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