1 // SPDX-License-Identifier: GPL-2.0
3 * Initialize MMU support.
5 * Copyright (C) 1998-2003 Hewlett-Packard Co
6 * David Mosberger-Tang <davidm@hpl.hp.com>
8 #include <linux/kernel.h>
9 #include <linux/init.h>
11 #include <linux/efi.h>
12 #include <linux/elf.h>
13 #include <linux/memblock.h>
15 #include <linux/sched/signal.h>
16 #include <linux/mmzone.h>
17 #include <linux/module.h>
18 #include <linux/personality.h>
19 #include <linux/reboot.h>
20 #include <linux/slab.h>
21 #include <linux/swap.h>
22 #include <linux/proc_fs.h>
23 #include <linux/bitops.h>
24 #include <linux/kexec.h>
28 #include <asm/machvec.h>
30 #include <asm/patch.h>
31 #include <asm/pgalloc.h>
33 #include <asm/sections.h>
35 #include <linux/uaccess.h>
36 #include <asm/unistd.h>
39 extern void ia64_tlb_init (void);
41 unsigned long MAX_DMA_ADDRESS = PAGE_OFFSET + 0x100000000UL;
43 #ifdef CONFIG_VIRTUAL_MEM_MAP
44 unsigned long VMALLOC_END = VMALLOC_END_INIT;
45 EXPORT_SYMBOL(VMALLOC_END);
46 struct page *vmem_map;
47 EXPORT_SYMBOL(vmem_map);
50 struct page *zero_page_memmap_ptr; /* map entry for zero page */
51 EXPORT_SYMBOL(zero_page_memmap_ptr);
54 __ia64_sync_icache_dcache (pte_t pte)
60 addr = (unsigned long) page_address(page);
62 if (test_bit(PG_arch_1, &page->flags))
63 return; /* i-cache is already coherent with d-cache */
65 flush_icache_range(addr, addr + (PAGE_SIZE << compound_order(page)));
66 set_bit(PG_arch_1, &page->flags); /* mark page as clean */
70 * Since DMA is i-cache coherent, any (complete) pages that were written via
71 * DMA can be marked as "clean" so that lazy_mmu_prot_update() doesn't have to
72 * flush them when they get mapped into an executable vm-area.
75 dma_mark_clean(void *addr, size_t size)
77 unsigned long pg_addr, end;
79 pg_addr = PAGE_ALIGN((unsigned long) addr);
80 end = (unsigned long) addr + size;
81 while (pg_addr + PAGE_SIZE <= end) {
82 struct page *page = virt_to_page(pg_addr);
83 set_bit(PG_arch_1, &page->flags);
89 ia64_set_rbs_bot (void)
91 unsigned long stack_size = rlimit_max(RLIMIT_STACK) & -16;
93 if (stack_size > MAX_USER_STACK_SIZE)
94 stack_size = MAX_USER_STACK_SIZE;
95 current->thread.rbs_bot = PAGE_ALIGN(current->mm->start_stack - stack_size);
99 * This performs some platform-dependent address space initialization.
100 * On IA-64, we want to setup the VM area for the register backing
101 * store (which grows upwards) and install the gateway page which is
102 * used for signal trampolines, etc.
105 ia64_init_addr_space (void)
107 struct vm_area_struct *vma;
112 * If we're out of memory and kmem_cache_alloc() returns NULL, we simply ignore
113 * the problem. When the process attempts to write to the register backing store
114 * for the first time, it will get a SEGFAULT in this case.
116 vma = vm_area_alloc(current->mm);
118 vma_set_anonymous(vma);
119 vma->vm_start = current->thread.rbs_bot & PAGE_MASK;
120 vma->vm_end = vma->vm_start + PAGE_SIZE;
121 vma->vm_flags = VM_DATA_DEFAULT_FLAGS|VM_GROWSUP|VM_ACCOUNT;
122 vma->vm_page_prot = vm_get_page_prot(vma->vm_flags);
123 down_write(¤t->mm->mmap_sem);
124 if (insert_vm_struct(current->mm, vma)) {
125 up_write(¤t->mm->mmap_sem);
129 up_write(¤t->mm->mmap_sem);
132 /* map NaT-page at address zero to speed up speculative dereferencing of NULL: */
133 if (!(current->personality & MMAP_PAGE_ZERO)) {
134 vma = vm_area_alloc(current->mm);
136 vma_set_anonymous(vma);
137 vma->vm_end = PAGE_SIZE;
138 vma->vm_page_prot = __pgprot(pgprot_val(PAGE_READONLY) | _PAGE_MA_NAT);
139 vma->vm_flags = VM_READ | VM_MAYREAD | VM_IO |
140 VM_DONTEXPAND | VM_DONTDUMP;
141 down_write(¤t->mm->mmap_sem);
142 if (insert_vm_struct(current->mm, vma)) {
143 up_write(¤t->mm->mmap_sem);
147 up_write(¤t->mm->mmap_sem);
155 free_reserved_area(ia64_imva(__init_begin), ia64_imva(__init_end),
156 -1, "unused kernel");
160 free_initrd_mem (unsigned long start, unsigned long end)
163 * EFI uses 4KB pages while the kernel can use 4KB or bigger.
164 * Thus EFI and the kernel may have different page sizes. It is
165 * therefore possible to have the initrd share the same page as
166 * the end of the kernel (given current setup).
168 * To avoid freeing/using the wrong page (kernel sized) we:
169 * - align up the beginning of initrd
170 * - align down the end of initrd
173 * |=============| a000
179 * |=============| 8000
182 * |/////////////| 7000
185 * |=============| 6000
188 * K=kernel using 8KB pages
190 * In this example, we must free page 8000 ONLY. So we must align up
191 * initrd_start and keep initrd_end as is.
193 start = PAGE_ALIGN(start);
194 end = end & PAGE_MASK;
197 printk(KERN_INFO "Freeing initrd memory: %ldkB freed\n", (end - start) >> 10);
199 for (; start < end; start += PAGE_SIZE) {
200 if (!virt_addr_valid(start))
202 free_reserved_page(virt_to_page(start));
207 * This installs a clean page in the kernel's page table.
209 static struct page * __init
210 put_kernel_page (struct page *page, unsigned long address, pgprot_t pgprot)
217 pgd = pgd_offset_k(address); /* note: this is NOT pgd_offset()! */
220 pud = pud_alloc(&init_mm, pgd, address);
223 pmd = pmd_alloc(&init_mm, pud, address);
226 pte = pte_alloc_kernel(pmd, address);
231 set_pte(pte, mk_pte(page, pgprot));
234 /* no need for flush_tlb */
244 * Map the gate page twice: once read-only to export the ELF
245 * headers etc. and once execute-only page to enable
246 * privilege-promotion via "epc":
248 page = virt_to_page(ia64_imva(__start_gate_section));
249 put_kernel_page(page, GATE_ADDR, PAGE_READONLY);
250 #ifdef HAVE_BUGGY_SEGREL
251 page = virt_to_page(ia64_imva(__start_gate_section + PAGE_SIZE));
252 put_kernel_page(page, GATE_ADDR + PAGE_SIZE, PAGE_GATE);
254 put_kernel_page(page, GATE_ADDR + PERCPU_PAGE_SIZE, PAGE_GATE);
255 /* Fill in the holes (if any) with read-only zero pages: */
259 for (addr = GATE_ADDR + PAGE_SIZE;
260 addr < GATE_ADDR + PERCPU_PAGE_SIZE;
263 put_kernel_page(ZERO_PAGE(0), addr,
265 put_kernel_page(ZERO_PAGE(0), addr + PERCPU_PAGE_SIZE,
273 static struct vm_area_struct gate_vma;
275 static int __init gate_vma_init(void)
277 vma_init(&gate_vma, NULL);
278 gate_vma.vm_start = FIXADDR_USER_START;
279 gate_vma.vm_end = FIXADDR_USER_END;
280 gate_vma.vm_flags = VM_READ | VM_MAYREAD | VM_EXEC | VM_MAYEXEC;
281 gate_vma.vm_page_prot = __P101;
285 __initcall(gate_vma_init);
287 struct vm_area_struct *get_gate_vma(struct mm_struct *mm)
292 int in_gate_area_no_mm(unsigned long addr)
294 if ((addr >= FIXADDR_USER_START) && (addr < FIXADDR_USER_END))
299 int in_gate_area(struct mm_struct *mm, unsigned long addr)
301 return in_gate_area_no_mm(addr);
304 void ia64_mmu_init(void *my_cpu_data)
306 unsigned long pta, impl_va_bits;
307 extern void tlb_init(void);
309 #ifdef CONFIG_DISABLE_VHPT
310 # define VHPT_ENABLE_BIT 0
312 # define VHPT_ENABLE_BIT 1
316 * Check if the virtually mapped linear page table (VMLPT) overlaps with a mapped
317 * address space. The IA-64 architecture guarantees that at least 50 bits of
318 * virtual address space are implemented but if we pick a large enough page size
319 * (e.g., 64KB), the mapped address space is big enough that it will overlap with
320 * VMLPT. I assume that once we run on machines big enough to warrant 64KB pages,
321 * IMPL_VA_MSB will be significantly bigger, so this is unlikely to become a
322 * problem in practice. Alternatively, we could truncate the top of the mapped
323 * address space to not permit mappings that would overlap with the VMLPT.
327 # define mapped_space_bits (3*(PAGE_SHIFT - pte_bits) + PAGE_SHIFT)
329 * The virtual page table has to cover the entire implemented address space within
330 * a region even though not all of this space may be mappable. The reason for
331 * this is that the Access bit and Dirty bit fault handlers perform
332 * non-speculative accesses to the virtual page table, so the address range of the
333 * virtual page table itself needs to be covered by virtual page table.
335 # define vmlpt_bits (impl_va_bits - PAGE_SHIFT + pte_bits)
336 # define POW2(n) (1ULL << (n))
338 impl_va_bits = ffz(~(local_cpu_data->unimpl_va_mask | (7UL << 61)));
340 if (impl_va_bits < 51 || impl_va_bits > 61)
341 panic("CPU has bogus IMPL_VA_MSB value of %lu!\n", impl_va_bits - 1);
343 * mapped_space_bits - PAGE_SHIFT is the total number of ptes we need,
344 * which must fit into "vmlpt_bits - pte_bits" slots. Second half of
345 * the test makes sure that our mapped space doesn't overlap the
346 * unimplemented hole in the middle of the region.
348 if ((mapped_space_bits - PAGE_SHIFT > vmlpt_bits - pte_bits) ||
349 (mapped_space_bits > impl_va_bits - 1))
350 panic("Cannot build a big enough virtual-linear page table"
351 " to cover mapped address space.\n"
352 " Try using a smaller page size.\n");
355 /* place the VMLPT at the end of each page-table mapped region: */
356 pta = POW2(61) - POW2(vmlpt_bits);
359 * Set the (virtually mapped linear) page table address. Bit
360 * 8 selects between the short and long format, bits 2-7 the
361 * size of the table, and bit 0 whether the VHPT walker is
364 ia64_set_pta(pta | (0 << 8) | (vmlpt_bits << 2) | VHPT_ENABLE_BIT);
368 #ifdef CONFIG_HUGETLB_PAGE
369 ia64_set_rr(HPAGE_REGION_BASE, HPAGE_SHIFT << 2);
374 #ifdef CONFIG_VIRTUAL_MEM_MAP
375 int vmemmap_find_next_valid_pfn(int node, int i)
377 unsigned long end_address, hole_next_pfn;
378 unsigned long stop_address;
379 pg_data_t *pgdat = NODE_DATA(node);
381 end_address = (unsigned long) &vmem_map[pgdat->node_start_pfn + i];
382 end_address = PAGE_ALIGN(end_address);
383 stop_address = (unsigned long) &vmem_map[pgdat_end_pfn(pgdat)];
391 pgd = pgd_offset_k(end_address);
392 if (pgd_none(*pgd)) {
393 end_address += PGDIR_SIZE;
397 pud = pud_offset(pgd, end_address);
398 if (pud_none(*pud)) {
399 end_address += PUD_SIZE;
403 pmd = pmd_offset(pud, end_address);
404 if (pmd_none(*pmd)) {
405 end_address += PMD_SIZE;
409 pte = pte_offset_kernel(pmd, end_address);
411 if (pte_none(*pte)) {
412 end_address += PAGE_SIZE;
414 if ((end_address < stop_address) &&
415 (end_address != ALIGN(end_address, 1UL << PMD_SHIFT)))
419 /* Found next valid vmem_map page */
421 } while (end_address < stop_address);
423 end_address = min(end_address, stop_address);
424 end_address = end_address - (unsigned long) vmem_map + sizeof(struct page) - 1;
425 hole_next_pfn = end_address / sizeof(struct page);
426 return hole_next_pfn - pgdat->node_start_pfn;
429 int __init create_mem_map_page_table(u64 start, u64 end, void *arg)
431 unsigned long address, start_page, end_page;
432 struct page *map_start, *map_end;
439 map_start = vmem_map + (__pa(start) >> PAGE_SHIFT);
440 map_end = vmem_map + (__pa(end) >> PAGE_SHIFT);
442 start_page = (unsigned long) map_start & PAGE_MASK;
443 end_page = PAGE_ALIGN((unsigned long) map_end);
444 node = paddr_to_nid(__pa(start));
446 for (address = start_page; address < end_page; address += PAGE_SIZE) {
447 pgd = pgd_offset_k(address);
449 pgd_populate(&init_mm, pgd, memblock_alloc_node(PAGE_SIZE, PAGE_SIZE, node));
450 pud = pud_offset(pgd, address);
453 pud_populate(&init_mm, pud, memblock_alloc_node(PAGE_SIZE, PAGE_SIZE, node));
454 pmd = pmd_offset(pud, address);
457 pmd_populate_kernel(&init_mm, pmd, memblock_alloc_node(PAGE_SIZE, PAGE_SIZE, node));
458 pte = pte_offset_kernel(pmd, address);
461 set_pte(pte, pfn_pte(__pa(memblock_alloc_node(PAGE_SIZE, PAGE_SIZE, node)) >> PAGE_SHIFT,
467 struct memmap_init_callback_data {
475 virtual_memmap_init(u64 start, u64 end, void *arg)
477 struct memmap_init_callback_data *args;
478 struct page *map_start, *map_end;
480 args = (struct memmap_init_callback_data *) arg;
481 map_start = vmem_map + (__pa(start) >> PAGE_SHIFT);
482 map_end = vmem_map + (__pa(end) >> PAGE_SHIFT);
484 if (map_start < args->start)
485 map_start = args->start;
486 if (map_end > args->end)
490 * We have to initialize "out of bounds" struct page elements that fit completely
491 * on the same pages that were allocated for the "in bounds" elements because they
492 * may be referenced later (and found to be "reserved").
494 map_start -= ((unsigned long) map_start & (PAGE_SIZE - 1)) / sizeof(struct page);
495 map_end += ((PAGE_ALIGN((unsigned long) map_end) - (unsigned long) map_end)
496 / sizeof(struct page));
498 if (map_start < map_end)
499 memmap_init_zone((unsigned long)(map_end - map_start),
500 args->nid, args->zone, page_to_pfn(map_start),
506 memmap_init (unsigned long size, int nid, unsigned long zone,
507 unsigned long start_pfn)
510 memmap_init_zone(size, nid, zone, start_pfn, MEMMAP_EARLY,
514 struct memmap_init_callback_data args;
516 start = pfn_to_page(start_pfn);
518 args.end = start + size;
522 efi_memmap_walk(virtual_memmap_init, &args);
527 ia64_pfn_valid (unsigned long pfn)
530 struct page *pg = pfn_to_page(pfn);
532 return (__get_user(byte, (char __user *) pg) == 0)
533 && ((((u64)pg & PAGE_MASK) == (((u64)(pg + 1) - 1) & PAGE_MASK))
534 || (__get_user(byte, (char __user *) (pg + 1) - 1) == 0));
536 EXPORT_SYMBOL(ia64_pfn_valid);
538 int __init find_largest_hole(u64 start, u64 end, void *arg)
542 static u64 last_end = PAGE_OFFSET;
544 /* NOTE: this algorithm assumes efi memmap table is ordered */
546 if (*max_gap < (start - last_end))
547 *max_gap = start - last_end;
552 #endif /* CONFIG_VIRTUAL_MEM_MAP */
554 int __init register_active_ranges(u64 start, u64 len, int nid)
556 u64 end = start + len;
559 if (start > crashk_res.start && start < crashk_res.end)
560 start = crashk_res.end;
561 if (end > crashk_res.start && end < crashk_res.end)
562 end = crashk_res.start;
566 memblock_add_node(__pa(start), end - start, nid);
571 find_max_min_low_pfn (u64 start, u64 end, void *arg)
573 unsigned long pfn_start, pfn_end;
574 #ifdef CONFIG_FLATMEM
575 pfn_start = (PAGE_ALIGN(__pa(start))) >> PAGE_SHIFT;
576 pfn_end = (PAGE_ALIGN(__pa(end - 1))) >> PAGE_SHIFT;
578 pfn_start = GRANULEROUNDDOWN(__pa(start)) >> PAGE_SHIFT;
579 pfn_end = GRANULEROUNDUP(__pa(end - 1)) >> PAGE_SHIFT;
581 min_low_pfn = min(min_low_pfn, pfn_start);
582 max_low_pfn = max(max_low_pfn, pfn_end);
587 * Boot command-line option "nolwsys" can be used to disable the use of any light-weight
588 * system call handler. When this option is in effect, all fsyscalls will end up bubbling
589 * down into the kernel and calling the normal (heavy-weight) syscall handler. This is
590 * useful for performance testing, but conceivably could also come in handy for debugging
594 static int nolwsys __initdata;
597 nolwsys_setup (char *s)
603 __setup("nolwsys", nolwsys_setup);
610 BUG_ON(PTRS_PER_PGD * sizeof(pgd_t) != PAGE_SIZE);
611 BUG_ON(PTRS_PER_PMD * sizeof(pmd_t) != PAGE_SIZE);
612 BUG_ON(PTRS_PER_PTE * sizeof(pte_t) != PAGE_SIZE);
616 * This needs to be called _after_ the command line has been parsed but _before_
617 * any drivers that may need the PCI DMA interface are initialized or bootmem has
623 #ifdef CONFIG_FLATMEM
627 set_max_mapnr(max_low_pfn);
628 high_memory = __va(max_low_pfn * PAGE_SIZE);
630 mem_init_print_info(NULL);
633 * For fsyscall entrpoints with no light-weight handler, use the ordinary
634 * (heavy-weight) handler, but mark it by setting bit 0, so the fsyscall entry
635 * code can tell them apart.
637 for (i = 0; i < NR_syscalls; ++i) {
638 extern unsigned long fsyscall_table[NR_syscalls];
639 extern unsigned long sys_call_table[NR_syscalls];
641 if (!fsyscall_table[i] || nolwsys)
642 fsyscall_table[i] = sys_call_table[i] | 1;
647 #ifdef CONFIG_MEMORY_HOTPLUG
648 int arch_add_memory(int nid, u64 start, u64 size, struct vmem_altmap *altmap,
651 unsigned long start_pfn = start >> PAGE_SHIFT;
652 unsigned long nr_pages = size >> PAGE_SHIFT;
655 ret = __add_pages(nid, start_pfn, nr_pages, altmap, want_memblock);
657 printk("%s: Problem encountered in __add_pages() as ret=%d\n",
663 #ifdef CONFIG_MEMORY_HOTREMOVE
664 int arch_remove_memory(u64 start, u64 size, struct vmem_altmap *altmap)
666 unsigned long start_pfn = start >> PAGE_SHIFT;
667 unsigned long nr_pages = size >> PAGE_SHIFT;
671 zone = page_zone(pfn_to_page(start_pfn));
672 ret = __remove_pages(zone, start_pfn, nr_pages, altmap);
674 pr_warn("%s: Problem encountered in __remove_pages() as"
675 " ret=%d\n", __func__, ret);