arm: add support for folded p4d page tables
[linux-2.6-microblaze.git] / arch / arm / mm / ioremap.c
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  *  linux/arch/arm/mm/ioremap.c
4  *
5  * Re-map IO memory to kernel address space so that we can access it.
6  *
7  * (C) Copyright 1995 1996 Linus Torvalds
8  *
9  * Hacked for ARM by Phil Blundell <philb@gnu.org>
10  * Hacked to allow all architectures to build, and various cleanups
11  * by Russell King
12  *
13  * This allows a driver to remap an arbitrary region of bus memory into
14  * virtual space.  One should *only* use readl, writel, memcpy_toio and
15  * so on with such remapped areas.
16  *
17  * Because the ARM only has a 32-bit address space we can't address the
18  * whole of the (physical) PCI space at once.  PCI huge-mode addressing
19  * allows us to circumvent this restriction by splitting PCI space into
20  * two 2GB chunks and mapping only one at a time into processor memory.
21  * We use MMU protection domains to trap any attempt to access the bank
22  * that is not currently mapped.  (This isn't fully implemented yet.)
23  */
24 #include <linux/module.h>
25 #include <linux/errno.h>
26 #include <linux/mm.h>
27 #include <linux/vmalloc.h>
28 #include <linux/io.h>
29 #include <linux/sizes.h>
30
31 #include <asm/cp15.h>
32 #include <asm/cputype.h>
33 #include <asm/cacheflush.h>
34 #include <asm/early_ioremap.h>
35 #include <asm/mmu_context.h>
36 #include <asm/pgalloc.h>
37 #include <asm/tlbflush.h>
38 #include <asm/system_info.h>
39
40 #include <asm/mach/map.h>
41 #include <asm/mach/pci.h>
42 #include "mm.h"
43
44
45 LIST_HEAD(static_vmlist);
46
47 static struct static_vm *find_static_vm_paddr(phys_addr_t paddr,
48                         size_t size, unsigned int mtype)
49 {
50         struct static_vm *svm;
51         struct vm_struct *vm;
52
53         list_for_each_entry(svm, &static_vmlist, list) {
54                 vm = &svm->vm;
55                 if (!(vm->flags & VM_ARM_STATIC_MAPPING))
56                         continue;
57                 if ((vm->flags & VM_ARM_MTYPE_MASK) != VM_ARM_MTYPE(mtype))
58                         continue;
59
60                 if (vm->phys_addr > paddr ||
61                         paddr + size - 1 > vm->phys_addr + vm->size - 1)
62                         continue;
63
64                 return svm;
65         }
66
67         return NULL;
68 }
69
70 struct static_vm *find_static_vm_vaddr(void *vaddr)
71 {
72         struct static_vm *svm;
73         struct vm_struct *vm;
74
75         list_for_each_entry(svm, &static_vmlist, list) {
76                 vm = &svm->vm;
77
78                 /* static_vmlist is ascending order */
79                 if (vm->addr > vaddr)
80                         break;
81
82                 if (vm->addr <= vaddr && vm->addr + vm->size > vaddr)
83                         return svm;
84         }
85
86         return NULL;
87 }
88
89 void __init add_static_vm_early(struct static_vm *svm)
90 {
91         struct static_vm *curr_svm;
92         struct vm_struct *vm;
93         void *vaddr;
94
95         vm = &svm->vm;
96         vm_area_add_early(vm);
97         vaddr = vm->addr;
98
99         list_for_each_entry(curr_svm, &static_vmlist, list) {
100                 vm = &curr_svm->vm;
101
102                 if (vm->addr > vaddr)
103                         break;
104         }
105         list_add_tail(&svm->list, &curr_svm->list);
106 }
107
108 int ioremap_page(unsigned long virt, unsigned long phys,
109                  const struct mem_type *mtype)
110 {
111         return ioremap_page_range(virt, virt + PAGE_SIZE, phys,
112                                   __pgprot(mtype->prot_pte));
113 }
114 EXPORT_SYMBOL(ioremap_page);
115
116 void __check_vmalloc_seq(struct mm_struct *mm)
117 {
118         unsigned int seq;
119
120         do {
121                 seq = init_mm.context.vmalloc_seq;
122                 memcpy(pgd_offset(mm, VMALLOC_START),
123                        pgd_offset_k(VMALLOC_START),
124                        sizeof(pgd_t) * (pgd_index(VMALLOC_END) -
125                                         pgd_index(VMALLOC_START)));
126                 mm->context.vmalloc_seq = seq;
127         } while (seq != init_mm.context.vmalloc_seq);
128 }
129
130 #if !defined(CONFIG_SMP) && !defined(CONFIG_ARM_LPAE)
131 /*
132  * Section support is unsafe on SMP - If you iounmap and ioremap a region,
133  * the other CPUs will not see this change until their next context switch.
134  * Meanwhile, (eg) if an interrupt comes in on one of those other CPUs
135  * which requires the new ioremap'd region to be referenced, the CPU will
136  * reference the _old_ region.
137  *
138  * Note that get_vm_area_caller() allocates a guard 4K page, so we need to
139  * mask the size back to 1MB aligned or we will overflow in the loop below.
140  */
141 static void unmap_area_sections(unsigned long virt, unsigned long size)
142 {
143         unsigned long addr = virt, end = virt + (size & ~(SZ_1M - 1));
144         pgd_t *pgd;
145         p4d_t *p4d;
146         pud_t *pud;
147         pmd_t *pmdp;
148
149         flush_cache_vunmap(addr, end);
150         pgd = pgd_offset_k(addr);
151         p4d = p4d_offset(pgd, addr);
152         pud = pud_offset(p4d, addr);
153         pmdp = pmd_offset(pud, addr);
154         do {
155                 pmd_t pmd = *pmdp;
156
157                 if (!pmd_none(pmd)) {
158                         /*
159                          * Clear the PMD from the page table, and
160                          * increment the vmalloc sequence so others
161                          * notice this change.
162                          *
163                          * Note: this is still racy on SMP machines.
164                          */
165                         pmd_clear(pmdp);
166                         init_mm.context.vmalloc_seq++;
167
168                         /*
169                          * Free the page table, if there was one.
170                          */
171                         if ((pmd_val(pmd) & PMD_TYPE_MASK) == PMD_TYPE_TABLE)
172                                 pte_free_kernel(&init_mm, pmd_page_vaddr(pmd));
173                 }
174
175                 addr += PMD_SIZE;
176                 pmdp += 2;
177         } while (addr < end);
178
179         /*
180          * Ensure that the active_mm is up to date - we want to
181          * catch any use-after-iounmap cases.
182          */
183         if (current->active_mm->context.vmalloc_seq != init_mm.context.vmalloc_seq)
184                 __check_vmalloc_seq(current->active_mm);
185
186         flush_tlb_kernel_range(virt, end);
187 }
188
189 static int
190 remap_area_sections(unsigned long virt, unsigned long pfn,
191                     size_t size, const struct mem_type *type)
192 {
193         unsigned long addr = virt, end = virt + size;
194         pgd_t *pgd;
195         p4d_t *p4d;
196         pud_t *pud;
197         pmd_t *pmd;
198
199         /*
200          * Remove and free any PTE-based mapping, and
201          * sync the current kernel mapping.
202          */
203         unmap_area_sections(virt, size);
204
205         pgd = pgd_offset_k(addr);
206         p4d = p4d_offset(pgd, addr);
207         pud = pud_offset(p4d, addr);
208         pmd = pmd_offset(pud, addr);
209         do {
210                 pmd[0] = __pmd(__pfn_to_phys(pfn) | type->prot_sect);
211                 pfn += SZ_1M >> PAGE_SHIFT;
212                 pmd[1] = __pmd(__pfn_to_phys(pfn) | type->prot_sect);
213                 pfn += SZ_1M >> PAGE_SHIFT;
214                 flush_pmd_entry(pmd);
215
216                 addr += PMD_SIZE;
217                 pmd += 2;
218         } while (addr < end);
219
220         return 0;
221 }
222
223 static int
224 remap_area_supersections(unsigned long virt, unsigned long pfn,
225                          size_t size, const struct mem_type *type)
226 {
227         unsigned long addr = virt, end = virt + size;
228         pgd_t *pgd;
229         p4d_t *p4d;
230         pud_t *pud;
231         pmd_t *pmd;
232
233         /*
234          * Remove and free any PTE-based mapping, and
235          * sync the current kernel mapping.
236          */
237         unmap_area_sections(virt, size);
238
239         pgd = pgd_offset_k(virt);
240         p4d = p4d_offset(pgd, addr);
241         pud = pud_offset(p4d, addr);
242         pmd = pmd_offset(pud, addr);
243         do {
244                 unsigned long super_pmd_val, i;
245
246                 super_pmd_val = __pfn_to_phys(pfn) | type->prot_sect |
247                                 PMD_SECT_SUPER;
248                 super_pmd_val |= ((pfn >> (32 - PAGE_SHIFT)) & 0xf) << 20;
249
250                 for (i = 0; i < 8; i++) {
251                         pmd[0] = __pmd(super_pmd_val);
252                         pmd[1] = __pmd(super_pmd_val);
253                         flush_pmd_entry(pmd);
254
255                         addr += PMD_SIZE;
256                         pmd += 2;
257                 }
258
259                 pfn += SUPERSECTION_SIZE >> PAGE_SHIFT;
260         } while (addr < end);
261
262         return 0;
263 }
264 #endif
265
266 static void __iomem * __arm_ioremap_pfn_caller(unsigned long pfn,
267         unsigned long offset, size_t size, unsigned int mtype, void *caller)
268 {
269         const struct mem_type *type;
270         int err;
271         unsigned long addr;
272         struct vm_struct *area;
273         phys_addr_t paddr = __pfn_to_phys(pfn);
274
275 #ifndef CONFIG_ARM_LPAE
276         /*
277          * High mappings must be supersection aligned
278          */
279         if (pfn >= 0x100000 && (paddr & ~SUPERSECTION_MASK))
280                 return NULL;
281 #endif
282
283         type = get_mem_type(mtype);
284         if (!type)
285                 return NULL;
286
287         /*
288          * Page align the mapping size, taking account of any offset.
289          */
290         size = PAGE_ALIGN(offset + size);
291
292         /*
293          * Try to reuse one of the static mapping whenever possible.
294          */
295         if (size && !(sizeof(phys_addr_t) == 4 && pfn >= 0x100000)) {
296                 struct static_vm *svm;
297
298                 svm = find_static_vm_paddr(paddr, size, mtype);
299                 if (svm) {
300                         addr = (unsigned long)svm->vm.addr;
301                         addr += paddr - svm->vm.phys_addr;
302                         return (void __iomem *) (offset + addr);
303                 }
304         }
305
306         /*
307          * Don't allow RAM to be mapped with mismatched attributes - this
308          * causes problems with ARMv6+
309          */
310         if (WARN_ON(pfn_valid(pfn) && mtype != MT_MEMORY_RW))
311                 return NULL;
312
313         area = get_vm_area_caller(size, VM_IOREMAP, caller);
314         if (!area)
315                 return NULL;
316         addr = (unsigned long)area->addr;
317         area->phys_addr = paddr;
318
319 #if !defined(CONFIG_SMP) && !defined(CONFIG_ARM_LPAE)
320         if (DOMAIN_IO == 0 &&
321             (((cpu_architecture() >= CPU_ARCH_ARMv6) && (get_cr() & CR_XP)) ||
322                cpu_is_xsc3()) && pfn >= 0x100000 &&
323                !((paddr | size | addr) & ~SUPERSECTION_MASK)) {
324                 area->flags |= VM_ARM_SECTION_MAPPING;
325                 err = remap_area_supersections(addr, pfn, size, type);
326         } else if (!((paddr | size | addr) & ~PMD_MASK)) {
327                 area->flags |= VM_ARM_SECTION_MAPPING;
328                 err = remap_area_sections(addr, pfn, size, type);
329         } else
330 #endif
331                 err = ioremap_page_range(addr, addr + size, paddr,
332                                          __pgprot(type->prot_pte));
333
334         if (err) {
335                 vunmap((void *)addr);
336                 return NULL;
337         }
338
339         flush_cache_vmap(addr, addr + size);
340         return (void __iomem *) (offset + addr);
341 }
342
343 void __iomem *__arm_ioremap_caller(phys_addr_t phys_addr, size_t size,
344         unsigned int mtype, void *caller)
345 {
346         phys_addr_t last_addr;
347         unsigned long offset = phys_addr & ~PAGE_MASK;
348         unsigned long pfn = __phys_to_pfn(phys_addr);
349
350         /*
351          * Don't allow wraparound or zero size
352          */
353         last_addr = phys_addr + size - 1;
354         if (!size || last_addr < phys_addr)
355                 return NULL;
356
357         return __arm_ioremap_pfn_caller(pfn, offset, size, mtype,
358                         caller);
359 }
360
361 /*
362  * Remap an arbitrary physical address space into the kernel virtual
363  * address space. Needed when the kernel wants to access high addresses
364  * directly.
365  *
366  * NOTE! We need to allow non-page-aligned mappings too: we will obviously
367  * have to convert them into an offset in a page-aligned mapping, but the
368  * caller shouldn't need to know that small detail.
369  */
370 void __iomem *
371 __arm_ioremap_pfn(unsigned long pfn, unsigned long offset, size_t size,
372                   unsigned int mtype)
373 {
374         return __arm_ioremap_pfn_caller(pfn, offset, size, mtype,
375                                         __builtin_return_address(0));
376 }
377 EXPORT_SYMBOL(__arm_ioremap_pfn);
378
379 void __iomem * (*arch_ioremap_caller)(phys_addr_t, size_t,
380                                       unsigned int, void *) =
381         __arm_ioremap_caller;
382
383 void __iomem *ioremap(resource_size_t res_cookie, size_t size)
384 {
385         return arch_ioremap_caller(res_cookie, size, MT_DEVICE,
386                                    __builtin_return_address(0));
387 }
388 EXPORT_SYMBOL(ioremap);
389
390 void __iomem *ioremap_cache(resource_size_t res_cookie, size_t size)
391 {
392         return arch_ioremap_caller(res_cookie, size, MT_DEVICE_CACHED,
393                                    __builtin_return_address(0));
394 }
395 EXPORT_SYMBOL(ioremap_cache);
396
397 void __iomem *ioremap_wc(resource_size_t res_cookie, size_t size)
398 {
399         return arch_ioremap_caller(res_cookie, size, MT_DEVICE_WC,
400                                    __builtin_return_address(0));
401 }
402 EXPORT_SYMBOL(ioremap_wc);
403
404 /*
405  * Remap an arbitrary physical address space into the kernel virtual
406  * address space as memory. Needed when the kernel wants to execute
407  * code in external memory. This is needed for reprogramming source
408  * clocks that would affect normal memory for example. Please see
409  * CONFIG_GENERIC_ALLOCATOR for allocating external memory.
410  */
411 void __iomem *
412 __arm_ioremap_exec(phys_addr_t phys_addr, size_t size, bool cached)
413 {
414         unsigned int mtype;
415
416         if (cached)
417                 mtype = MT_MEMORY_RWX;
418         else
419                 mtype = MT_MEMORY_RWX_NONCACHED;
420
421         return __arm_ioremap_caller(phys_addr, size, mtype,
422                         __builtin_return_address(0));
423 }
424
425 void *arch_memremap_wb(phys_addr_t phys_addr, size_t size)
426 {
427         return (__force void *)arch_ioremap_caller(phys_addr, size,
428                                                    MT_MEMORY_RW,
429                                                    __builtin_return_address(0));
430 }
431
432 void __iounmap(volatile void __iomem *io_addr)
433 {
434         void *addr = (void *)(PAGE_MASK & (unsigned long)io_addr);
435         struct static_vm *svm;
436
437         /* If this is a static mapping, we must leave it alone */
438         svm = find_static_vm_vaddr(addr);
439         if (svm)
440                 return;
441
442 #if !defined(CONFIG_SMP) && !defined(CONFIG_ARM_LPAE)
443         {
444                 struct vm_struct *vm;
445
446                 vm = find_vm_area(addr);
447
448                 /*
449                  * If this is a section based mapping we need to handle it
450                  * specially as the VM subsystem does not know how to handle
451                  * such a beast.
452                  */
453                 if (vm && (vm->flags & VM_ARM_SECTION_MAPPING))
454                         unmap_area_sections((unsigned long)vm->addr, vm->size);
455         }
456 #endif
457
458         vunmap(addr);
459 }
460
461 void (*arch_iounmap)(volatile void __iomem *) = __iounmap;
462
463 void iounmap(volatile void __iomem *cookie)
464 {
465         arch_iounmap(cookie);
466 }
467 EXPORT_SYMBOL(iounmap);
468
469 #ifdef CONFIG_PCI
470 static int pci_ioremap_mem_type = MT_DEVICE;
471
472 void pci_ioremap_set_mem_type(int mem_type)
473 {
474         pci_ioremap_mem_type = mem_type;
475 }
476
477 int pci_ioremap_io(unsigned int offset, phys_addr_t phys_addr)
478 {
479         BUG_ON(offset + SZ_64K - 1 > IO_SPACE_LIMIT);
480
481         return ioremap_page_range(PCI_IO_VIRT_BASE + offset,
482                                   PCI_IO_VIRT_BASE + offset + SZ_64K,
483                                   phys_addr,
484                                   __pgprot(get_mem_type(pci_ioremap_mem_type)->prot_pte));
485 }
486 EXPORT_SYMBOL_GPL(pci_ioremap_io);
487
488 void __iomem *pci_remap_cfgspace(resource_size_t res_cookie, size_t size)
489 {
490         return arch_ioremap_caller(res_cookie, size, MT_UNCACHED,
491                                    __builtin_return_address(0));
492 }
493 EXPORT_SYMBOL_GPL(pci_remap_cfgspace);
494 #endif
495
496 /*
497  * Must be called after early_fixmap_init
498  */
499 void __init early_ioremap_init(void)
500 {
501         early_ioremap_setup();
502 }