot is the same:
 885	 * Convert protection attributes to 4k-format, as cpa->mask* are set
 886	 * up accordingly.
 887	 */
 888
 889	/* Clear PSE (aka _PAGE_PAT) and move PAT bit to correct position */
 890	req_prot = pgprot_large_2_4k(old_prot);
 891
 892	pgprot_val(req_prot) &= ~pgprot_val(cpa->mask_clr);
 893	pgprot_val(req_prot) |= pgprot_val(cpa->mask_set);
 894
 895	/*
 896	 * req_prot is in format of 4k pages. It must be converted to large
 897	 * page format: the caching mode includes the PAT bit located at
 898	 * different bit positions in the two formats.
 899	 */
 900	req_prot = pgprot_4k_2_large(req_prot);
 901	req_prot = pgprot_clear_protnone_bits(req_prot);
 902	if (pgprot_val(req_prot) & _PAGE_PRESENT)
 903		pgprot_val(req_prot) |= _PAGE_PSE;
 904
 905	/*
 906	 * old_pfn points to the large page base pfn. So we need to add the
 907	 * offset of the virtual address:
 908	 */
 909	pfn = old_pfn + ((address & (psize - 1)) >> PAGE_SHIFT);
 910	cpa->pfn = pfn;
 911
 912	/*
 913	 * Calculate the large page base address and the number of 4K pages
 914	 * in the large page
 915	 */
 916	lpaddr = address & pmask;
 917	numpages = psize >> PAGE_SHIFT;
 918
 919	/*
 920	 * Sanity check that the existing mapping is correct versus the static
 921	 * protections. static_protections() guards against !PRESENT, so no
 922	 * extra conditional required here.
 923	 */
 924	chk_prot = static_protections(old_prot, lpaddr, old_pfn, numpages,
 925				      psize, CPA_CONFLICT);
 926
 927	if (WARN_ON_ONCE(pgprot_val(chk_prot) != pgprot_val(old_prot))) {
 928		/*
 929		 * Split the large page and tell the split code to
 930		 * enforce static protections.
 931		 */
 932		cpa->force_static_prot = 1;
 933		return 1;
 934	}
 935
 936	/*
 937	 * Optimization: If the requested pgprot is the same as the current
 938	 * pgprot, then the large page can be preserved and no updates are
 939	 * required independent of alignment and length of the requested
 940	 * range. The above already established that the current pgprot is
 941	 * correct, which in consequence makes the requested pgprot correct
 942	 * as well if it is the same. The static protection scan below will
 943	 * not come to a different conclusion.
 944	 */
 945	if (pgprot_val(req_prot) == pgprot_val(old_prot)) {
 946		cpa_inc_lp_sameprot(level);
 947		return 0;
 948	}
 949
 950	/*
 951	 * If the requested range does not cover the full page, split it up
 952	 */
 953	if (address != lpaddr || cpa->numpages != numpages)
 954		return 1;
 955
 956	/*
 957	 * Check whether the requested pgprot is conflicting with a static
 958	 * protection requirement in the large page.
 959	 */
 960	new_prot = static_protections(req_prot, lpaddr, old_pfn, numpages,
 961				      psize, CPA_DETECT);
 962
 963	new_prot = verify_rwx(old_prot, new_prot, lpaddr, old_pfn, numpages);
 
 964
 965	/*
 966	 * If there is a conflict, split the large page.
 967	 *
 968	 * There used to be a 4k wise evaluation trying really hard to
 969	 * preserve the large pages, but experimentation has shown, that this
 970	 * does not help at all. There might be corner cases which would
 971	 * preserve one large page occasionally, but it's really not worth the
 972	 * extra code and cycles for the common case.
 973	 */
 974	if (pgprot_val(req_prot) != pgprot_val(new_prot))
 975		return 1;
 976
 977	/* All checks passed. Update the large page mapping. */
 978	new_pte = pfn_pte(old_pfn, new_prot);
 979	__set_pmd_pte(kpte, address, new_pte);
 980	cpa->flags |= CPA_FLUSHTLB;
 981	cpa_inc_lp_preserved(level);
 982	return 0;
 983}
 984
 985static int should_split_large_page(pte_t *kpte, unsigned long address,
 986				   struct cpa_data *cpa)
 987{
 988	int do_split;
 989
 990	if (cpa->force_split)
 991		return 1;
 992
 993	spin_lock(&pgd_lock);
 994	do_split = __should_split_large_page(kpte, address, cpa);
 995	spin_unlock(&pgd_lock);
 996
 997	return do_split;
 998}
 999
1000static void split_set_pte(struct cpa_data *cpa, pte_t *pte, unsigned long pfn,
1001			  pgprot_t ref_prot, unsigned long address,
1002			  unsigned long size)
1003{
1004	unsigned int npg = PFN_DOWN(size);
1005	pgprot_t prot;
1006
1007	/*
1008	 * If should_split_large_page() discovered an inconsistent mapping,
1009	 * remove the invalid protection in the split mapping.
1010	 */
1011	if (!cpa->force_static_prot)
1012		goto set;
1013
1014	/* Hand in lpsize = 0 to enforce the protection mechanism */
1015	prot = static_protections(ref_prot, address, pfn, npg, 0, CPA_PROTECT);
1016
1017	if (pgprot_val(prot) == pgprot_val(ref_prot))
1018		goto set;
1019
1020	/*
1021	 * If this is splitting a PMD, fix it up. PUD splits cannot be
1022	 * fixed trivially as that would require to rescan the newly
1023	 * installed PMD mappings after returning from split_large_page()
1024	 * so an eventual further split can allocate the necessary PTE
1025	 * pages. Warn for now and revisit it in case this actually
1026	 * happens.
1027	 */
1028	if (size == PAGE_SIZE)
1029		ref_prot = prot;
1030	else
1031		pr_warn_once("CPA: Cannot fixup static protections for PUD split\n");
1032set:
1033	set_pte(pte, pfn_pte(pfn, ref_prot));
1034}
1035
1036static int
1037__split_large_page(struct cpa_data *cpa, pte_t *kpte, unsigned long address,
1038		   struct page *base)
1039{
1040	unsigned long lpaddr, lpinc, ref_pfn, pfn, pfninc = 1;
1041	pte_t *pbase = (pte_t *)page_address(base);
1042	unsigned int i, level;
1043	pgprot_t ref_prot;
 
1044	pte_t *tmp;
1045
1046	spin_lock(&pgd_lock);
1047	/*
1048	 * Check for races, another CPU might have split this page
1049	 * up for us already:
1050	 */
1051	tmp = _lookup_address_cpa(cpa, address, &level);
1052	if (tmp != kpte) {
1053		spin_unlock(&pgd_lock);
1054		return 1;
1055	}
1056
1057	paravirt_alloc_pte(&init_mm, page_to_pfn(base));
1058
1059	switch (level) {
1060	case PG_LEVEL_2M:
1061		ref_prot = pmd_pgprot(*(pmd_t *)kpte);
1062		/*
1063		 * Clear PSE (aka _PAGE_PAT) and move
1064		 * PAT bit to correct position.
1065		 */
1066		ref_prot = pgprot_large_2_4k(ref_prot);
1067		ref_pfn = pmd_pfn(*(pmd_t *)kpte);
1068		lpaddr = address & PMD_MASK;
1069		lpinc = PAGE_SIZE;
1070		break;
1071
1072	case PG_LEVEL_1G:
1073		ref_prot = pud_pgprot(*(pud_t *)kpte);
1074		ref_pfn = pud_pfn(*(pud_t *)kpte);
1075		pfninc = PMD_SIZE >> PAGE_SHIFT;
1076		lpaddr = address & PUD_MASK;
1077		lpinc = PMD_SIZE;
1078		/*
1079		 * Clear the PSE flags if the PRESENT flag is not set
1080		 * otherwise pmd_present/pmd_huge will return true
1081		 * even on a non present pmd.
1082		 */
1083		if (!(pgprot_val(ref_prot) & _PAGE_PRESENT))
1084			pgprot_val(ref_prot) &= ~_PAGE_PSE;
1085		break;
1086
1087	default:
1088		spin_unlock(&pgd_lock);
1089		return 1;
1090	}
1091
1092	ref_prot = pgprot_clear_protnone_bits(ref_prot);
1093
1094	/*
1095	 * Get the target pfn from the original entry:
1096	 */
1097	pfn = ref_pfn;
1098	for (i = 0; i < PTRS_PER_PTE; i++, pfn += pfninc, lpaddr += lpinc)
1099		split_set_pte(cpa, pbase + i, pfn, ref_prot, lpaddr, lpinc);
1100
1101	if (virt_addr_valid(address)) {
1102		unsigned long pfn = PFN_DOWN(__pa(address));
1103
1104		if (pfn_range_is_mapped(pfn, pfn + 1))
1105			split_page_count(level);
1106	}
1107
1108	/*
1109	 * Install the new, split up pagetable.
1110	 *
1111	 * We use the standard kernel pagetable protections for the new
1112	 * pagetable protections, the actual ptes set above control the
1113	 * primary protection behavior:
1114	 */
1115	__set_pmd_pte(kpte, address, mk_pte(base, __pgprot(_KERNPG_TABLE)));
1116
1117	/*
1118	 * Do a global flush tlb after splitting the large page
1119	 * and before we do the actual change page attribute in the PTE.
1120	 *
1121	 * Without this, we violate the TLB application note, that says:
1122	 * "The TLBs may contain both ordinary and large-page
1123	 *  translations for a 4-KByte range of linear addresses. This
1124	 *  may occur if software modifies the paging structures so that
1125	 *  the page size used for the address range changes. If the two
1126	 *  translations differ with respect to page frame or attributes
1127	 *  (e.g., permissions), processor behavior is undefined and may
1128	 *  be implementation-specific."
1129	 *
1130	 * We do this global tlb flush inside the cpa_lock, so that we
1131	 * don't allow any other cpu, with stale tlb entries change the
1132	 * page attribute in parallel, that also falls into the
1133	 * just split large page entry.
1134	 */
1135	flush_tlb_all();
1136	spin_unlock(&pgd_lock);
1137
1138	return 0;
1139}
1140
1141static int split_large_page(struct cpa_data *cpa, pte_t *kpte,
1142			    unsigned long address)
1143{
1144	struct page *base;
1145
1146	if (!debug_pagealloc_enabled())
1147		spin_unlock(&cpa_lock);
1148	base = alloc_pages(GFP_KERNEL, 0);
1149	if (!debug_pagealloc_enabled())
1150		spin_lock(&cpa_lock);
1151	if (!base)
1152		return -ENOMEM;
1153
1154	if (__split_large_page(cpa, kpte, address, base))
1155		__free_page(base);
1156
1157	return 0;
1158}
1159
1160static bool try_to_free_pte_page(pte_t *pte)
1161{
1162	int i;
1163
1164	for (i = 0; i < PTRS_PER_PTE; i++)
1165		if (!pte_none(pte[i]))
1166			return false;
1167
1168	free_page((unsigned long)pte);
1169	return true;
1170}
1171
1172static bool try_to_free_pmd_page(pmd_t *pmd)
1173{
1174	int i;
1175
1176	for (i = 0; i < PTRS_PER_PMD; i++)
1177		if (!pmd_none(pmd[i]))
1178			return false;
1179
1180	free_page((unsigned long)pmd);
1181	return true;
1182}
1183
1184static bool unmap_pte_range(pmd_t *pmd, unsigned long start, unsigned long end)
1185{
1186	pte_t *pte = pte_offset_kernel(pmd, start);
1187
1188	while (start < end) {
1189		set_pte(pte, __pte(0));
1190
1191		start += PAGE_SIZE;
1192		pte++;
1193	}
1194
1195	if (try_to_free_pte_page((pte_t *)pmd_page_vaddr(*pmd))) {
1196		pmd_clear(pmd);
1197		return true;
1198	}
1199	return false;
1200}
1201
1202static void __unmap_pmd_range(pud_t *pud, pmd_t *pmd,
1203			      unsigned long start, unsigned long end)
1204{
1205	if (unmap_pte_range(pmd, start, end))
1206		if (try_to_free_pmd_page(pud_pgtable(*pud)))
1207			pud_clear(pud);
1208}
1209
1210static void unmap_pmd_range(pud_t *pud, unsigned long start, unsigned long end)
1211{
1212	pmd_t *pmd = pmd_offset(pud, start);
1213
1214	/*
1215	 * Not on a 2MB page boundary?
1216	 */
1217	if (start & (PMD_SIZE - 1)) {
1218		unsigned long next_page = (start + PMD_SIZE) & PMD_MASK;
1219		unsigned long pre_end = min_t(unsigned long, end, next_page);
1220
1221		__unmap_pmd_range(pud, pmd, start, pre_end);
1222
1223		start = pre_end;
1224		pmd++;
1225	}
1226
1227	/*
1228	 * Try to unmap in 2M chunks.
1229	 */
1230	while (end - start >= PMD_SIZE) {
1231		if (pmd_large(*pmd))
1232			pmd_clear(pmd);
1233		else
1234			__unmap_pmd_range(pud, pmd, start, start + PMD_SIZE);
1235
1236		start += PMD_SIZE;
1237		pmd++;
1238	}
1239
1240	/*
1241	 * 4K leftovers?
1242	 */
1243	if (start < end)
1244		return __unmap_pmd_range(pud, pmd, start, end);
1245
1246	/*
1247	 * Try again to free the PMD page if haven't succeeded above.
1248	 */
1249	if (!pud_none(*pud))
1250		if (try_to_free_pmd_page(pud_pgtable(*pud)))
1251			pud_clear(pud);
1252}
1253
1254static void unmap_pud_range(p4d_t *p4d, unsigned long start, unsigned long end)
1255{
1256	pud_t *pud = pud_offset(p4d, start);
1257
1258	/*
1259	 * Not on a GB page boundary?
1260	 */
1261	if (start & (PUD_SIZE - 1)) {
1262		unsigned long next_page = (start + PUD_SIZE) & PUD_MASK;
1263		unsigned long pre_end	= min_t(unsigned long, end, next_page);
1264
1265		unmap_pmd_range(pud, start, pre_end);
1266
1267		start = pre_end;
1268		pud++;
1269	}
1270
1271	/*
1272	 * Try to unmap in 1G chunks?
1273	 */
1274	while (end - start >= PUD_SIZE) {
1275
1276		if (pud_large(*pud))
1277			pud_clear(pud);
1278		else
1279			unmap_pmd_range(pud, start, start + PUD_SIZE);
1280
1281		start += PUD_SIZE;
1282		pud++;
1283	}
1284
1285	/*
1286	 * 2M leftovers?
1287	 */
1288	if (start < end)
1289		unmap_pmd_range(pud, start, end);
1290
1291	/*
1292	 * No need to try to free the PUD page because we'll free it in
1293	 * populate_pgd's error path
1294	 */
1295}
1296
1297static int alloc_pte_page(pmd_t *pmd)
1298{
1299	pte_t *pte = (pte_t *)get_zeroed_page(GFP_KERNEL);
1300	if (!pte)
1301		return -1;
1302
1303	set_pmd(pmd, __pmd(__pa(pte) | _KERNPG_TABLE));
1304	return 0;
1305}
1306
1307static int alloc_pmd_page(pud_t *pud)
1308{
1309	pmd_t *pmd = (pmd_t *)get_zeroed_page(GFP_KERNEL);
1310	if (!pmd)
1311		return -1;
1312
1313	set_pud(pud, __pud(__pa(pmd) | _KERNPG_TABLE));
1314	return 0;
1315}
1316
1317static void populate_pte(struct cpa_data *cpa,
1318			 unsigned long start, unsigned long end,
1319			 unsigned num_pages, pmd_t *pmd, pgprot_t pgprot)
1320{
1321	pte_t *pte;
1322
1323	pte = pte_offset_kernel(pmd, start);
1324
1325	pgprot = pgprot_clear_protnone_bits(pgprot);
1326
1327	while (num_pages-- && start < end) {
1328		set_pte(pte, pfn_pte(cpa->pfn, pgprot));
1329
1330		start	 += PAGE_SIZE;
1331		cpa->pfn++;
1332		pte++;
1333	}
1334}
1335
1336static long populate_pmd(struct cpa_data *cpa,
1337			 unsigned long start, unsigned long end,
1338			 unsigned num_pages, pud_t *pud, pgprot_t pgprot)
1339{
1340	long cur_pages = 0;
1341	pmd_t *pmd;
1342	pgprot_t pmd_pgprot;
1343
1344	/*
1345	 * Not on a 2M boundary?
1346	 */
1347	if (start & (PMD_SIZE - 1)) {
1348		unsigned long pre_end = start + (num_pages << PAGE_SHIFT);
1349		unsigned long next_page = (start + PMD_SIZE) & PMD_MASK;
1350
1351		pre_end   = min_t(unsigned long, pre_end, next_page);
1352		cur_pages = (pre_end - start) >> PAGE_SHIFT;
1353		cur_pages = min_t(unsigned int, num_pages, cur_pages);
1354
1355		/*
1356		 * Need a PTE page?
1357		 */
1358		pmd = pmd_offset(pud, start);
1359		if (pmd_none(*pmd))
1360			if (alloc_pte_page(pmd))
1361				return -1;
1362
1363		populate_pte(cpa, start, pre_end, cur_pages, pmd, pgprot);
1364
1365		start = pre_end;
1366	}
1367
1368	/*
1369	 * We mapped them all?
1370	 */
1371	if (num_pages == cur_pages)
1372		return cur_pages;
1373
1374	pmd_pgprot = pgprot_4k_2_large(pgprot);
1375
1376	while (end - start >= PMD_SIZE) {
1377
1378		/*
1379		 * We cannot use a 1G page so allocate a PMD page if needed.
1380		 */
1381		if (pud_none(*pud))
1382			if (alloc_pmd_page(pud))
1383				return -1;
1384
1385		pmd = pmd_offset(pud, start);
1386
1387		set_pmd(pmd, pmd_mkhuge(pfn_pmd(cpa->pfn,
1388					canon_pgprot(pmd_pgprot))));
1389
1390		start	  += PMD_SIZE;
1391		cpa->pfn  += PMD_SIZE >> PAGE_SHIFT;
1392		cur_pages += PMD_SIZE >> PAGE_SHIFT;
1393	}
1394
1395	/*
1396	 * Map trailing 4K pages.
1397	 */
1398	if (start < end) {
1399		pmd = pmd_offset(pud, start);
1400		if (pmd_none(*pmd))
1401			if (alloc_pte_page(pmd))
1402				return -1;
1403
1404		populate_pte(cpa, start, end, num_pages - cur_pages,
1405			     pmd, pgprot);
1406	}
1407	return num_pages;
1408}
1409
1410static int populate_pud(struct cpa_data *cpa, unsigned long start, p4d_t *p4d,
1411			pgprot_t pgprot)
1412{
1413	pud_t *pud;
1414	unsigned long end;
1415	long cur_pages = 0;
1416	pgprot_t pud_pgprot;
1417
1418	end = start + (cpa->numpages << PAGE_SHIFT);
1419
1420	/*
1421	 * Not on a Gb page boundary? => map everything up to it with
1422	 * smaller pages.
1423	 */
1424	if (start & (PUD_SIZE - 1)) {
1425		unsigned long pre_end;
1426		unsigned long next_page = (start + PUD_SIZE) & PUD_MASK;
1427
1428		pre_end   = min_t(unsigned long, end, next_page);
1429		cur_pages = (pre_end - start) >> PAGE_SHIFT;
1430		cur_pages = min_t(int, (int)cpa->numpages, cur_pages);
1431
1432		pud = pud_offset(p4d, start);
1433
1434		/*
1435		 * Need a PMD page?
1436		 */
1437		if (pud_none(*pud))
1438			if (alloc_pmd_page(pud))
1439				return -1;
1440
1441		cur_pages = populate_pmd(cpa, start, pre_end, cur_pages,
1442					 pud, pgprot);
1443		if (cur_pages < 0)
1444			return cur_pages;
1445
1446		start = pre_end;
1447	}
1448
1449	/* We mapped them all? */
1450	if (cpa->numpages == cur_pages)
1451		return cur_pages;
1452
1453	pud = pud_offset(p4d, start);
1454	pud_pgprot = pgprot_4k_2_large(pgprot);
1455
1456	/*
1457	 * Map everything starting from the Gb boundary, possibly with 1G pages
1458	 */
1459	while (boot_cpu_has(X86_FEATURE_GBPAGES) && end - start >= PUD_SIZE) {
1460		set_pud(pud, pud_mkhuge(pfn_pud(cpa->pfn,
1461				   canon_pgprot(pud_pgprot))));
1462
1463		start	  += PUD_SIZE;
1464		cpa->pfn  += PUD_SIZE >> PAGE_SHIFT;
1465		cur_pages += PUD_SIZE >> PAGE_SHIFT;
1466		pud++;
1467	}
1468
1469	/* Map trailing leftover */
1470	if (start < end) {
1471		long tmp;
1472
1473		pud = pud_offset(p4d, start);
1474		if (pud_none(*pud))
1475			if (alloc_pmd_page(pud))
1476				return -1;
1477
1478		tmp = populate_pmd(cpa, start, end, cpa->numpages - cur_pages,
1479				   pud, pgprot);
1480		if (tmp < 0)
1481			return cur_pages;
1482
1483		cur_pages += tmp;
1484	}
1485	return cur_pages;
1486}
1487
1488/*
1489 * Restrictions for kernel page table do not necessarily apply when mapping in
1490 * an alternate PGD.
1491 */
1492static int populate_pgd(struct cpa_data *cpa, unsigned long addr)
1493{
1494	pgprot_t pgprot = __pgprot(_KERNPG_TABLE);
1495	pud_t *pud = NULL;	/* shut up gcc */
1496	p4d_t *p4d;
1497	pgd_t *pgd_entry;
1498	long ret;
1499
1500	pgd_entry = cpa->pgd + pgd_index(addr);
1501
1502	if (pgd_none(*pgd_entry)) {
1503		p4d = (p4d_t *)get_zeroed_page(GFP_KERNEL);
1504		if (!p4d)
1505			return -1;
1506
1507		set_pgd(pgd_entry, __pgd(__pa(p4d) | _KERNPG_TABLE));
1508	}
1509
1510	/*
1511	 * Allocate a PUD page and hand it down for mapping.
1512	 */
1513	p4d = p4d_offset(pgd_entry, addr);
1514	if (p4d_none(*p4d)) {
1515		pud = (pud_t *)get_zeroed_page(GFP_KERNEL);
1516		if (!pud)
1517			return -1;
1518
1519		set_p4d(p4d, __p4d(__pa(pud) | _KERNPG_TABLE));
1520	}
1521
1522	pgprot_val(pgprot) &= ~pgprot_val(cpa->mask_clr);
1523	pgprot_val(pgprot) |=  pgprot_val(cpa->mask_set);
1524
1525	ret = populate_pud(cpa, addr, p4d, pgprot);
1526	if (ret < 0) {
1527		/*
1528		 * Leave the PUD page in place in case some other CPU or thread
1529		 * already found it, but remove any useless entries we just
1530		 * added to it.
1531		 */
1532		unmap_pud_range(p4d, addr,
1533				addr + (cpa->numpages << PAGE_SHIFT));
1534		return ret;
1535	}
1536
1537	cpa->numpages = ret;
1538	return 0;
1539}
1540
1541static int __cpa_process_fault(struct cpa_data *cpa, unsigned long vaddr,
1542			       int primary)
1543{
1544	if (cpa->pgd) {
1545		/*
1546		 * Right now, we only execute this code path when mapping
1547		 * the EFI virtual memory map regions, no other users
1548		 * provide a ->pgd value. This may change in the future.
1549		 */
1550		return populate_pgd(cpa, vaddr);
1551	}
1552
1553	/*
1554	 * Ignore all non primary paths.
1555	 */
1556	if (!primary) {
1557		cpa->numpages = 1;
1558		return 0;
1559	}
1560
1561	/*
1562	 * Ignore the NULL PTE for kernel identity mapping, as it is expected
1563	 * to have holes.
1564	 * Also set numpages to '1' indicating that we processed cpa req for
1565	 * one virtual address page and its pfn. TBD: numpages can be set based
1566	 * on the initial value and the level returned by lookup_address().
1567	 */
1568	if (within(vaddr, PAGE_OFFSET,
1569		   PAGE_OFFSET + (max_pfn_mapped << PAGE_SHIFT))) {
1570		cpa->numpages = 1;
1571		cpa->pfn = __pa(vaddr) >> PAGE_SHIFT;
1572		return 0;
1573
1574	} else if (__cpa_pfn_in_highmap(cpa->pfn)) {
1575		/* Faults in the highmap are OK, so do not warn: */
1576		return -EFAULT;
1577	} else {
1578		WARN(1, KERN_WARNING "CPA: called for zero pte. "
1579			"vaddr = %lx cpa->vaddr = %lx\n", vaddr,
1580			*cpa->vaddr);
1581
1582		return -EFAULT;
1583	}
1584}
1585
1586static int __change_page_attr(struct cpa_data *cpa, int primary)
1587{
1588	unsigned long address;
1589	int do_split, err;
1590	unsigned int level;
1591	pte_t *kpte, old_pte;
 
1592
1593	address = __cpa_addr(cpa, cpa->curpage);
1594repeat:
1595	kpte = _lookup_address_cpa(cpa, address, &level);
1596	if (!kpte)
1597		return __cpa_process_fault(cpa, address, primary);
1598
1599	old_pte = *kpte;
1600	if (pte_none(old_pte))
1601		return __cpa_process_fault(cpa, address, primary);
1602
1603	if (level == PG_LEVEL_4K) {
1604		pte_t new_pte;
1605		pgprot_t old_prot = pte_pgprot(old_pte);
1606		pgprot_t new_prot = pte_pgprot(old_pte);
1607		unsigned long pfn = pte_pfn(old_pte);
1608
1609		pgprot_val(new_prot) &= ~pgprot_val(cpa->mask_clr);
1610		pgprot_val(new_prot) |= pgprot_val(cpa->mask_set);
1611
1612		cpa_inc_4k_install();
1613		/* Hand in lpsize = 0 to enforce the protection mechanism */
1614		new_prot = static_protections(new_prot, address, pfn, 1, 0,
1615					      CPA_PROTECT);
1616
1617		new_prot = verify_rwx(old_prot, new_prot, address, pfn, 1);
 
1618
1619		new_prot = pgprot_clear_protnone_bits(new_prot);
1620
1621		/*
1622		 * We need to keep the pfn from the existing PTE,
1623		 * after all we're only going to change it's attributes
1624		 * not the memory it points to
1625		 */
1626		new_pte = pfn_pte(pfn, new_prot);
1627		cpa->pfn = pfn;
1628		/*
1629		 * Do we really change anything ?
1630		 */
1631		if (pte_val(old_pte) != pte_val(new_pte)) {
1632			set_pte_atomic(kpte, new_pte);
1633			cpa->flags |= CPA_FLUSHTLB;
1634		}
1635		cpa->numpages = 1;
1636		return 0;
1637	}
1638
1639	/*
1640	 * Check, whether we can keep the large page intact
1641	 * and just change the pte:
1642	 */
1643	do_split = should_split_large_page(kpte, address, cpa);
1644	/*
1645	 * When the range fits into the existing large page,
1646	 * return. cp->numpages and cpa->tlbflush have been updated in
1647	 * try_large_page:
1648	 */
1649	if (do_split <= 0)
1650		return do_split;
1651
1652	/*
1653	 * We have to split the large page:
1654	 */
1655	err = split_large_page(cpa, kpte, address);
1656	if (!err)
1657		goto repeat;
1658
1659	return err;
1660}
1661
1662static int __change_page_attr_set_clr(struct cpa_data *cpa, int primary);
1663
1664/*
1665 * Check the directmap and "high kernel map" 'aliases'.
1666 */
1667static int cpa_process_alias(struct cpa_data *cpa)
1668{
1669	struct cpa_data alias_cpa;
1670	unsigned long laddr = (unsigned long)__va(cpa->pfn << PAGE_SHIFT);
1671	unsigned long vaddr;
1672	int ret;
1673
1674	if (!pfn_range_is_mapped(cpa->pfn, cpa->pfn + 1))
1675		return 0;
1676
1677	/*
1678	 * No need to redo, when the primary call touched the direct
1679	 * mapping already:
1680	 */
1681	vaddr = __cpa_addr(cpa, cpa->curpage);
1682	if (!(within(vaddr, PAGE_OFFSET,
1683		    PAGE_OFFSET + (max_pfn_mapped << PAGE_SHIFT)))) {
1684
1685		alias_cpa = *cpa;
1686		alias_cpa.vaddr = &laddr;
1687		alias_cpa.flags &= ~(CPA_PAGES_ARRAY | CPA_ARRAY);
1688		alias_cpa.curpage = 0;
1689
1690		/* Directmap always has NX set, do not modify. */
1691		if (__supported_pte_mask & _PAGE_NX) {
1692			alias_cpa.mask_clr.pgprot &= ~_PAGE_NX;
1693			alias_cpa.mask_set.pgprot &= ~_PAGE_NX;
1694		}
1695
1696		cpa->force_flush_all = 1;
1697
1698		ret = __change_page_attr_set_clr(&alias_cpa, 0);
1699		if (ret)
1700			return ret;
1701	}
1702
1703#ifdef CONFIG_X86_64
1704	/*
1705	 * If the primary call didn't touch the high mapping already
1706	 * and the physical address is inside the kernel map, we need
1707	 * to touch the high mapped kernel as well:
1708	 */
1709	if (!within(vaddr, (unsigned long)_text, _brk_end) &&
1710	    __cpa_pfn_in_highmap(cpa->pfn)) {
1711		unsigned long temp_cpa_vaddr = (cpa->pfn << PAGE_SHIFT) +
1712					       __START_KERNEL_map - phys_base;
1713		alias_cpa = *cpa;
1714		alias_cpa.vaddr = &temp_cpa_vaddr;
1715		alias_cpa.flags &= ~(CPA_PAGES_ARRAY | CPA_ARRAY);
1716		alias_cpa.curpage = 0;
1717
1718		/*
1719		 * [_text, _brk_end) also covers data, do not modify NX except
1720		 * in cases where the highmap is the primary target.
1721		 */
1722		if (__supported_pte_mask & _PAGE_NX) {
1723			alias_cpa.mask_clr.pgprot &= ~_PAGE_NX;
1724			alias_cpa.mask_set.pgprot &= ~_PAGE_NX;
1725		}
1726
1727		cpa->force_flush_all = 1;
1728		/*
1729		 * The high mapping range is imprecise, so ignore the
1730		 * return value.
1731		 */
1732		__change_page_attr_set_clr(&alias_cpa, 0);
1733	}
1734#endif
1735
1736	return 0;
1737}
1738
1739static int __change_page_attr_set_clr(struct cpa_data *cpa, int primary)
1740{
1741	unsigned long numpages = cpa->numpages;
1742	unsigned long rempages = numpages;
1743	int ret = 0;
1744
1745	/*
1746	 * No changes, easy!
1747	 */
1748	if (!(pgprot_val(cpa->mask_set) | pgprot_val(cpa->mask_clr)) &&
1749	    !cpa->force_split)
1750		return ret;
1751
1752	while (rempages) {
1753		/*
1754		 * Store the remaining nr of pages for the large page
1755		 * preservation check.
1756		 */
1757		cpa->numpages = rempages;
1758		/* for array changes, we can't use large page */
1759		if (cpa->flags & (CPA_ARRAY | CPA_PAGES_ARRAY))
1760			cpa->numpages = 1;
1761
1762		if (!debug_pagealloc_enabled())
1763			spin_lock(&cpa_lock);
1764		ret = __change_page_attr(cpa, primary);
1765		if (!debug_pagealloc_enabled())
1766			spin_unlock(&cpa_lock);
1767		if (ret)
1768			goto out;
1769
1770		if (primary && !(cpa->flags & CPA_NO_CHECK_ALIAS)) {
1771			ret = cpa_process_alias(cpa);
1772			if (ret)
1773				goto out;
1774		}
1775
1776		/*
1777		 * Adjust the number of pages with the result of the
1778		 * CPA operation. Either a large page has been
1779		 * preserved or a single page update happened.
1780		 */
1781		BUG_ON(cpa->numpages > rempages || !cpa->numpages);
1782		rempages -= cpa->numpages;
1783		cpa->curpage += cpa->numpages;
1784	}
1785
1786out:
1787	/* Restore the original numpages */
1788	cpa->numpages = numpages;
1789	return ret;
1790}
1791
1792static int change_page_attr_set_clr(unsigned long *addr, int numpages,
1793				    pgprot_t mask_set, pgprot_t mask_clr,
1794				    int force_split, int in_flag,
1795				    struct page **pages)
1796{
1797	struct cpa_data cpa;
1798	int ret, cache;
1799
1800	memset(&cpa, 0, sizeof(cpa));
1801
1802	/*
1803	 * Check, if we are requested to set a not supported
1804	 * feature.  Clearing non-supported features is OK.
1805	 */
1806	mask_set = canon_pgprot(mask_set);
1807
1808	if (!pgprot_val(mask_set) && !pgprot_val(mask_clr) && !force_split)
1809		return 0;
1810
1811	/* Ensure we are PAGE_SIZE aligned */
1812	if (in_flag & CPA_ARRAY) {
1813		int i;
1814		for (i = 0; i < numpages; i++) {
1815			if (addr[i] & ~PAGE_MASK) {
1816				addr[i] &= PAGE_MASK;
1817				WARN_ON_ONCE(1);
1818			}
1819		}
1820	} else if (!(in_flag & CPA_PAGES_ARRAY)) {
1821		/*
1822		 * in_flag of CPA_PAGES_ARRAY implies it is aligned.
1823		 * No need to check in that case
1824		 */
1825		if (*addr & ~PAGE_MASK) {
1826			*addr &= PAGE_MASK;
1827			/*
1828			 * People should not be passing in unaligned addresses:
1829			 */
1830			WARN_ON_ONCE(1);
1831		}
1832	}
1833
1834	/* Must avoid aliasing mappings in the highmem code */
1835	kmap_flush_unused();
1836
1837	vm_unmap_aliases();
1838
1839	cpa.vaddr = addr;
1840	cpa.pages = pages;
1841	cpa.numpages = numpages;
1842	cpa.mask_set = mask_set;
1843	cpa.mask_clr = mask_clr;
1844	cpa.flags = in_flag;
1845	cpa.curpage = 0;
1846	cpa.force_split = force_split;
1847
1848	ret = __change_page_attr_set_clr(&cpa, 1);
1849
1850	/*
1851	 * Check whether we really changed something:
1852	 */
1853	if (!(cpa.flags & CPA_FLUSHTLB))
1854		goto out;
1855
1856	/*
1857	 * No need to flush, when we did not set any of the caching
1858	 * attributes:
1859	 */
1860	cache = !!pgprot2cachemode(mask_set);
1861
1862	/*
1863	 * On error; flush everything to be sure.
1864	 */
1865	if (ret) {
1866		cpa_flush_all(cache);
1867		goto out;
1868	}
1869
1870	cpa_flush(&cpa, cache);
1871out:
1872	return ret;
1873}
1874
1875static inline int change_page_attr_set(unsigned long *addr, int numpages,
1876				       pgprot_t mask, int array)
1877{
1878	return change_page_attr_set_clr(addr, numpages, mask, __pgprot(0), 0,
1879		(array ? CPA_ARRAY : 0), NULL);
1880}
1881
1882static inline int change_page_attr_clear(unsigned long *addr, int numpages,
1883					 pgprot_t mask, int array)
1884{
1885	return change_page_attr_set_clr(addr, numpages, __pgprot(0), mask, 0,
1886		(array ? CPA_ARRAY : 0), NULL);
1887}
1888
1889static inline int cpa_set_pages_array(struct page **pages, int numpages,
1890				       pgprot_t mask)
1891{
1892	return change_page_attr_set_clr(NULL, numpages, mask, __pgprot(0), 0,
1893		CPA_PAGES_ARRAY, pages);
1894}
1895
1896static inline int cpa_clear_pages_array(struct page **pages, int numpages,
1897					 pgprot_t mask)
1898{
1899	return change_page_attr_set_clr(NULL, numpages, __pgprot(0), mask, 0,
1900		CPA_PAGES_ARRAY, pages);
1901}
1902
1903/*
1904 * __set_memory_prot is an internal helper for callers that have been passed
1905 * a pgprot_t value from upper layers and a reservation has already been taken.
1906 * If you want to set the pgprot to a specific page protocol, use the
1907 * set_memory_xx() functions.
1908 */
1909int __set_memory_prot(unsigned long addr, int numpages, pgprot_t prot)
1910{
1911	return change_page_attr_set_clr(&addr, numpages, prot,
1912					__pgprot(~pgprot_val(prot)), 0, 0,
1913					NULL);
1914}
1915
1916int _set_memory_uc(unsigned long addr, int numpages)
1917{
1918	/*
1919	 * for now UC MINUS. see comments in ioremap()
1920	 * If you really need strong UC use ioremap_uc(), but note
1921	 * that you cannot override IO areas with set_memory_*() as
1922	 * these helpers cannot work with IO memory.
1923	 */
1924	return change_page_attr_set(&addr, numpages,
1925				    cachemode2pgprot(_PAGE_CACHE_MODE_UC_MINUS),
1926				    0);
1927}
1928
1929int set_memory_uc(unsigned long addr, int numpages)
1930{
1931	int ret;
1932
1933	/*
1934	 * for now UC MINUS. see comments in ioremap()
1935	 */
1936	ret = memtype_reserve(__pa(addr), __pa(addr) + numpages * PAGE_SIZE,
1937			      _PAGE_CACHE_MODE_UC_MINUS, NULL);
1938	if (ret)
1939		goto out_err;
1940
1941	ret = _set_memory_uc(addr, numpages);
1942	if (ret)
1943		goto out_free;
1944
1945	return 0;
1946
1947out_free:
1948	memtype_free(__pa(addr), __pa(addr) + numpages * PAGE_SIZE);
1949out_err:
1950	return ret;
1951}
1952EXPORT_SYMBOL(set_memory_uc);
1953
1954int _set_memory_wc(unsigned long addr, int numpages)
1955{
1956	int ret;
1957
1958	ret = change_page_attr_set(&addr, numpages,
1959				   cachemode2pgprot(_PAGE_CACHE_MODE_UC_MINUS),
1960				   0);
1961	if (!ret) {
1962		ret = change_page_attr_set_clr(&addr, numpages,
1963					       cachemode2pgprot(_PAGE_CACHE_MODE_WC),
1964					       __pgprot(_PAGE_CACHE_MASK),
1965					       0, 0, NULL);
1966	}
1967	return ret;
1968}
1969
1970int set_memory_wc(unsigned long addr, int numpages)
1971{
1972	int ret;
1973
1974	ret = memtype_reserve(__pa(addr), __pa(addr) + numpages * PAGE_SIZE,
1975		_PAGE_CACHE_MODE_WC, NULL);
1976	if (ret)
1977		return ret;
1978
1979	ret = _set_memory_wc(addr, numpages);
1980	if (ret)
1981		memtype_free(__pa(addr), __pa(addr) + numpages * PAGE_SIZE);
1982
1983	return ret;
1984}
1985EXPORT_SYMBOL(set_memory_wc);
1986
1987int _set_memory_wt(unsigned long addr, int numpages)
1988{
1989	return change_page_attr_set(&addr, numpages,
1990				    cachemode2pgprot(_PAGE_CACHE_MODE_WT), 0);
1991}
1992
1993int _set_memory_wb(unsigned long addr, int numpages)
1994{
1995	/* WB cache mode is hard wired to all cache attribute bits being 0 */
1996	return change_page_attr_clear(&addr, numpages,
1997				      __pgprot(_PAGE_CACHE_MASK), 0);
1998}
1999
2000int set_memory_wb(unsigned long addr, int numpages)
2001{
2002	int ret;
2003
2004	ret = _set_memory_wb(addr, numpages);
2005	if (ret)
2006		return ret;
2007
2008	memtype_free(__pa(addr), __pa(addr) + numpages * PAGE_SIZE);
2009	return 0;
2010}
2011EXPORT_SYMBOL(set_memory_wb);
2012
2013/* Prevent speculative access to a page by marking it not-present */
2014#ifdef CONFIG_X86_64
2015int set_mce_nospec(unsigned long pfn)
2016{
2017	unsigned long decoy_addr;
2018	int rc;
2019
2020	/* SGX pages are not in the 1:1 map */
2021	if (arch_is_platform_page(pfn << PAGE_SHIFT))
2022		return 0;
2023	/*
2024	 * We would like to just call:
2025	 *      set_memory_XX((unsigned long)pfn_to_kaddr(pfn), 1);
2026	 * but doing that would radically increase the odds of a
2027	 * speculative access to the poison page because we'd have
2028	 * the virtual address of the kernel 1:1 mapping sitting
2029	 * around in registers.
2030	 * Instead we get tricky.  We create a non-canonical address
2031	 * that looks just like the one we want, but has bit 63 flipped.
2032	 * This relies on set_memory_XX() properly sanitizing any __pa()
2033	 * results with __PHYSICAL_MASK or PTE_PFN_MASK.
2034	 */
2035	decoy_addr = (pfn << PAGE_SHIFT) + (PAGE_OFFSET ^ BIT(63));
2036
2037	rc = set_memory_np(decoy_addr, 1);
2038	if (rc)
2039		pr_warn("Could not invalidate pfn=0x%lx from 1:1 map\n", pfn);
2040	return rc;
2041}
2042
2043static int set_memory_p(unsigned long *addr, int numpages)
2044{
2045	return change_page_attr_set(addr, numpages, __pgprot(_PAGE_PRESENT), 0);
2046}
2047
2048/* Restore full speculative operation to the pfn. */
2049int clear_mce_nospec(unsigned long pfn)
2050{
2051	unsigned long addr = (unsigned long) pfn_to_kaddr(pfn);
2052
2053	return set_memory_p(&addr, 1);
2054}
2055EXPORT_SYMBOL_GPL(clear_mce_nospec);
2056#endif /* CONFIG_X86_64 */
2057
2058int set_memory_x(unsigned long addr, int numpages)
2059{
2060	if (!(__supported_pte_mask & _PAGE_NX))
2061		return 0;
2062
2063	return change_page_attr_clear(&addr, numpages, __pgprot(_PAGE_NX), 0);
2064}
2065
2066int set_memory_nx(unsigned long addr, int numpages)
2067{
2068	if (!(__supported_pte_mask & _PAGE_NX))
2069		return 0;
2070
2071	return change_page_attr_set(&addr, numpages, __pgprot(_PAGE_NX), 0);
2072}
2073
2074int set_memory_ro(unsigned long addr, int numpages)
2075{
2076	return change_page_attr_clear(&addr, numpages, __pgprot(_PAGE_RW), 0);
2077}
2078
2079int set_memory_rox(unsigned long addr, int numpages)
2080{
2081	pgprot_t clr = __pgprot(_PAGE_RW);
2082
2083	if (__supported_pte_mask & _PAGE_NX)
2084		clr.pgprot |= _PAGE_NX;
2085
2086	return change_page_attr_clear(&addr, numpages, clr, 0);
2087}
2088
2089int set_memory_rw(unsigned long addr, int numpages)
2090{
2091	return change_page_attr_set(&addr, numpages, __pgprot(_PAGE_RW), 0);
2092}
2093
2094int set_memory_np(unsigned long addr, int numpages)
2095{
2096	return change_page_attr_clear(&addr, numpages, __pgprot(_PAGE_PRESENT), 0);
2097}
2098
2099int set_memory_np_noalias(unsigned long addr, int numpages)
2100{
2101	return change_page_attr_set_clr(&addr, numpages, __pgprot(0),
2102					__pgprot(_PAGE_PRESENT), 0,
2103					CPA_NO_CHECK_ALIAS, NULL);
2104}
2105
 
 
 
 
 
2106int set_memory_4k(unsigned long addr, int numpages)
2107{
2108	return change_page_attr_set_clr(&addr, numpages, __pgprot(0),
2109					__pgprot(0), 1, 0, NULL);
2110}
2111
2112int set_memory_nonglobal(unsigned long addr, int numpages)
2113{
2114	return change_page_attr_clear(&addr, numpages,
2115				      __pgprot(_PAGE_GLOBAL), 0);
2116}
2117
2118int set_memory_global(unsigned long addr, int numpages)
2119{
2120	return change_page_attr_set(&addr, numpages,
2121				    __pgprot(_PAGE_GLOBAL), 0);
2122}
2123
2124/*
2125 * __set_memory_enc_pgtable() is used for the hypervisors that get
2126 * informed about "encryption" status via page tables.
2127 */
2128static int __set_memory_enc_pgtable(unsigned long addr, int numpages, bool enc)
2129{
2130	pgprot_t empty = __pgprot(0);
2131	struct cpa_data cpa;
2132	int ret;
2133
2134	/* Should not be working on unaligned addresses */
2135	if (WARN_ONCE(addr & ~PAGE_MASK, "misaligned address: %#lx\n", addr))
2136		addr &= PAGE_MASK;
2137
2138	memset(&cpa, 0, sizeof(cpa));
2139	cpa.vaddr = &addr;
2140	cpa.numpages = numpages;
2141	cpa.mask_set = enc ? pgprot_encrypted(empty) : pgprot_decrypted(empty);
2142	cpa.mask_clr = enc ? pgprot_decrypted(empty) : pgprot_encrypted(empty);
2143	cpa.pgd = init_mm.pgd;
2144
2145	/* Must avoid aliasing mappings in the highmem code */
2146	kmap_flush_unused();
2147	vm_unmap_aliases();
2148
2149	/* Flush the caches as needed before changing the encryption attribute. */
2150	if (x86_platform.guest.enc_tlb_flush_required(enc))
2151		cpa_flush(&cpa, x86_platform.guest.enc_cache_flush_required());
2152
2153	/* Notify hypervisor that we are about to set/clr encryption attribute. */
2154	x86_platform.guest.enc_status_change_prepare(addr, numpages, enc);
 
2155
2156	ret = __change_page_attr_set_clr(&cpa, 1);
2157
2158	/*
2159	 * After changing the encryption attribute, we need to flush TLBs again
2160	 * in case any speculative TLB caching occurred (but no need to flush
2161	 * caches again).  We could just use cpa_flush_all(), but in case TLB
2162	 * flushing gets optimized in the cpa_flush() path use the same logic
2163	 * as above.
2164	 */
2165	cpa_flush(&cpa, 0);
2166
 
 
 
2167	/* Notify hypervisor that we have successfully set/clr encryption attribute. */
2168	if (!ret) {
2169		if (!x86_platform.guest.enc_status_change_finish(addr, numpages, enc))
2170			ret = -EIO;
2171	}
2172
2173	return ret;
 
 
 
 
 
 
2174}
2175
2176static int __set_memory_enc_dec(unsigned long addr, int numpages, bool enc)
2177{
2178	if (hv_is_isolation_supported())
2179		return hv_set_mem_host_visibility(addr, numpages, !enc);
2180
2181	if (cc_platform_has(CC_ATTR_MEM_ENCRYPT))
2182		return __set_memory_enc_pgtable(addr, numpages, enc);
2183
2184	return 0;
2185}
2186
2187int set_memory_encrypted(unsigned long addr, int numpages)
2188{
2189	return __set_memory_enc_dec(addr, numpages, true);
2190}
2191EXPORT_SYMBOL_GPL(set_memory_encrypted);
2192
2193int set_memory_decrypted(unsigned long addr, int numpages)
2194{
2195	return __set_memory_enc_dec(addr, numpages, false);
2196}
2197EXPORT_SYMBOL_GPL(set_memory_decrypted);
2198
2199int set_pages_uc(struct page *page, int numpages)
2200{
2201	unsigned long addr = (unsigned long)page_address(page);
2202
2203	return set_memory_uc(addr, numpages);
2204}
2205EXPORT_SYMBOL(set_pages_uc);
2206
2207static int _set_pages_array(struct page **pages, int numpages,
2208		enum page_cache_mode new_type)
2209{
2210	unsigned long start;
2211	unsigned long end;
2212	enum page_cache_mode set_type;
2213	int i;
2214	int free_idx;
2215	int ret;
2216
2217	for (i = 0; i < numpages; i++) {
2218		if (PageHighMem(pages[i]))
2219			continue;
2220		start = page_to_pfn(pages[i]) << PAGE_SHIFT;
2221		end = start + PAGE_SIZE;
2222		if (memtype_reserve(start, end, new_type, NULL))
2223			goto err_out;
2224	}
2225
2226	/* If WC, set to UC- first and then WC */
2227	set_type = (new_type == _PAGE_CACHE_MODE_WC) ?
2228				_PAGE_CACHE_MODE_UC_MINUS : new_type;
2229
2230	ret = cpa_set_pages_array(pages, numpages,
2231				  cachemode2pgprot(set_type));
2232	if (!ret && new_type == _PAGE_CACHE_MODE_WC)
2233		ret = change_page_attr_set_clr(NULL, numpages,
2234					       cachemode2pgprot(
2235						_PAGE_CACHE_MODE_WC),
2236					       __pgprot(_PAGE_CACHE_MASK),
2237					       0, CPA_PAGES_ARRAY, pages);
2238	if (ret)
2239		goto err_out;
2240	return 0; /* Success */
2241err_out:
2242	free_idx = i;
2243	for (i = 0; i < free_idx; i++) {
2244		if (PageHighMem(pages[i]))
2245			continue;
2246		start = page_to_pfn(pages[i]) << PAGE_SHIFT;
2247		end = start + PAGE_SIZE;
2248		memtype_free(start, end);
2249	}
2250	return -EINVAL;
2251}
2252
2253int set_pages_array_uc(struct page **pages, int numpages)
2254{
2255	return _set_pages_array(pages, numpages, _PAGE_CACHE_MODE_UC_MINUS);
2256}
2257EXPORT_SYMBOL(set_pages_array_uc);
2258
2259int set_pages_array_wc(struct page **pages, int numpages)
2260{
2261	return _set_pages_array(pages, numpages, _PAGE_CACHE_MODE_WC);
2262}
2263EXPORT_SYMBOL(set_pages_array_wc);
2264
2265int set_pages_wb(struct page *page, int numpages)
2266{
2267	unsigned long addr = (unsigned long)page_address(page);
2268
2269	return set_memory_wb(addr, numpages);
2270}
2271EXPORT_SYMBOL(set_pages_wb);
2272
2273int set_pages_array_wb(struct page **pages, int numpages)
2274{
2275	int retval;
2276	unsigned long start;
2277	unsigned long end;
2278	int i;
2279
2280	/* WB cache mode is hard wired to all cache attribute bits being 0 */
2281	retval = cpa_clear_pages_array(pages, numpages,
2282			__pgprot(_PAGE_CACHE_MASK));
2283	if (retval)
2284		return retval;
2285
2286	for (i = 0; i < numpages; i++) {
2287		if (PageHighMem(pages[i]))
2288			continue;
2289		start = page_to_pfn(pages[i]) << PAGE_SHIFT;
2290		end = start + PAGE_SIZE;
2291		memtype_free(start, end);
2292	}
2293
2294	return 0;
2295}
2296EXPORT_SYMBOL(set_pages_array_wb);
2297
2298int set_pages_ro(struct page *page, int numpages)
2299{
2300	unsigned long addr = (unsigned long)page_address(page);
2301
2302	return set_memory_ro(addr, numpages);
2303}
2304
2305int set_pages_rw(struct page *page, int numpages)
2306{
2307	unsigned long addr = (unsigned long)page_address(page);
2308
2309	return set_memory_rw(addr, numpages);
2310}
2311
2312static int __set_pages_p(struct page *page, int numpages)
2313{
2314	unsigned long tempaddr = (unsigned long) page_address(page);
2315	struct cpa_data cpa = { .vaddr = &tempaddr,
2316				.pgd = NULL,
2317				.numpages = numpages,
2318				.mask_set = __pgprot(_PAGE_PRESENT | _PAGE_RW),
2319				.mask_clr = __pgprot(0),
2320				.flags = CPA_NO_CHECK_ALIAS };
2321
2322	/*
2323	 * No alias checking needed for setting present flag. otherwise,
2324	 * we may need to break large pages for 64-bit kernel text
2325	 * mappings (this adds to complexity if we want to do this from
2326	 * atomic context especially). Let's keep it simple!
2327	 */
2328	return __change_page_attr_set_clr(&cpa, 1);
2329}
2330
2331static int __set_pages_np(struct page *page, int numpages)
2332{
2333	unsigned long tempaddr = (unsigned long) page_address(page);
2334	struct cpa_data cpa = { .vaddr = &tempaddr,
2335				.pgd = NULL,
2336				.numpages = numpages,
2337				.mask_set = __pgprot(0),
2338				.mask_clr = __pgprot(_PAGE_PRESENT | _PAGE_RW),
2339				.flags = CPA_NO_CHECK_ALIAS };
2340
2341	/*
2342	 * No alias checking needed for setting not present flag. otherwise,
2343	 * we may need to break large pages for 64-bit kernel text
2344	 * mappings (this adds to complexity if we want to do this from
2345	 * atomic context especially). Let's keep it simple!
2346	 */
2347	return __change_page_attr_set_clr(&cpa, 1);
2348}
2349
2350int set_direct_map_invalid_noflush(struct page *page)
2351{
2352	return __set_pages_np(page, 1);
2353}
2354
2355int set_direct_map_default_noflush(struct page *page)
2356{
2357	return __set_pages_p(page, 1);
2358}
2359
2360#ifdef CONFIG_DEBUG_PAGEALLOC
2361void __kernel_map_pages(struct page *page, int numpages, int enable)
2362{
2363	if (PageHighMem(page))
2364		return;
2365	if (!enable) {
2366		debug_check_no_locks_freed(page_address(page),
2367					   numpages * PAGE_SIZE);
2368	}
2369
2370	/*
2371	 * The return value is ignored as the calls cannot fail.
2372	 * Large pages for identity mappings are not used at boot time
2373	 * and hence no memory allocations during large page split.
2374	 */
2375	if (enable)
2376		__set_pages_p(page, numpages);
2377	else
2378		__set_pages_np(page, numpages);
2379
2380	/*
2381	 * We should perform an IPI and flush all tlbs,
2382	 * but that can deadlock->flush only current cpu.
2383	 * Preemption needs to be disabled around __flush_tlb_all() due to
2384	 * CR3 reload in __native_flush_tlb().
2385	 */
2386	preempt_disable();
2387	__flush_tlb_all();
2388	preempt_enable();
2389
2390	arch_flush_lazy_mmu_mode();
2391}
2392#endif /* CONFIG_DEBUG_PAGEALLOC */
2393
2394bool kernel_page_present(struct page *page)
2395{
2396	unsigned int level;
2397	pte_t *pte;
2398
2399	if (PageHighMem(page))
2400		return false;
2401
2402	pte = lookup_address((unsigned long)page_address(page), &level);
2403	return (pte_val(*pte) & _PAGE_PRESENT);
2404}
2405
2406int __init kernel_map_pages_in_pgd(pgd_t *pgd, u64 pfn, unsigned long address,
2407				   unsigned numpages, unsigned long page_flags)
2408{
2409	int retval = -EINVAL;
2410
2411	struct cpa_data cpa = {
2412		.vaddr = &address,
2413		.pfn = pfn,
2414		.pgd = pgd,
2415		.numpages = numpages,
2416		.mask_set = __pgprot(0),
2417		.mask_clr = __pgprot(~page_flags & (_PAGE_NX|_PAGE_RW)),
2418		.flags = CPA_NO_CHECK_ALIAS,
2419	};
2420
2421	WARN_ONCE(num_online_cpus() > 1, "Don't call after initializing SMP");
2422
2423	if (!(__supported_pte_mask & _PAGE_NX))
2424		goto out;
2425
2426	if (!(page_flags & _PAGE_ENC))
2427		cpa.mask_clr = pgprot_encrypted(cpa.mask_clr);
2428
2429	cpa.mask_set = __pgprot(_PAGE_PRESENT | page_flags);
2430
2431	retval = __change_page_attr_set_clr(&cpa, 1);
2432	__flush_tlb_all();
2433
2434out:
2435	return retval;
2436}
2437
2438/*
2439 * __flush_tlb_all() flushes mappings only on current CPU and hence this
2440 * function shouldn't be used in an SMP environment. Presently, it's used only
2441 * during boot (way before smp_init()) by EFI subsystem and hence is ok.
2442 */
2443int __init kernel_unmap_pages_in_pgd(pgd_t *pgd, unsigned long address,
2444				     unsigned long numpages)
2445{
2446	int retval;
2447
2448	/*
2449	 * The typical sequence for unmapping is to find a pte through
2450	 * lookup_address_in_pgd() (ideally, it should never return NULL because
2451	 * the address is already mapped) and change it's protections. As pfn is
2452	 * the *target* of a mapping, it's not useful while unmapping.
2453	 */
2454	struct cpa_data cpa = {
2455		.vaddr		= &address,
2456		.pfn		= 0,
2457		.pgd		= pgd,
2458		.numpages	= numpages,
2459		.mask_set	= __pgprot(0),
2460		.mask_clr	= __pgprot(_PAGE_PRESENT | _PAGE_RW),
2461		.flags		= CPA_NO_CHECK_ALIAS,
2462	};
2463
2464	WARN_ONCE(num_online_cpus() > 1, "Don't call after initializing SMP");
2465
2466	retval = __change_page_attr_set_clr(&cpa, 1);
2467	__flush_tlb_all();
2468
2469	return retval;
2470}
2471
2472/*
2473 * The testcases use internal knowledge of the implementation that shouldn't
2474 * be exposed to the rest of the kernel. Include these directly here.
2475 */
2476#ifdef CONFIG_CPA_DEBUG
2477#include "cpa-test.c"
2478#endif