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