Loading...
1// SPDX-License-Identifier: GPL-2.0-only
2/*
3 * Re-map IO memory to kernel address space so that we can access it.
4 * This is needed for high PCI addresses that aren't mapped in the
5 * 640k-1MB IO memory area on PC's
6 *
7 * (C) Copyright 1995 1996 Linus Torvalds
8 */
9
10#include <linux/memblock.h>
11#include <linux/init.h>
12#include <linux/io.h>
13#include <linux/ioport.h>
14#include <linux/slab.h>
15#include <linux/vmalloc.h>
16#include <linux/mmiotrace.h>
17#include <linux/mem_encrypt.h>
18#include <linux/efi.h>
19
20#include <asm/set_memory.h>
21#include <asm/e820/api.h>
22#include <asm/efi.h>
23#include <asm/fixmap.h>
24#include <asm/pgtable.h>
25#include <asm/tlbflush.h>
26#include <asm/pgalloc.h>
27#include <asm/pat.h>
28#include <asm/setup.h>
29
30#include "physaddr.h"
31
32/*
33 * Descriptor controlling ioremap() behavior.
34 */
35struct ioremap_desc {
36 unsigned int flags;
37};
38
39/*
40 * Fix up the linear direct mapping of the kernel to avoid cache attribute
41 * conflicts.
42 */
43int ioremap_change_attr(unsigned long vaddr, unsigned long size,
44 enum page_cache_mode pcm)
45{
46 unsigned long nrpages = size >> PAGE_SHIFT;
47 int err;
48
49 switch (pcm) {
50 case _PAGE_CACHE_MODE_UC:
51 default:
52 err = _set_memory_uc(vaddr, nrpages);
53 break;
54 case _PAGE_CACHE_MODE_WC:
55 err = _set_memory_wc(vaddr, nrpages);
56 break;
57 case _PAGE_CACHE_MODE_WT:
58 err = _set_memory_wt(vaddr, nrpages);
59 break;
60 case _PAGE_CACHE_MODE_WB:
61 err = _set_memory_wb(vaddr, nrpages);
62 break;
63 }
64
65 return err;
66}
67
68/* Does the range (or a subset of) contain normal RAM? */
69static unsigned int __ioremap_check_ram(struct resource *res)
70{
71 unsigned long start_pfn, stop_pfn;
72 unsigned long i;
73
74 if ((res->flags & IORESOURCE_SYSTEM_RAM) != IORESOURCE_SYSTEM_RAM)
75 return 0;
76
77 start_pfn = (res->start + PAGE_SIZE - 1) >> PAGE_SHIFT;
78 stop_pfn = (res->end + 1) >> PAGE_SHIFT;
79 if (stop_pfn > start_pfn) {
80 for (i = 0; i < (stop_pfn - start_pfn); ++i)
81 if (pfn_valid(start_pfn + i) &&
82 !PageReserved(pfn_to_page(start_pfn + i)))
83 return IORES_MAP_SYSTEM_RAM;
84 }
85
86 return 0;
87}
88
89/*
90 * In a SEV guest, NONE and RESERVED should not be mapped encrypted because
91 * there the whole memory is already encrypted.
92 */
93static unsigned int __ioremap_check_encrypted(struct resource *res)
94{
95 if (!sev_active())
96 return 0;
97
98 switch (res->desc) {
99 case IORES_DESC_NONE:
100 case IORES_DESC_RESERVED:
101 break;
102 default:
103 return IORES_MAP_ENCRYPTED;
104 }
105
106 return 0;
107}
108
109static int __ioremap_collect_map_flags(struct resource *res, void *arg)
110{
111 struct ioremap_desc *desc = arg;
112
113 if (!(desc->flags & IORES_MAP_SYSTEM_RAM))
114 desc->flags |= __ioremap_check_ram(res);
115
116 if (!(desc->flags & IORES_MAP_ENCRYPTED))
117 desc->flags |= __ioremap_check_encrypted(res);
118
119 return ((desc->flags & (IORES_MAP_SYSTEM_RAM | IORES_MAP_ENCRYPTED)) ==
120 (IORES_MAP_SYSTEM_RAM | IORES_MAP_ENCRYPTED));
121}
122
123/*
124 * To avoid multiple resource walks, this function walks resources marked as
125 * IORESOURCE_MEM and IORESOURCE_BUSY and looking for system RAM and/or a
126 * resource described not as IORES_DESC_NONE (e.g. IORES_DESC_ACPI_TABLES).
127 */
128static void __ioremap_check_mem(resource_size_t addr, unsigned long size,
129 struct ioremap_desc *desc)
130{
131 u64 start, end;
132
133 start = (u64)addr;
134 end = start + size - 1;
135 memset(desc, 0, sizeof(struct ioremap_desc));
136
137 walk_mem_res(start, end, desc, __ioremap_collect_map_flags);
138}
139
140/*
141 * Remap an arbitrary physical address space into the kernel virtual
142 * address space. It transparently creates kernel huge I/O mapping when
143 * the physical address is aligned by a huge page size (1GB or 2MB) and
144 * the requested size is at least the huge page size.
145 *
146 * NOTE: MTRRs can override PAT memory types with a 4KB granularity.
147 * Therefore, the mapping code falls back to use a smaller page toward 4KB
148 * when a mapping range is covered by non-WB type of MTRRs.
149 *
150 * NOTE! We need to allow non-page-aligned mappings too: we will obviously
151 * have to convert them into an offset in a page-aligned mapping, but the
152 * caller shouldn't need to know that small detail.
153 */
154static void __iomem *
155__ioremap_caller(resource_size_t phys_addr, unsigned long size,
156 enum page_cache_mode pcm, void *caller, bool encrypted)
157{
158 unsigned long offset, vaddr;
159 resource_size_t last_addr;
160 const resource_size_t unaligned_phys_addr = phys_addr;
161 const unsigned long unaligned_size = size;
162 struct ioremap_desc io_desc;
163 struct vm_struct *area;
164 enum page_cache_mode new_pcm;
165 pgprot_t prot;
166 int retval;
167 void __iomem *ret_addr;
168
169 /* Don't allow wraparound or zero size */
170 last_addr = phys_addr + size - 1;
171 if (!size || last_addr < phys_addr)
172 return NULL;
173
174 if (!phys_addr_valid(phys_addr)) {
175 printk(KERN_WARNING "ioremap: invalid physical address %llx\n",
176 (unsigned long long)phys_addr);
177 WARN_ON_ONCE(1);
178 return NULL;
179 }
180
181 __ioremap_check_mem(phys_addr, size, &io_desc);
182
183 /*
184 * Don't allow anybody to remap normal RAM that we're using..
185 */
186 if (io_desc.flags & IORES_MAP_SYSTEM_RAM) {
187 WARN_ONCE(1, "ioremap on RAM at %pa - %pa\n",
188 &phys_addr, &last_addr);
189 return NULL;
190 }
191
192 /*
193 * Mappings have to be page-aligned
194 */
195 offset = phys_addr & ~PAGE_MASK;
196 phys_addr &= PHYSICAL_PAGE_MASK;
197 size = PAGE_ALIGN(last_addr+1) - phys_addr;
198
199 retval = reserve_memtype(phys_addr, (u64)phys_addr + size,
200 pcm, &new_pcm);
201 if (retval) {
202 printk(KERN_ERR "ioremap reserve_memtype failed %d\n", retval);
203 return NULL;
204 }
205
206 if (pcm != new_pcm) {
207 if (!is_new_memtype_allowed(phys_addr, size, pcm, new_pcm)) {
208 printk(KERN_ERR
209 "ioremap error for 0x%llx-0x%llx, requested 0x%x, got 0x%x\n",
210 (unsigned long long)phys_addr,
211 (unsigned long long)(phys_addr + size),
212 pcm, new_pcm);
213 goto err_free_memtype;
214 }
215 pcm = new_pcm;
216 }
217
218 /*
219 * If the page being mapped is in memory and SEV is active then
220 * make sure the memory encryption attribute is enabled in the
221 * resulting mapping.
222 */
223 prot = PAGE_KERNEL_IO;
224 if ((io_desc.flags & IORES_MAP_ENCRYPTED) || encrypted)
225 prot = pgprot_encrypted(prot);
226
227 switch (pcm) {
228 case _PAGE_CACHE_MODE_UC:
229 default:
230 prot = __pgprot(pgprot_val(prot) |
231 cachemode2protval(_PAGE_CACHE_MODE_UC));
232 break;
233 case _PAGE_CACHE_MODE_UC_MINUS:
234 prot = __pgprot(pgprot_val(prot) |
235 cachemode2protval(_PAGE_CACHE_MODE_UC_MINUS));
236 break;
237 case _PAGE_CACHE_MODE_WC:
238 prot = __pgprot(pgprot_val(prot) |
239 cachemode2protval(_PAGE_CACHE_MODE_WC));
240 break;
241 case _PAGE_CACHE_MODE_WT:
242 prot = __pgprot(pgprot_val(prot) |
243 cachemode2protval(_PAGE_CACHE_MODE_WT));
244 break;
245 case _PAGE_CACHE_MODE_WB:
246 break;
247 }
248
249 /*
250 * Ok, go for it..
251 */
252 area = get_vm_area_caller(size, VM_IOREMAP, caller);
253 if (!area)
254 goto err_free_memtype;
255 area->phys_addr = phys_addr;
256 vaddr = (unsigned long) area->addr;
257
258 if (kernel_map_sync_memtype(phys_addr, size, pcm))
259 goto err_free_area;
260
261 if (ioremap_page_range(vaddr, vaddr + size, phys_addr, prot))
262 goto err_free_area;
263
264 ret_addr = (void __iomem *) (vaddr + offset);
265 mmiotrace_ioremap(unaligned_phys_addr, unaligned_size, ret_addr);
266
267 /*
268 * Check if the request spans more than any BAR in the iomem resource
269 * tree.
270 */
271 if (iomem_map_sanity_check(unaligned_phys_addr, unaligned_size))
272 pr_warn("caller %pS mapping multiple BARs\n", caller);
273
274 return ret_addr;
275err_free_area:
276 free_vm_area(area);
277err_free_memtype:
278 free_memtype(phys_addr, phys_addr + size);
279 return NULL;
280}
281
282/**
283 * ioremap_nocache - map bus memory into CPU space
284 * @phys_addr: bus address of the memory
285 * @size: size of the resource to map
286 *
287 * ioremap_nocache performs a platform specific sequence of operations to
288 * make bus memory CPU accessible via the readb/readw/readl/writeb/
289 * writew/writel functions and the other mmio helpers. The returned
290 * address is not guaranteed to be usable directly as a virtual
291 * address.
292 *
293 * This version of ioremap ensures that the memory is marked uncachable
294 * on the CPU as well as honouring existing caching rules from things like
295 * the PCI bus. Note that there are other caches and buffers on many
296 * busses. In particular driver authors should read up on PCI writes
297 *
298 * It's useful if some control registers are in such an area and
299 * write combining or read caching is not desirable:
300 *
301 * Must be freed with iounmap.
302 */
303void __iomem *ioremap_nocache(resource_size_t phys_addr, unsigned long size)
304{
305 /*
306 * Ideally, this should be:
307 * pat_enabled() ? _PAGE_CACHE_MODE_UC : _PAGE_CACHE_MODE_UC_MINUS;
308 *
309 * Till we fix all X drivers to use ioremap_wc(), we will use
310 * UC MINUS. Drivers that are certain they need or can already
311 * be converted over to strong UC can use ioremap_uc().
312 */
313 enum page_cache_mode pcm = _PAGE_CACHE_MODE_UC_MINUS;
314
315 return __ioremap_caller(phys_addr, size, pcm,
316 __builtin_return_address(0), false);
317}
318EXPORT_SYMBOL(ioremap_nocache);
319
320/**
321 * ioremap_uc - map bus memory into CPU space as strongly uncachable
322 * @phys_addr: bus address of the memory
323 * @size: size of the resource to map
324 *
325 * ioremap_uc performs a platform specific sequence of operations to
326 * make bus memory CPU accessible via the readb/readw/readl/writeb/
327 * writew/writel functions and the other mmio helpers. The returned
328 * address is not guaranteed to be usable directly as a virtual
329 * address.
330 *
331 * This version of ioremap ensures that the memory is marked with a strong
332 * preference as completely uncachable on the CPU when possible. For non-PAT
333 * systems this ends up setting page-attribute flags PCD=1, PWT=1. For PAT
334 * systems this will set the PAT entry for the pages as strong UC. This call
335 * will honor existing caching rules from things like the PCI bus. Note that
336 * there are other caches and buffers on many busses. In particular driver
337 * authors should read up on PCI writes.
338 *
339 * It's useful if some control registers are in such an area and
340 * write combining or read caching is not desirable:
341 *
342 * Must be freed with iounmap.
343 */
344void __iomem *ioremap_uc(resource_size_t phys_addr, unsigned long size)
345{
346 enum page_cache_mode pcm = _PAGE_CACHE_MODE_UC;
347
348 return __ioremap_caller(phys_addr, size, pcm,
349 __builtin_return_address(0), false);
350}
351EXPORT_SYMBOL_GPL(ioremap_uc);
352
353/**
354 * ioremap_wc - map memory into CPU space write combined
355 * @phys_addr: bus address of the memory
356 * @size: size of the resource to map
357 *
358 * This version of ioremap ensures that the memory is marked write combining.
359 * Write combining allows faster writes to some hardware devices.
360 *
361 * Must be freed with iounmap.
362 */
363void __iomem *ioremap_wc(resource_size_t phys_addr, unsigned long size)
364{
365 return __ioremap_caller(phys_addr, size, _PAGE_CACHE_MODE_WC,
366 __builtin_return_address(0), false);
367}
368EXPORT_SYMBOL(ioremap_wc);
369
370/**
371 * ioremap_wt - map memory into CPU space write through
372 * @phys_addr: bus address of the memory
373 * @size: size of the resource to map
374 *
375 * This version of ioremap ensures that the memory is marked write through.
376 * Write through stores data into memory while keeping the cache up-to-date.
377 *
378 * Must be freed with iounmap.
379 */
380void __iomem *ioremap_wt(resource_size_t phys_addr, unsigned long size)
381{
382 return __ioremap_caller(phys_addr, size, _PAGE_CACHE_MODE_WT,
383 __builtin_return_address(0), false);
384}
385EXPORT_SYMBOL(ioremap_wt);
386
387void __iomem *ioremap_encrypted(resource_size_t phys_addr, unsigned long size)
388{
389 return __ioremap_caller(phys_addr, size, _PAGE_CACHE_MODE_WB,
390 __builtin_return_address(0), true);
391}
392EXPORT_SYMBOL(ioremap_encrypted);
393
394void __iomem *ioremap_cache(resource_size_t phys_addr, unsigned long size)
395{
396 return __ioremap_caller(phys_addr, size, _PAGE_CACHE_MODE_WB,
397 __builtin_return_address(0), false);
398}
399EXPORT_SYMBOL(ioremap_cache);
400
401void __iomem *ioremap_prot(resource_size_t phys_addr, unsigned long size,
402 unsigned long prot_val)
403{
404 return __ioremap_caller(phys_addr, size,
405 pgprot2cachemode(__pgprot(prot_val)),
406 __builtin_return_address(0), false);
407}
408EXPORT_SYMBOL(ioremap_prot);
409
410/**
411 * iounmap - Free a IO remapping
412 * @addr: virtual address from ioremap_*
413 *
414 * Caller must ensure there is only one unmapping for the same pointer.
415 */
416void iounmap(volatile void __iomem *addr)
417{
418 struct vm_struct *p, *o;
419
420 if ((void __force *)addr <= high_memory)
421 return;
422
423 /*
424 * The PCI/ISA range special-casing was removed from __ioremap()
425 * so this check, in theory, can be removed. However, there are
426 * cases where iounmap() is called for addresses not obtained via
427 * ioremap() (vga16fb for example). Add a warning so that these
428 * cases can be caught and fixed.
429 */
430 if ((void __force *)addr >= phys_to_virt(ISA_START_ADDRESS) &&
431 (void __force *)addr < phys_to_virt(ISA_END_ADDRESS)) {
432 WARN(1, "iounmap() called for ISA range not obtained using ioremap()\n");
433 return;
434 }
435
436 mmiotrace_iounmap(addr);
437
438 addr = (volatile void __iomem *)
439 (PAGE_MASK & (unsigned long __force)addr);
440
441 /* Use the vm area unlocked, assuming the caller
442 ensures there isn't another iounmap for the same address
443 in parallel. Reuse of the virtual address is prevented by
444 leaving it in the global lists until we're done with it.
445 cpa takes care of the direct mappings. */
446 p = find_vm_area((void __force *)addr);
447
448 if (!p) {
449 printk(KERN_ERR "iounmap: bad address %p\n", addr);
450 dump_stack();
451 return;
452 }
453
454 free_memtype(p->phys_addr, p->phys_addr + get_vm_area_size(p));
455
456 /* Finally remove it */
457 o = remove_vm_area((void __force *)addr);
458 BUG_ON(p != o || o == NULL);
459 kfree(p);
460}
461EXPORT_SYMBOL(iounmap);
462
463int __init arch_ioremap_p4d_supported(void)
464{
465 return 0;
466}
467
468int __init arch_ioremap_pud_supported(void)
469{
470#ifdef CONFIG_X86_64
471 return boot_cpu_has(X86_FEATURE_GBPAGES);
472#else
473 return 0;
474#endif
475}
476
477int __init arch_ioremap_pmd_supported(void)
478{
479 return boot_cpu_has(X86_FEATURE_PSE);
480}
481
482/*
483 * Convert a physical pointer to a virtual kernel pointer for /dev/mem
484 * access
485 */
486void *xlate_dev_mem_ptr(phys_addr_t phys)
487{
488 unsigned long start = phys & PAGE_MASK;
489 unsigned long offset = phys & ~PAGE_MASK;
490 void *vaddr;
491
492 /* memremap() maps if RAM, otherwise falls back to ioremap() */
493 vaddr = memremap(start, PAGE_SIZE, MEMREMAP_WB);
494
495 /* Only add the offset on success and return NULL if memremap() failed */
496 if (vaddr)
497 vaddr += offset;
498
499 return vaddr;
500}
501
502void unxlate_dev_mem_ptr(phys_addr_t phys, void *addr)
503{
504 memunmap((void *)((unsigned long)addr & PAGE_MASK));
505}
506
507/*
508 * Examine the physical address to determine if it is an area of memory
509 * that should be mapped decrypted. If the memory is not part of the
510 * kernel usable area it was accessed and created decrypted, so these
511 * areas should be mapped decrypted. And since the encryption key can
512 * change across reboots, persistent memory should also be mapped
513 * decrypted.
514 *
515 * If SEV is active, that implies that BIOS/UEFI also ran encrypted so
516 * only persistent memory should be mapped decrypted.
517 */
518static bool memremap_should_map_decrypted(resource_size_t phys_addr,
519 unsigned long size)
520{
521 int is_pmem;
522
523 /*
524 * Check if the address is part of a persistent memory region.
525 * This check covers areas added by E820, EFI and ACPI.
526 */
527 is_pmem = region_intersects(phys_addr, size, IORESOURCE_MEM,
528 IORES_DESC_PERSISTENT_MEMORY);
529 if (is_pmem != REGION_DISJOINT)
530 return true;
531
532 /*
533 * Check if the non-volatile attribute is set for an EFI
534 * reserved area.
535 */
536 if (efi_enabled(EFI_BOOT)) {
537 switch (efi_mem_type(phys_addr)) {
538 case EFI_RESERVED_TYPE:
539 if (efi_mem_attributes(phys_addr) & EFI_MEMORY_NV)
540 return true;
541 break;
542 default:
543 break;
544 }
545 }
546
547 /* Check if the address is outside kernel usable area */
548 switch (e820__get_entry_type(phys_addr, phys_addr + size - 1)) {
549 case E820_TYPE_RESERVED:
550 case E820_TYPE_ACPI:
551 case E820_TYPE_NVS:
552 case E820_TYPE_UNUSABLE:
553 /* For SEV, these areas are encrypted */
554 if (sev_active())
555 break;
556 /* Fallthrough */
557
558 case E820_TYPE_PRAM:
559 return true;
560 default:
561 break;
562 }
563
564 return false;
565}
566
567/*
568 * Examine the physical address to determine if it is EFI data. Check
569 * it against the boot params structure and EFI tables and memory types.
570 */
571static bool memremap_is_efi_data(resource_size_t phys_addr,
572 unsigned long size)
573{
574 u64 paddr;
575
576 /* Check if the address is part of EFI boot/runtime data */
577 if (!efi_enabled(EFI_BOOT))
578 return false;
579
580 paddr = boot_params.efi_info.efi_memmap_hi;
581 paddr <<= 32;
582 paddr |= boot_params.efi_info.efi_memmap;
583 if (phys_addr == paddr)
584 return true;
585
586 paddr = boot_params.efi_info.efi_systab_hi;
587 paddr <<= 32;
588 paddr |= boot_params.efi_info.efi_systab;
589 if (phys_addr == paddr)
590 return true;
591
592 if (efi_is_table_address(phys_addr))
593 return true;
594
595 switch (efi_mem_type(phys_addr)) {
596 case EFI_BOOT_SERVICES_DATA:
597 case EFI_RUNTIME_SERVICES_DATA:
598 return true;
599 default:
600 break;
601 }
602
603 return false;
604}
605
606/*
607 * Examine the physical address to determine if it is boot data by checking
608 * it against the boot params setup_data chain.
609 */
610static bool memremap_is_setup_data(resource_size_t phys_addr,
611 unsigned long size)
612{
613 struct setup_data *data;
614 u64 paddr, paddr_next;
615
616 paddr = boot_params.hdr.setup_data;
617 while (paddr) {
618 unsigned int len;
619
620 if (phys_addr == paddr)
621 return true;
622
623 data = memremap(paddr, sizeof(*data),
624 MEMREMAP_WB | MEMREMAP_DEC);
625
626 paddr_next = data->next;
627 len = data->len;
628
629 memunmap(data);
630
631 if ((phys_addr > paddr) && (phys_addr < (paddr + len)))
632 return true;
633
634 paddr = paddr_next;
635 }
636
637 return false;
638}
639
640/*
641 * Examine the physical address to determine if it is boot data by checking
642 * it against the boot params setup_data chain (early boot version).
643 */
644static bool __init early_memremap_is_setup_data(resource_size_t phys_addr,
645 unsigned long size)
646{
647 struct setup_data *data;
648 u64 paddr, paddr_next;
649
650 paddr = boot_params.hdr.setup_data;
651 while (paddr) {
652 unsigned int len;
653
654 if (phys_addr == paddr)
655 return true;
656
657 data = early_memremap_decrypted(paddr, sizeof(*data));
658
659 paddr_next = data->next;
660 len = data->len;
661
662 early_memunmap(data, sizeof(*data));
663
664 if ((phys_addr > paddr) && (phys_addr < (paddr + len)))
665 return true;
666
667 paddr = paddr_next;
668 }
669
670 return false;
671}
672
673/*
674 * Architecture function to determine if RAM remap is allowed. By default, a
675 * RAM remap will map the data as encrypted. Determine if a RAM remap should
676 * not be done so that the data will be mapped decrypted.
677 */
678bool arch_memremap_can_ram_remap(resource_size_t phys_addr, unsigned long size,
679 unsigned long flags)
680{
681 if (!mem_encrypt_active())
682 return true;
683
684 if (flags & MEMREMAP_ENC)
685 return true;
686
687 if (flags & MEMREMAP_DEC)
688 return false;
689
690 if (sme_active()) {
691 if (memremap_is_setup_data(phys_addr, size) ||
692 memremap_is_efi_data(phys_addr, size))
693 return false;
694 }
695
696 return !memremap_should_map_decrypted(phys_addr, size);
697}
698
699/*
700 * Architecture override of __weak function to adjust the protection attributes
701 * used when remapping memory. By default, early_memremap() will map the data
702 * as encrypted. Determine if an encrypted mapping should not be done and set
703 * the appropriate protection attributes.
704 */
705pgprot_t __init early_memremap_pgprot_adjust(resource_size_t phys_addr,
706 unsigned long size,
707 pgprot_t prot)
708{
709 bool encrypted_prot;
710
711 if (!mem_encrypt_active())
712 return prot;
713
714 encrypted_prot = true;
715
716 if (sme_active()) {
717 if (early_memremap_is_setup_data(phys_addr, size) ||
718 memremap_is_efi_data(phys_addr, size))
719 encrypted_prot = false;
720 }
721
722 if (encrypted_prot && memremap_should_map_decrypted(phys_addr, size))
723 encrypted_prot = false;
724
725 return encrypted_prot ? pgprot_encrypted(prot)
726 : pgprot_decrypted(prot);
727}
728
729bool phys_mem_access_encrypted(unsigned long phys_addr, unsigned long size)
730{
731 return arch_memremap_can_ram_remap(phys_addr, size, 0);
732}
733
734#ifdef CONFIG_AMD_MEM_ENCRYPT
735/* Remap memory with encryption */
736void __init *early_memremap_encrypted(resource_size_t phys_addr,
737 unsigned long size)
738{
739 return early_memremap_prot(phys_addr, size, __PAGE_KERNEL_ENC);
740}
741
742/*
743 * Remap memory with encryption and write-protected - cannot be called
744 * before pat_init() is called
745 */
746void __init *early_memremap_encrypted_wp(resource_size_t phys_addr,
747 unsigned long size)
748{
749 /* Be sure the write-protect PAT entry is set for write-protect */
750 if (__pte2cachemode_tbl[_PAGE_CACHE_MODE_WP] != _PAGE_CACHE_MODE_WP)
751 return NULL;
752
753 return early_memremap_prot(phys_addr, size, __PAGE_KERNEL_ENC_WP);
754}
755
756/* Remap memory without encryption */
757void __init *early_memremap_decrypted(resource_size_t phys_addr,
758 unsigned long size)
759{
760 return early_memremap_prot(phys_addr, size, __PAGE_KERNEL_NOENC);
761}
762
763/*
764 * Remap memory without encryption and write-protected - cannot be called
765 * before pat_init() is called
766 */
767void __init *early_memremap_decrypted_wp(resource_size_t phys_addr,
768 unsigned long size)
769{
770 /* Be sure the write-protect PAT entry is set for write-protect */
771 if (__pte2cachemode_tbl[_PAGE_CACHE_MODE_WP] != _PAGE_CACHE_MODE_WP)
772 return NULL;
773
774 return early_memremap_prot(phys_addr, size, __PAGE_KERNEL_NOENC_WP);
775}
776#endif /* CONFIG_AMD_MEM_ENCRYPT */
777
778static pte_t bm_pte[PAGE_SIZE/sizeof(pte_t)] __page_aligned_bss;
779
780static inline pmd_t * __init early_ioremap_pmd(unsigned long addr)
781{
782 /* Don't assume we're using swapper_pg_dir at this point */
783 pgd_t *base = __va(read_cr3_pa());
784 pgd_t *pgd = &base[pgd_index(addr)];
785 p4d_t *p4d = p4d_offset(pgd, addr);
786 pud_t *pud = pud_offset(p4d, addr);
787 pmd_t *pmd = pmd_offset(pud, addr);
788
789 return pmd;
790}
791
792static inline pte_t * __init early_ioremap_pte(unsigned long addr)
793{
794 return &bm_pte[pte_index(addr)];
795}
796
797bool __init is_early_ioremap_ptep(pte_t *ptep)
798{
799 return ptep >= &bm_pte[0] && ptep < &bm_pte[PAGE_SIZE/sizeof(pte_t)];
800}
801
802void __init early_ioremap_init(void)
803{
804 pmd_t *pmd;
805
806#ifdef CONFIG_X86_64
807 BUILD_BUG_ON((fix_to_virt(0) + PAGE_SIZE) & ((1 << PMD_SHIFT) - 1));
808#else
809 WARN_ON((fix_to_virt(0) + PAGE_SIZE) & ((1 << PMD_SHIFT) - 1));
810#endif
811
812 early_ioremap_setup();
813
814 pmd = early_ioremap_pmd(fix_to_virt(FIX_BTMAP_BEGIN));
815 memset(bm_pte, 0, sizeof(bm_pte));
816 pmd_populate_kernel(&init_mm, pmd, bm_pte);
817
818 /*
819 * The boot-ioremap range spans multiple pmds, for which
820 * we are not prepared:
821 */
822#define __FIXADDR_TOP (-PAGE_SIZE)
823 BUILD_BUG_ON((__fix_to_virt(FIX_BTMAP_BEGIN) >> PMD_SHIFT)
824 != (__fix_to_virt(FIX_BTMAP_END) >> PMD_SHIFT));
825#undef __FIXADDR_TOP
826 if (pmd != early_ioremap_pmd(fix_to_virt(FIX_BTMAP_END))) {
827 WARN_ON(1);
828 printk(KERN_WARNING "pmd %p != %p\n",
829 pmd, early_ioremap_pmd(fix_to_virt(FIX_BTMAP_END)));
830 printk(KERN_WARNING "fix_to_virt(FIX_BTMAP_BEGIN): %08lx\n",
831 fix_to_virt(FIX_BTMAP_BEGIN));
832 printk(KERN_WARNING "fix_to_virt(FIX_BTMAP_END): %08lx\n",
833 fix_to_virt(FIX_BTMAP_END));
834
835 printk(KERN_WARNING "FIX_BTMAP_END: %d\n", FIX_BTMAP_END);
836 printk(KERN_WARNING "FIX_BTMAP_BEGIN: %d\n",
837 FIX_BTMAP_BEGIN);
838 }
839}
840
841void __init __early_set_fixmap(enum fixed_addresses idx,
842 phys_addr_t phys, pgprot_t flags)
843{
844 unsigned long addr = __fix_to_virt(idx);
845 pte_t *pte;
846
847 if (idx >= __end_of_fixed_addresses) {
848 BUG();
849 return;
850 }
851 pte = early_ioremap_pte(addr);
852
853 /* Sanitize 'prot' against any unsupported bits: */
854 pgprot_val(flags) &= __supported_pte_mask;
855
856 if (pgprot_val(flags))
857 set_pte(pte, pfn_pte(phys >> PAGE_SHIFT, flags));
858 else
859 pte_clear(&init_mm, addr, pte);
860 __flush_tlb_one_kernel(addr);
861}
1// SPDX-License-Identifier: GPL-2.0-only
2/*
3 * Re-map IO memory to kernel address space so that we can access it.
4 * This is needed for high PCI addresses that aren't mapped in the
5 * 640k-1MB IO memory area on PC's
6 *
7 * (C) Copyright 1995 1996 Linus Torvalds
8 */
9
10#include <linux/memblock.h>
11#include <linux/init.h>
12#include <linux/io.h>
13#include <linux/ioport.h>
14#include <linux/slab.h>
15#include <linux/vmalloc.h>
16#include <linux/mmiotrace.h>
17#include <linux/cc_platform.h>
18#include <linux/efi.h>
19#include <linux/pgtable.h>
20#include <linux/kmsan.h>
21
22#include <asm/set_memory.h>
23#include <asm/e820/api.h>
24#include <asm/efi.h>
25#include <asm/fixmap.h>
26#include <asm/tlbflush.h>
27#include <asm/pgalloc.h>
28#include <asm/memtype.h>
29#include <asm/setup.h>
30
31#include "physaddr.h"
32
33/*
34 * Descriptor controlling ioremap() behavior.
35 */
36struct ioremap_desc {
37 unsigned int flags;
38};
39
40/*
41 * Fix up the linear direct mapping of the kernel to avoid cache attribute
42 * conflicts.
43 */
44int ioremap_change_attr(unsigned long vaddr, unsigned long size,
45 enum page_cache_mode pcm)
46{
47 unsigned long nrpages = size >> PAGE_SHIFT;
48 int err;
49
50 switch (pcm) {
51 case _PAGE_CACHE_MODE_UC:
52 default:
53 err = _set_memory_uc(vaddr, nrpages);
54 break;
55 case _PAGE_CACHE_MODE_WC:
56 err = _set_memory_wc(vaddr, nrpages);
57 break;
58 case _PAGE_CACHE_MODE_WT:
59 err = _set_memory_wt(vaddr, nrpages);
60 break;
61 case _PAGE_CACHE_MODE_WB:
62 err = _set_memory_wb(vaddr, nrpages);
63 break;
64 }
65
66 return err;
67}
68
69/* Does the range (or a subset of) contain normal RAM? */
70static unsigned int __ioremap_check_ram(struct resource *res)
71{
72 unsigned long start_pfn, stop_pfn;
73 unsigned long i;
74
75 if ((res->flags & IORESOURCE_SYSTEM_RAM) != IORESOURCE_SYSTEM_RAM)
76 return 0;
77
78 start_pfn = (res->start + PAGE_SIZE - 1) >> PAGE_SHIFT;
79 stop_pfn = (res->end + 1) >> PAGE_SHIFT;
80 if (stop_pfn > start_pfn) {
81 for (i = 0; i < (stop_pfn - start_pfn); ++i)
82 if (pfn_valid(start_pfn + i) &&
83 !PageReserved(pfn_to_page(start_pfn + i)))
84 return IORES_MAP_SYSTEM_RAM;
85 }
86
87 return 0;
88}
89
90/*
91 * In a SEV guest, NONE and RESERVED should not be mapped encrypted because
92 * there the whole memory is already encrypted.
93 */
94static unsigned int __ioremap_check_encrypted(struct resource *res)
95{
96 if (!cc_platform_has(CC_ATTR_GUEST_MEM_ENCRYPT))
97 return 0;
98
99 switch (res->desc) {
100 case IORES_DESC_NONE:
101 case IORES_DESC_RESERVED:
102 break;
103 default:
104 return IORES_MAP_ENCRYPTED;
105 }
106
107 return 0;
108}
109
110/*
111 * The EFI runtime services data area is not covered by walk_mem_res(), but must
112 * be mapped encrypted when SEV is active.
113 */
114static void __ioremap_check_other(resource_size_t addr, struct ioremap_desc *desc)
115{
116 if (!cc_platform_has(CC_ATTR_GUEST_MEM_ENCRYPT))
117 return;
118
119 if (x86_platform.hyper.is_private_mmio(addr)) {
120 desc->flags |= IORES_MAP_ENCRYPTED;
121 return;
122 }
123
124 if (!IS_ENABLED(CONFIG_EFI))
125 return;
126
127 if (efi_mem_type(addr) == EFI_RUNTIME_SERVICES_DATA ||
128 (efi_mem_type(addr) == EFI_BOOT_SERVICES_DATA &&
129 efi_mem_attributes(addr) & EFI_MEMORY_RUNTIME))
130 desc->flags |= IORES_MAP_ENCRYPTED;
131}
132
133static int __ioremap_collect_map_flags(struct resource *res, void *arg)
134{
135 struct ioremap_desc *desc = arg;
136
137 if (!(desc->flags & IORES_MAP_SYSTEM_RAM))
138 desc->flags |= __ioremap_check_ram(res);
139
140 if (!(desc->flags & IORES_MAP_ENCRYPTED))
141 desc->flags |= __ioremap_check_encrypted(res);
142
143 return ((desc->flags & (IORES_MAP_SYSTEM_RAM | IORES_MAP_ENCRYPTED)) ==
144 (IORES_MAP_SYSTEM_RAM | IORES_MAP_ENCRYPTED));
145}
146
147/*
148 * To avoid multiple resource walks, this function walks resources marked as
149 * IORESOURCE_MEM and IORESOURCE_BUSY and looking for system RAM and/or a
150 * resource described not as IORES_DESC_NONE (e.g. IORES_DESC_ACPI_TABLES).
151 *
152 * After that, deal with misc other ranges in __ioremap_check_other() which do
153 * not fall into the above category.
154 */
155static void __ioremap_check_mem(resource_size_t addr, unsigned long size,
156 struct ioremap_desc *desc)
157{
158 u64 start, end;
159
160 start = (u64)addr;
161 end = start + size - 1;
162 memset(desc, 0, sizeof(struct ioremap_desc));
163
164 walk_mem_res(start, end, desc, __ioremap_collect_map_flags);
165
166 __ioremap_check_other(addr, desc);
167}
168
169/*
170 * Remap an arbitrary physical address space into the kernel virtual
171 * address space. It transparently creates kernel huge I/O mapping when
172 * the physical address is aligned by a huge page size (1GB or 2MB) and
173 * the requested size is at least the huge page size.
174 *
175 * NOTE: MTRRs can override PAT memory types with a 4KB granularity.
176 * Therefore, the mapping code falls back to use a smaller page toward 4KB
177 * when a mapping range is covered by non-WB type of MTRRs.
178 *
179 * NOTE! We need to allow non-page-aligned mappings too: we will obviously
180 * have to convert them into an offset in a page-aligned mapping, but the
181 * caller shouldn't need to know that small detail.
182 */
183static void __iomem *
184__ioremap_caller(resource_size_t phys_addr, unsigned long size,
185 enum page_cache_mode pcm, void *caller, bool encrypted)
186{
187 unsigned long offset, vaddr;
188 resource_size_t last_addr;
189 const resource_size_t unaligned_phys_addr = phys_addr;
190 const unsigned long unaligned_size = size;
191 struct ioremap_desc io_desc;
192 struct vm_struct *area;
193 enum page_cache_mode new_pcm;
194 pgprot_t prot;
195 int retval;
196 void __iomem *ret_addr;
197
198 /* Don't allow wraparound or zero size */
199 last_addr = phys_addr + size - 1;
200 if (!size || last_addr < phys_addr)
201 return NULL;
202
203 if (!phys_addr_valid(phys_addr)) {
204 printk(KERN_WARNING "ioremap: invalid physical address %llx\n",
205 (unsigned long long)phys_addr);
206 WARN_ON_ONCE(1);
207 return NULL;
208 }
209
210 __ioremap_check_mem(phys_addr, size, &io_desc);
211
212 /*
213 * Don't allow anybody to remap normal RAM that we're using..
214 */
215 if (io_desc.flags & IORES_MAP_SYSTEM_RAM) {
216 WARN_ONCE(1, "ioremap on RAM at %pa - %pa\n",
217 &phys_addr, &last_addr);
218 return NULL;
219 }
220
221 /*
222 * Mappings have to be page-aligned
223 */
224 offset = phys_addr & ~PAGE_MASK;
225 phys_addr &= PAGE_MASK;
226 size = PAGE_ALIGN(last_addr+1) - phys_addr;
227
228 /*
229 * Mask out any bits not part of the actual physical
230 * address, like memory encryption bits.
231 */
232 phys_addr &= PHYSICAL_PAGE_MASK;
233
234 retval = memtype_reserve(phys_addr, (u64)phys_addr + size,
235 pcm, &new_pcm);
236 if (retval) {
237 printk(KERN_ERR "ioremap memtype_reserve failed %d\n", retval);
238 return NULL;
239 }
240
241 if (pcm != new_pcm) {
242 if (!is_new_memtype_allowed(phys_addr, size, pcm, new_pcm)) {
243 printk(KERN_ERR
244 "ioremap error for 0x%llx-0x%llx, requested 0x%x, got 0x%x\n",
245 (unsigned long long)phys_addr,
246 (unsigned long long)(phys_addr + size),
247 pcm, new_pcm);
248 goto err_free_memtype;
249 }
250 pcm = new_pcm;
251 }
252
253 /*
254 * If the page being mapped is in memory and SEV is active then
255 * make sure the memory encryption attribute is enabled in the
256 * resulting mapping.
257 * In TDX guests, memory is marked private by default. If encryption
258 * is not requested (using encrypted), explicitly set decrypt
259 * attribute in all IOREMAPPED memory.
260 */
261 prot = PAGE_KERNEL_IO;
262 if ((io_desc.flags & IORES_MAP_ENCRYPTED) || encrypted)
263 prot = pgprot_encrypted(prot);
264 else
265 prot = pgprot_decrypted(prot);
266
267 switch (pcm) {
268 case _PAGE_CACHE_MODE_UC:
269 default:
270 prot = __pgprot(pgprot_val(prot) |
271 cachemode2protval(_PAGE_CACHE_MODE_UC));
272 break;
273 case _PAGE_CACHE_MODE_UC_MINUS:
274 prot = __pgprot(pgprot_val(prot) |
275 cachemode2protval(_PAGE_CACHE_MODE_UC_MINUS));
276 break;
277 case _PAGE_CACHE_MODE_WC:
278 prot = __pgprot(pgprot_val(prot) |
279 cachemode2protval(_PAGE_CACHE_MODE_WC));
280 break;
281 case _PAGE_CACHE_MODE_WT:
282 prot = __pgprot(pgprot_val(prot) |
283 cachemode2protval(_PAGE_CACHE_MODE_WT));
284 break;
285 case _PAGE_CACHE_MODE_WB:
286 break;
287 }
288
289 /*
290 * Ok, go for it..
291 */
292 area = get_vm_area_caller(size, VM_IOREMAP, caller);
293 if (!area)
294 goto err_free_memtype;
295 area->phys_addr = phys_addr;
296 vaddr = (unsigned long) area->addr;
297
298 if (memtype_kernel_map_sync(phys_addr, size, pcm))
299 goto err_free_area;
300
301 if (ioremap_page_range(vaddr, vaddr + size, phys_addr, prot))
302 goto err_free_area;
303
304 ret_addr = (void __iomem *) (vaddr + offset);
305 mmiotrace_ioremap(unaligned_phys_addr, unaligned_size, ret_addr);
306
307 /*
308 * Check if the request spans more than any BAR in the iomem resource
309 * tree.
310 */
311 if (iomem_map_sanity_check(unaligned_phys_addr, unaligned_size))
312 pr_warn("caller %pS mapping multiple BARs\n", caller);
313
314 return ret_addr;
315err_free_area:
316 free_vm_area(area);
317err_free_memtype:
318 memtype_free(phys_addr, phys_addr + size);
319 return NULL;
320}
321
322/**
323 * ioremap - map bus memory into CPU space
324 * @phys_addr: bus address of the memory
325 * @size: size of the resource to map
326 *
327 * ioremap performs a platform specific sequence of operations to
328 * make bus memory CPU accessible via the readb/readw/readl/writeb/
329 * writew/writel functions and the other mmio helpers. The returned
330 * address is not guaranteed to be usable directly as a virtual
331 * address.
332 *
333 * This version of ioremap ensures that the memory is marked uncachable
334 * on the CPU as well as honouring existing caching rules from things like
335 * the PCI bus. Note that there are other caches and buffers on many
336 * busses. In particular driver authors should read up on PCI writes
337 *
338 * It's useful if some control registers are in such an area and
339 * write combining or read caching is not desirable:
340 *
341 * Must be freed with iounmap.
342 */
343void __iomem *ioremap(resource_size_t phys_addr, unsigned long size)
344{
345 /*
346 * Ideally, this should be:
347 * pat_enabled() ? _PAGE_CACHE_MODE_UC : _PAGE_CACHE_MODE_UC_MINUS;
348 *
349 * Till we fix all X drivers to use ioremap_wc(), we will use
350 * UC MINUS. Drivers that are certain they need or can already
351 * be converted over to strong UC can use ioremap_uc().
352 */
353 enum page_cache_mode pcm = _PAGE_CACHE_MODE_UC_MINUS;
354
355 return __ioremap_caller(phys_addr, size, pcm,
356 __builtin_return_address(0), false);
357}
358EXPORT_SYMBOL(ioremap);
359
360/**
361 * ioremap_uc - map bus memory into CPU space as strongly uncachable
362 * @phys_addr: bus address of the memory
363 * @size: size of the resource to map
364 *
365 * ioremap_uc performs a platform specific sequence of operations to
366 * make bus memory CPU accessible via the readb/readw/readl/writeb/
367 * writew/writel functions and the other mmio helpers. The returned
368 * address is not guaranteed to be usable directly as a virtual
369 * address.
370 *
371 * This version of ioremap ensures that the memory is marked with a strong
372 * preference as completely uncachable on the CPU when possible. For non-PAT
373 * systems this ends up setting page-attribute flags PCD=1, PWT=1. For PAT
374 * systems this will set the PAT entry for the pages as strong UC. This call
375 * will honor existing caching rules from things like the PCI bus. Note that
376 * there are other caches and buffers on many busses. In particular driver
377 * authors should read up on PCI writes.
378 *
379 * It's useful if some control registers are in such an area and
380 * write combining or read caching is not desirable:
381 *
382 * Must be freed with iounmap.
383 */
384void __iomem *ioremap_uc(resource_size_t phys_addr, unsigned long size)
385{
386 enum page_cache_mode pcm = _PAGE_CACHE_MODE_UC;
387
388 return __ioremap_caller(phys_addr, size, pcm,
389 __builtin_return_address(0), false);
390}
391EXPORT_SYMBOL_GPL(ioremap_uc);
392
393/**
394 * ioremap_wc - map memory into CPU space write combined
395 * @phys_addr: bus address of the memory
396 * @size: size of the resource to map
397 *
398 * This version of ioremap ensures that the memory is marked write combining.
399 * Write combining allows faster writes to some hardware devices.
400 *
401 * Must be freed with iounmap.
402 */
403void __iomem *ioremap_wc(resource_size_t phys_addr, unsigned long size)
404{
405 return __ioremap_caller(phys_addr, size, _PAGE_CACHE_MODE_WC,
406 __builtin_return_address(0), false);
407}
408EXPORT_SYMBOL(ioremap_wc);
409
410/**
411 * ioremap_wt - map memory into CPU space write through
412 * @phys_addr: bus address of the memory
413 * @size: size of the resource to map
414 *
415 * This version of ioremap ensures that the memory is marked write through.
416 * Write through stores data into memory while keeping the cache up-to-date.
417 *
418 * Must be freed with iounmap.
419 */
420void __iomem *ioremap_wt(resource_size_t phys_addr, unsigned long size)
421{
422 return __ioremap_caller(phys_addr, size, _PAGE_CACHE_MODE_WT,
423 __builtin_return_address(0), false);
424}
425EXPORT_SYMBOL(ioremap_wt);
426
427void __iomem *ioremap_encrypted(resource_size_t phys_addr, unsigned long size)
428{
429 return __ioremap_caller(phys_addr, size, _PAGE_CACHE_MODE_WB,
430 __builtin_return_address(0), true);
431}
432EXPORT_SYMBOL(ioremap_encrypted);
433
434void __iomem *ioremap_cache(resource_size_t phys_addr, unsigned long size)
435{
436 return __ioremap_caller(phys_addr, size, _PAGE_CACHE_MODE_WB,
437 __builtin_return_address(0), false);
438}
439EXPORT_SYMBOL(ioremap_cache);
440
441void __iomem *ioremap_prot(resource_size_t phys_addr, unsigned long size,
442 unsigned long prot_val)
443{
444 return __ioremap_caller(phys_addr, size,
445 pgprot2cachemode(__pgprot(prot_val)),
446 __builtin_return_address(0), false);
447}
448EXPORT_SYMBOL(ioremap_prot);
449
450/**
451 * iounmap - Free a IO remapping
452 * @addr: virtual address from ioremap_*
453 *
454 * Caller must ensure there is only one unmapping for the same pointer.
455 */
456void iounmap(volatile void __iomem *addr)
457{
458 struct vm_struct *p, *o;
459
460 if ((void __force *)addr <= high_memory)
461 return;
462
463 /*
464 * The PCI/ISA range special-casing was removed from __ioremap()
465 * so this check, in theory, can be removed. However, there are
466 * cases where iounmap() is called for addresses not obtained via
467 * ioremap() (vga16fb for example). Add a warning so that these
468 * cases can be caught and fixed.
469 */
470 if ((void __force *)addr >= phys_to_virt(ISA_START_ADDRESS) &&
471 (void __force *)addr < phys_to_virt(ISA_END_ADDRESS)) {
472 WARN(1, "iounmap() called for ISA range not obtained using ioremap()\n");
473 return;
474 }
475
476 mmiotrace_iounmap(addr);
477
478 addr = (volatile void __iomem *)
479 (PAGE_MASK & (unsigned long __force)addr);
480
481 /* Use the vm area unlocked, assuming the caller
482 ensures there isn't another iounmap for the same address
483 in parallel. Reuse of the virtual address is prevented by
484 leaving it in the global lists until we're done with it.
485 cpa takes care of the direct mappings. */
486 p = find_vm_area((void __force *)addr);
487
488 if (!p) {
489 printk(KERN_ERR "iounmap: bad address %p\n", addr);
490 dump_stack();
491 return;
492 }
493
494 kmsan_iounmap_page_range((unsigned long)addr,
495 (unsigned long)addr + get_vm_area_size(p));
496 memtype_free(p->phys_addr, p->phys_addr + get_vm_area_size(p));
497
498 /* Finally remove it */
499 o = remove_vm_area((void __force *)addr);
500 BUG_ON(p != o || o == NULL);
501 kfree(p);
502}
503EXPORT_SYMBOL(iounmap);
504
505/*
506 * Convert a physical pointer to a virtual kernel pointer for /dev/mem
507 * access
508 */
509void *xlate_dev_mem_ptr(phys_addr_t phys)
510{
511 unsigned long start = phys & PAGE_MASK;
512 unsigned long offset = phys & ~PAGE_MASK;
513 void *vaddr;
514
515 /* memremap() maps if RAM, otherwise falls back to ioremap() */
516 vaddr = memremap(start, PAGE_SIZE, MEMREMAP_WB);
517
518 /* Only add the offset on success and return NULL if memremap() failed */
519 if (vaddr)
520 vaddr += offset;
521
522 return vaddr;
523}
524
525void unxlate_dev_mem_ptr(phys_addr_t phys, void *addr)
526{
527 memunmap((void *)((unsigned long)addr & PAGE_MASK));
528}
529
530#ifdef CONFIG_AMD_MEM_ENCRYPT
531/*
532 * Examine the physical address to determine if it is an area of memory
533 * that should be mapped decrypted. If the memory is not part of the
534 * kernel usable area it was accessed and created decrypted, so these
535 * areas should be mapped decrypted. And since the encryption key can
536 * change across reboots, persistent memory should also be mapped
537 * decrypted.
538 *
539 * If SEV is active, that implies that BIOS/UEFI also ran encrypted so
540 * only persistent memory should be mapped decrypted.
541 */
542static bool memremap_should_map_decrypted(resource_size_t phys_addr,
543 unsigned long size)
544{
545 int is_pmem;
546
547 /*
548 * Check if the address is part of a persistent memory region.
549 * This check covers areas added by E820, EFI and ACPI.
550 */
551 is_pmem = region_intersects(phys_addr, size, IORESOURCE_MEM,
552 IORES_DESC_PERSISTENT_MEMORY);
553 if (is_pmem != REGION_DISJOINT)
554 return true;
555
556 /*
557 * Check if the non-volatile attribute is set for an EFI
558 * reserved area.
559 */
560 if (efi_enabled(EFI_BOOT)) {
561 switch (efi_mem_type(phys_addr)) {
562 case EFI_RESERVED_TYPE:
563 if (efi_mem_attributes(phys_addr) & EFI_MEMORY_NV)
564 return true;
565 break;
566 default:
567 break;
568 }
569 }
570
571 /* Check if the address is outside kernel usable area */
572 switch (e820__get_entry_type(phys_addr, phys_addr + size - 1)) {
573 case E820_TYPE_RESERVED:
574 case E820_TYPE_ACPI:
575 case E820_TYPE_NVS:
576 case E820_TYPE_UNUSABLE:
577 /* For SEV, these areas are encrypted */
578 if (cc_platform_has(CC_ATTR_GUEST_MEM_ENCRYPT))
579 break;
580 fallthrough;
581
582 case E820_TYPE_PRAM:
583 return true;
584 default:
585 break;
586 }
587
588 return false;
589}
590
591/*
592 * Examine the physical address to determine if it is EFI data. Check
593 * it against the boot params structure and EFI tables and memory types.
594 */
595static bool memremap_is_efi_data(resource_size_t phys_addr,
596 unsigned long size)
597{
598 u64 paddr;
599
600 /* Check if the address is part of EFI boot/runtime data */
601 if (!efi_enabled(EFI_BOOT))
602 return false;
603
604 paddr = boot_params.efi_info.efi_memmap_hi;
605 paddr <<= 32;
606 paddr |= boot_params.efi_info.efi_memmap;
607 if (phys_addr == paddr)
608 return true;
609
610 paddr = boot_params.efi_info.efi_systab_hi;
611 paddr <<= 32;
612 paddr |= boot_params.efi_info.efi_systab;
613 if (phys_addr == paddr)
614 return true;
615
616 if (efi_is_table_address(phys_addr))
617 return true;
618
619 switch (efi_mem_type(phys_addr)) {
620 case EFI_BOOT_SERVICES_DATA:
621 case EFI_RUNTIME_SERVICES_DATA:
622 return true;
623 default:
624 break;
625 }
626
627 return false;
628}
629
630/*
631 * Examine the physical address to determine if it is boot data by checking
632 * it against the boot params setup_data chain.
633 */
634static bool memremap_is_setup_data(resource_size_t phys_addr,
635 unsigned long size)
636{
637 struct setup_indirect *indirect;
638 struct setup_data *data;
639 u64 paddr, paddr_next;
640
641 paddr = boot_params.hdr.setup_data;
642 while (paddr) {
643 unsigned int len;
644
645 if (phys_addr == paddr)
646 return true;
647
648 data = memremap(paddr, sizeof(*data),
649 MEMREMAP_WB | MEMREMAP_DEC);
650 if (!data) {
651 pr_warn("failed to memremap setup_data entry\n");
652 return false;
653 }
654
655 paddr_next = data->next;
656 len = data->len;
657
658 if ((phys_addr > paddr) && (phys_addr < (paddr + len))) {
659 memunmap(data);
660 return true;
661 }
662
663 if (data->type == SETUP_INDIRECT) {
664 memunmap(data);
665 data = memremap(paddr, sizeof(*data) + len,
666 MEMREMAP_WB | MEMREMAP_DEC);
667 if (!data) {
668 pr_warn("failed to memremap indirect setup_data\n");
669 return false;
670 }
671
672 indirect = (struct setup_indirect *)data->data;
673
674 if (indirect->type != SETUP_INDIRECT) {
675 paddr = indirect->addr;
676 len = indirect->len;
677 }
678 }
679
680 memunmap(data);
681
682 if ((phys_addr > paddr) && (phys_addr < (paddr + len)))
683 return true;
684
685 paddr = paddr_next;
686 }
687
688 return false;
689}
690
691/*
692 * Examine the physical address to determine if it is boot data by checking
693 * it against the boot params setup_data chain (early boot version).
694 */
695static bool __init early_memremap_is_setup_data(resource_size_t phys_addr,
696 unsigned long size)
697{
698 struct setup_indirect *indirect;
699 struct setup_data *data;
700 u64 paddr, paddr_next;
701
702 paddr = boot_params.hdr.setup_data;
703 while (paddr) {
704 unsigned int len, size;
705
706 if (phys_addr == paddr)
707 return true;
708
709 data = early_memremap_decrypted(paddr, sizeof(*data));
710 if (!data) {
711 pr_warn("failed to early memremap setup_data entry\n");
712 return false;
713 }
714
715 size = sizeof(*data);
716
717 paddr_next = data->next;
718 len = data->len;
719
720 if ((phys_addr > paddr) && (phys_addr < (paddr + len))) {
721 early_memunmap(data, sizeof(*data));
722 return true;
723 }
724
725 if (data->type == SETUP_INDIRECT) {
726 size += len;
727 early_memunmap(data, sizeof(*data));
728 data = early_memremap_decrypted(paddr, size);
729 if (!data) {
730 pr_warn("failed to early memremap indirect setup_data\n");
731 return false;
732 }
733
734 indirect = (struct setup_indirect *)data->data;
735
736 if (indirect->type != SETUP_INDIRECT) {
737 paddr = indirect->addr;
738 len = indirect->len;
739 }
740 }
741
742 early_memunmap(data, size);
743
744 if ((phys_addr > paddr) && (phys_addr < (paddr + len)))
745 return true;
746
747 paddr = paddr_next;
748 }
749
750 return false;
751}
752
753/*
754 * Architecture function to determine if RAM remap is allowed. By default, a
755 * RAM remap will map the data as encrypted. Determine if a RAM remap should
756 * not be done so that the data will be mapped decrypted.
757 */
758bool arch_memremap_can_ram_remap(resource_size_t phys_addr, unsigned long size,
759 unsigned long flags)
760{
761 if (!cc_platform_has(CC_ATTR_MEM_ENCRYPT))
762 return true;
763
764 if (flags & MEMREMAP_ENC)
765 return true;
766
767 if (flags & MEMREMAP_DEC)
768 return false;
769
770 if (cc_platform_has(CC_ATTR_HOST_MEM_ENCRYPT)) {
771 if (memremap_is_setup_data(phys_addr, size) ||
772 memremap_is_efi_data(phys_addr, size))
773 return false;
774 }
775
776 return !memremap_should_map_decrypted(phys_addr, size);
777}
778
779/*
780 * Architecture override of __weak function to adjust the protection attributes
781 * used when remapping memory. By default, early_memremap() will map the data
782 * as encrypted. Determine if an encrypted mapping should not be done and set
783 * the appropriate protection attributes.
784 */
785pgprot_t __init early_memremap_pgprot_adjust(resource_size_t phys_addr,
786 unsigned long size,
787 pgprot_t prot)
788{
789 bool encrypted_prot;
790
791 if (!cc_platform_has(CC_ATTR_MEM_ENCRYPT))
792 return prot;
793
794 encrypted_prot = true;
795
796 if (cc_platform_has(CC_ATTR_HOST_MEM_ENCRYPT)) {
797 if (early_memremap_is_setup_data(phys_addr, size) ||
798 memremap_is_efi_data(phys_addr, size))
799 encrypted_prot = false;
800 }
801
802 if (encrypted_prot && memremap_should_map_decrypted(phys_addr, size))
803 encrypted_prot = false;
804
805 return encrypted_prot ? pgprot_encrypted(prot)
806 : pgprot_decrypted(prot);
807}
808
809bool phys_mem_access_encrypted(unsigned long phys_addr, unsigned long size)
810{
811 return arch_memremap_can_ram_remap(phys_addr, size, 0);
812}
813
814/* Remap memory with encryption */
815void __init *early_memremap_encrypted(resource_size_t phys_addr,
816 unsigned long size)
817{
818 return early_memremap_prot(phys_addr, size, __PAGE_KERNEL_ENC);
819}
820
821/*
822 * Remap memory with encryption and write-protected - cannot be called
823 * before pat_init() is called
824 */
825void __init *early_memremap_encrypted_wp(resource_size_t phys_addr,
826 unsigned long size)
827{
828 if (!x86_has_pat_wp())
829 return NULL;
830 return early_memremap_prot(phys_addr, size, __PAGE_KERNEL_ENC_WP);
831}
832
833/* Remap memory without encryption */
834void __init *early_memremap_decrypted(resource_size_t phys_addr,
835 unsigned long size)
836{
837 return early_memremap_prot(phys_addr, size, __PAGE_KERNEL_NOENC);
838}
839
840/*
841 * Remap memory without encryption and write-protected - cannot be called
842 * before pat_init() is called
843 */
844void __init *early_memremap_decrypted_wp(resource_size_t phys_addr,
845 unsigned long size)
846{
847 if (!x86_has_pat_wp())
848 return NULL;
849 return early_memremap_prot(phys_addr, size, __PAGE_KERNEL_NOENC_WP);
850}
851#endif /* CONFIG_AMD_MEM_ENCRYPT */
852
853static pte_t bm_pte[PAGE_SIZE/sizeof(pte_t)] __page_aligned_bss;
854
855static inline pmd_t * __init early_ioremap_pmd(unsigned long addr)
856{
857 /* Don't assume we're using swapper_pg_dir at this point */
858 pgd_t *base = __va(read_cr3_pa());
859 pgd_t *pgd = &base[pgd_index(addr)];
860 p4d_t *p4d = p4d_offset(pgd, addr);
861 pud_t *pud = pud_offset(p4d, addr);
862 pmd_t *pmd = pmd_offset(pud, addr);
863
864 return pmd;
865}
866
867static inline pte_t * __init early_ioremap_pte(unsigned long addr)
868{
869 return &bm_pte[pte_index(addr)];
870}
871
872bool __init is_early_ioremap_ptep(pte_t *ptep)
873{
874 return ptep >= &bm_pte[0] && ptep < &bm_pte[PAGE_SIZE/sizeof(pte_t)];
875}
876
877void __init early_ioremap_init(void)
878{
879 pmd_t *pmd;
880
881#ifdef CONFIG_X86_64
882 BUILD_BUG_ON((fix_to_virt(0) + PAGE_SIZE) & ((1 << PMD_SHIFT) - 1));
883#else
884 WARN_ON((fix_to_virt(0) + PAGE_SIZE) & ((1 << PMD_SHIFT) - 1));
885#endif
886
887 early_ioremap_setup();
888
889 pmd = early_ioremap_pmd(fix_to_virt(FIX_BTMAP_BEGIN));
890 memset(bm_pte, 0, sizeof(bm_pte));
891 pmd_populate_kernel(&init_mm, pmd, bm_pte);
892
893 /*
894 * The boot-ioremap range spans multiple pmds, for which
895 * we are not prepared:
896 */
897#define __FIXADDR_TOP (-PAGE_SIZE)
898 BUILD_BUG_ON((__fix_to_virt(FIX_BTMAP_BEGIN) >> PMD_SHIFT)
899 != (__fix_to_virt(FIX_BTMAP_END) >> PMD_SHIFT));
900#undef __FIXADDR_TOP
901 if (pmd != early_ioremap_pmd(fix_to_virt(FIX_BTMAP_END))) {
902 WARN_ON(1);
903 printk(KERN_WARNING "pmd %p != %p\n",
904 pmd, early_ioremap_pmd(fix_to_virt(FIX_BTMAP_END)));
905 printk(KERN_WARNING "fix_to_virt(FIX_BTMAP_BEGIN): %08lx\n",
906 fix_to_virt(FIX_BTMAP_BEGIN));
907 printk(KERN_WARNING "fix_to_virt(FIX_BTMAP_END): %08lx\n",
908 fix_to_virt(FIX_BTMAP_END));
909
910 printk(KERN_WARNING "FIX_BTMAP_END: %d\n", FIX_BTMAP_END);
911 printk(KERN_WARNING "FIX_BTMAP_BEGIN: %d\n",
912 FIX_BTMAP_BEGIN);
913 }
914}
915
916void __init __early_set_fixmap(enum fixed_addresses idx,
917 phys_addr_t phys, pgprot_t flags)
918{
919 unsigned long addr = __fix_to_virt(idx);
920 pte_t *pte;
921
922 if (idx >= __end_of_fixed_addresses) {
923 BUG();
924 return;
925 }
926 pte = early_ioremap_pte(addr);
927
928 /* Sanitize 'prot' against any unsupported bits: */
929 pgprot_val(flags) &= __supported_pte_mask;
930
931 if (pgprot_val(flags))
932 set_pte(pte, pfn_pte(phys >> PAGE_SHIFT, flags));
933 else
934 pte_clear(&init_mm, addr, pte);
935 flush_tlb_one_kernel(addr);
936}