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 (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}
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}