2 * Copyright (C) 1995 Linus Torvalds
3 * Copyright 2010 Tilera Corporation. All Rights Reserved.
5 * This program is free software; you can redistribute it and/or
6 * modify it under the terms of the GNU General Public License
7 * as published by the Free Software Foundation, version 2.
9 * This program is distributed in the hope that it will be useful, but
10 * WITHOUT ANY WARRANTY; without even the implied warranty of
11 * MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
12 * NON INFRINGEMENT. See the GNU General Public License for
16 #include <linux/module.h>
17 #include <linux/signal.h>
18 #include <linux/sched.h>
19 #include <linux/kernel.h>
20 #include <linux/errno.h>
21 #include <linux/string.h>
22 #include <linux/types.h>
23 #include <linux/ptrace.h>
24 #include <linux/mman.h>
26 #include <linux/hugetlb.h>
27 #include <linux/swap.h>
28 #include <linux/smp.h>
29 #include <linux/init.h>
30 #include <linux/highmem.h>
31 #include <linux/pagemap.h>
32 #include <linux/poison.h>
33 #include <linux/bootmem.h>
34 #include <linux/slab.h>
35 #include <linux/proc_fs.h>
36 #include <linux/efi.h>
37 #include <linux/memory_hotplug.h>
38 #include <linux/uaccess.h>
39 #include <asm/mmu_context.h>
40 #include <asm/processor.h>
41 #include <asm/pgtable.h>
42 #include <asm/pgalloc.h>
44 #include <asm/fixmap.h>
46 #include <asm/tlbflush.h>
47 #include <asm/sections.h>
48 #include <asm/setup.h>
49 #include <asm/homecache.h>
50 #include <hv/hypervisor.h>
51 #include <arch/chip.h>
55 #define clear_pgd(pmdptr) (*(pmdptr) = hv_pte(0))
58 unsigned long VMALLOC_RESERVE = CONFIG_VMALLOC_RESERVE;
59 EXPORT_SYMBOL(VMALLOC_RESERVE);
62 /* Create an L2 page table */
63 static pte_t * __init alloc_pte(void)
65 return __alloc_bootmem(L2_KERNEL_PGTABLE_SIZE, HV_PAGE_TABLE_ALIGN, 0);
69 * L2 page tables per controller. We allocate these all at once from
70 * the bootmem allocator and store them here. This saves on kernel L2
71 * page table memory, compared to allocating a full 64K page per L2
72 * page table, and also means that in cases where we use huge pages,
73 * we are guaranteed to later be able to shatter those huge pages and
74 * switch to using these page tables instead, without requiring
75 * further allocation. Each l2_ptes[] entry points to the first page
76 * table for the first hugepage-size piece of memory on the
77 * controller; other page tables are just indexed directly, i.e. the
78 * L2 page tables are contiguous in memory for each controller.
80 static pte_t *l2_ptes[MAX_NUMNODES];
81 static int num_l2_ptes[MAX_NUMNODES];
83 static void init_prealloc_ptes(int node, int pages)
85 BUG_ON(pages & (PTRS_PER_PTE - 1));
87 num_l2_ptes[node] = pages;
88 l2_ptes[node] = __alloc_bootmem(pages * sizeof(pte_t),
89 HV_PAGE_TABLE_ALIGN, 0);
93 pte_t *get_prealloc_pte(unsigned long pfn)
95 int node = pfn_to_nid(pfn);
96 pfn &= ~(-1UL << (NR_PA_HIGHBIT_SHIFT - PAGE_SHIFT));
97 BUG_ON(node >= MAX_NUMNODES);
98 BUG_ON(pfn >= num_l2_ptes[node]);
99 return &l2_ptes[node][pfn];
103 * What caching do we expect pages from the heap to have when
104 * they are allocated during bootup? (Once we've installed the
105 * "real" swapper_pg_dir.)
107 static int initial_heap_home(void)
110 return PAGE_HOME_HASH;
111 return smp_processor_id();
115 * Place a pointer to an L2 page table in a middle page
118 static void __init assign_pte(pmd_t *pmd, pte_t *page_table)
120 phys_addr_t pa = __pa(page_table);
121 unsigned long l2_ptfn = pa >> HV_LOG2_PAGE_TABLE_ALIGN;
122 pte_t pteval = hv_pte_set_ptfn(__pgprot(_PAGE_TABLE), l2_ptfn);
123 BUG_ON((pa & (HV_PAGE_TABLE_ALIGN-1)) != 0);
124 pteval = pte_set_home(pteval, initial_heap_home());
125 *(pte_t *)pmd = pteval;
126 if (page_table != (pte_t *)pmd_page_vaddr(*pmd))
132 static inline pmd_t *alloc_pmd(void)
134 return __alloc_bootmem(L1_KERNEL_PGTABLE_SIZE, HV_PAGE_TABLE_ALIGN, 0);
137 static inline void assign_pmd(pud_t *pud, pmd_t *pmd)
139 assign_pte((pmd_t *)pud, (pte_t *)pmd);
142 #endif /* __tilegx__ */
144 /* Replace the given pmd with a full PTE table. */
145 void __init shatter_pmd(pmd_t *pmd)
147 pte_t *pte = get_prealloc_pte(pte_pfn(*(pte_t *)pmd));
148 assign_pte(pmd, pte);
152 static pmd_t *__init get_pmd(pgd_t pgtables[], unsigned long va)
154 pud_t *pud = pud_offset(&pgtables[pgd_index(va)], va);
156 assign_pmd(pud, alloc_pmd());
157 return pmd_offset(pud, va);
160 static pmd_t *__init get_pmd(pgd_t pgtables[], unsigned long va)
162 return pmd_offset(pud_offset(&pgtables[pgd_index(va)], va), va);
167 * This function initializes a certain range of kernel virtual memory
168 * with new bootmem page tables, everywhere page tables are missing in
173 * NOTE: The pagetables are allocated contiguous on the physical space
174 * so we can cache the place of the first one and move around without
175 * checking the pgd every time.
177 static void __init page_table_range_init(unsigned long start,
178 unsigned long end, pgd_t *pgd)
181 start = round_down(start, PMD_SIZE);
182 end = round_up(end, PMD_SIZE);
183 for (vaddr = start; vaddr < end; vaddr += PMD_SIZE) {
184 pmd_t *pmd = get_pmd(pgd, vaddr);
186 assign_pte(pmd, alloc_pte());
191 static int __initdata ktext_hash = 1; /* .text pages */
192 static int __initdata kdata_hash = 1; /* .data and .bss pages */
193 int __write_once hash_default = 1; /* kernel allocator pages */
194 EXPORT_SYMBOL(hash_default);
195 int __write_once kstack_hash = 1; /* if no homecaching, use h4h */
198 * CPUs to use to for striping the pages of kernel data. If hash-for-home
199 * is available, this is only relevant if kcache_hash sets up the
200 * .data and .bss to be page-homed, and we don't want the default mode
201 * of using the full set of kernel cpus for the striping.
203 static __initdata struct cpumask kdata_mask;
204 static __initdata int kdata_arg_seen;
206 int __write_once kdata_huge; /* if no homecaching, small pages */
209 /* Combine a generic pgprot_t with cache home to get a cache-aware pgprot. */
210 static pgprot_t __init construct_pgprot(pgprot_t prot, int home)
212 prot = pte_set_home(prot, home);
213 if (home == PAGE_HOME_IMMUTABLE) {
215 prot = hv_pte_set_mode(prot, HV_PTE_MODE_CACHE_HASH_L3);
217 prot = hv_pte_set_mode(prot, HV_PTE_MODE_CACHE_NO_L3);
223 * For a given kernel data VA, how should it be cached?
224 * We return the complete pgprot_t with caching bits set.
226 static pgprot_t __init init_pgprot(ulong address)
230 enum { CODE_DELTA = MEM_SV_START - PAGE_OFFSET };
232 /* For kdata=huge, everything is just hash-for-home. */
234 return construct_pgprot(PAGE_KERNEL, PAGE_HOME_HASH);
237 * We map the aliased pages of permanent text so we can
238 * update them if necessary, for ftrace, etc.
240 if (address < (ulong) _sinittext - CODE_DELTA)
241 return construct_pgprot(PAGE_KERNEL, PAGE_HOME_HASH);
243 /* We map read-only data non-coherent for performance. */
244 if ((address >= (ulong) __start_rodata &&
245 address < (ulong) __end_rodata) ||
246 address == (ulong) empty_zero_page) {
247 return construct_pgprot(PAGE_KERNEL_RO, PAGE_HOME_IMMUTABLE);
251 /* Force the atomic_locks[] array page to be hash-for-home. */
252 if (address == (ulong) atomic_locks)
253 return construct_pgprot(PAGE_KERNEL, PAGE_HOME_HASH);
257 * Everything else that isn't data or bss is heap, so mark it
258 * with the initial heap home (hash-for-home, or this cpu). This
259 * includes any addresses after the loaded image and any address before
260 * __init_end, since we already captured the case of text before
261 * _sinittext, and __pa(einittext) is approximately __pa(__init_begin).
263 * All the LOWMEM pages that we mark this way will get their
264 * struct page homecache properly marked later, in set_page_homes().
265 * The HIGHMEM pages we leave with a default zero for their
266 * homes, but with a zero free_time we don't have to actually
267 * do a flush action the first time we use them, either.
269 if (address >= (ulong) _end || address < (ulong) __init_end)
270 return construct_pgprot(PAGE_KERNEL, initial_heap_home());
272 /* Use hash-for-home if requested for data/bss. */
274 return construct_pgprot(PAGE_KERNEL, PAGE_HOME_HASH);
277 * Otherwise we just hand out consecutive cpus. To avoid
278 * requiring this function to hold state, we just walk forward from
279 * __end_rodata by PAGE_SIZE, skipping the readonly and init data, to
280 * reach the requested address, while walking cpu home around
281 * kdata_mask. This is typically no more than a dozen or so iterations.
283 page = (((ulong)__end_rodata) + PAGE_SIZE - 1) & PAGE_MASK;
284 BUG_ON(address < page || address >= (ulong)_end);
285 cpu = cpumask_first(&kdata_mask);
286 for (; page < address; page += PAGE_SIZE) {
287 if (page >= (ulong)&init_thread_union &&
288 page < (ulong)&init_thread_union + THREAD_SIZE)
290 if (page == (ulong)empty_zero_page)
293 if (page == (ulong)atomic_locks)
296 cpu = cpumask_next(cpu, &kdata_mask);
298 cpu = cpumask_first(&kdata_mask);
300 return construct_pgprot(PAGE_KERNEL, cpu);
304 * This function sets up how we cache the kernel text. If we have
305 * hash-for-home support, normally that is used instead (see the
306 * kcache_hash boot flag for more information). But if we end up
307 * using a page-based caching technique, this option sets up the
308 * details of that. In addition, the "ktext=nocache" option may
309 * always be used to disable local caching of text pages, if desired.
312 static int __initdata ktext_arg_seen;
313 static int __initdata ktext_small;
314 static int __initdata ktext_local;
315 static int __initdata ktext_all;
316 static int __initdata ktext_nondataplane;
317 static int __initdata ktext_nocache;
318 static struct cpumask __initdata ktext_mask;
320 static int __init setup_ktext(char *str)
325 /* If you have a leading "nocache", turn off ktext caching */
326 if (strncmp(str, "nocache", 7) == 0) {
328 pr_info("ktext: disabling local caching of kernel text\n");
338 /* Default setting: use a huge page */
339 if (strcmp(str, "huge") == 0)
340 pr_info("ktext: using one huge locally cached page\n");
342 /* Pay TLB cost but get no cache benefit: cache small pages locally */
343 else if (strcmp(str, "local") == 0) {
346 pr_info("ktext: using small pages with local caching\n");
349 /* Neighborhood cache ktext pages on all cpus. */
350 else if (strcmp(str, "all") == 0) {
353 pr_info("ktext: using maximal caching neighborhood\n");
357 /* Neighborhood ktext pages on specified mask */
358 else if (cpulist_parse(str, &ktext_mask) == 0) {
359 if (cpumask_weight(&ktext_mask) > 1) {
361 pr_info("ktext: using caching neighborhood %*pbl with small pages\n",
362 cpumask_pr_args(&ktext_mask));
364 pr_info("ktext: caching on cpu %*pbl with one huge page\n",
365 cpumask_pr_args(&ktext_mask));
375 early_param("ktext", setup_ktext);
378 static inline pgprot_t ktext_set_nocache(pgprot_t prot)
381 prot = hv_pte_set_nc(prot);
383 prot = hv_pte_set_no_alloc_l2(prot);
387 /* Temporary page table we use for staging. */
388 static pgd_t pgtables[PTRS_PER_PGD]
389 __attribute__((aligned(HV_PAGE_TABLE_ALIGN)));
392 * This maps the physical memory to kernel virtual address space, a total
393 * of max_low_pfn pages, by creating page tables starting from address
396 * This routine transitions us from using a set of compiled-in large
397 * pages to using some more precise caching, including removing access
398 * to code pages mapped at PAGE_OFFSET (executed only at MEM_SV_START)
399 * marking read-only data as locally cacheable, striping the remaining
400 * .data and .bss across all the available tiles, and removing access
401 * to pages above the top of RAM (thus ensuring a page fault from a bad
402 * virtual address rather than a hypervisor shoot down for accessing
403 * memory outside the assigned limits).
405 static void __init kernel_physical_mapping_init(pgd_t *pgd_base)
407 unsigned long long irqmask;
408 unsigned long address, pfn;
412 const struct cpumask *my_cpu_mask = cpumask_of(smp_processor_id());
413 struct cpumask kstripe_mask;
416 if (ktext_arg_seen && ktext_hash) {
417 pr_warn("warning: \"ktext\" boot argument ignored if \"kcache_hash\" sets up text hash-for-home\n");
421 if (kdata_arg_seen && kdata_hash) {
422 pr_warn("warning: \"kdata\" boot argument ignored if \"kcache_hash\" sets up data hash-for-home\n");
425 if (kdata_huge && !hash_default) {
426 pr_warn("warning: disabling \"kdata=huge\"; requires kcache_hash=all or =allbutstack\n");
431 * Set up a mask for cpus to use for kernel striping.
432 * This is normally all cpus, but minus dataplane cpus if any.
433 * If the dataplane covers the whole chip, we stripe over
434 * the whole chip too.
436 cpumask_copy(&kstripe_mask, cpu_possible_mask);
438 kdata_mask = kstripe_mask;
440 /* Allocate and fill in L2 page tables */
441 for (i = 0; i < MAX_NUMNODES; ++i) {
442 #ifdef CONFIG_HIGHMEM
443 unsigned long end_pfn = node_lowmem_end_pfn[i];
445 unsigned long end_pfn = node_end_pfn[i];
447 unsigned long end_huge_pfn = 0;
449 /* Pre-shatter the last huge page to allow per-cpu pages. */
451 end_huge_pfn = end_pfn - (HPAGE_SIZE >> PAGE_SHIFT);
453 pfn = node_start_pfn[i];
455 /* Allocate enough memory to hold L2 page tables for node. */
456 init_prealloc_ptes(i, end_pfn - pfn);
458 address = (unsigned long) pfn_to_kaddr(pfn);
459 while (pfn < end_pfn) {
460 BUG_ON(address & (HPAGE_SIZE-1));
461 pmd = get_pmd(pgtables, address);
462 pte = get_prealloc_pte(pfn);
463 if (pfn < end_huge_pfn) {
464 pgprot_t prot = init_pgprot(address);
465 *(pte_t *)pmd = pte_mkhuge(pfn_pte(pfn, prot));
466 for (pte_ofs = 0; pte_ofs < PTRS_PER_PTE;
467 pfn++, pte_ofs++, address += PAGE_SIZE)
468 pte[pte_ofs] = pfn_pte(pfn, prot);
471 printk(KERN_DEBUG "pre-shattered huge page at %#lx\n",
473 for (pte_ofs = 0; pte_ofs < PTRS_PER_PTE;
474 pfn++, pte_ofs++, address += PAGE_SIZE) {
475 pgprot_t prot = init_pgprot(address);
476 pte[pte_ofs] = pfn_pte(pfn, prot);
478 assign_pte(pmd, pte);
484 * Set or check ktext_map now that we have cpu_possible_mask
485 * and kstripe_mask to work with.
488 cpumask_copy(&ktext_mask, cpu_possible_mask);
489 else if (ktext_nondataplane)
490 ktext_mask = kstripe_mask;
491 else if (!cpumask_empty(&ktext_mask)) {
492 /* Sanity-check any mask that was requested */
494 cpumask_andnot(&bad, &ktext_mask, cpu_possible_mask);
495 cpumask_and(&ktext_mask, &ktext_mask, cpu_possible_mask);
496 if (!cpumask_empty(&bad))
497 pr_info("ktext: not using unavailable cpus %*pbl\n",
498 cpumask_pr_args(&bad));
499 if (cpumask_empty(&ktext_mask)) {
500 pr_warn("ktext: no valid cpus; caching on %d\n",
502 cpumask_copy(&ktext_mask,
503 cpumask_of(smp_processor_id()));
507 address = MEM_SV_START;
508 pmd = get_pmd(pgtables, address);
509 pfn = 0; /* code starts at PA 0 */
511 /* Allocate an L2 PTE for the kernel text */
513 pgprot_t prot = construct_pgprot(PAGE_KERNEL_EXEC,
514 PAGE_HOME_IMMUTABLE);
518 prot = hv_pte_set_mode(prot,
519 HV_PTE_MODE_UNCACHED);
521 prot = hv_pte_set_mode(prot,
522 HV_PTE_MODE_CACHE_NO_L3);
524 prot = hv_pte_set_mode(prot,
525 HV_PTE_MODE_CACHE_TILE_L3);
526 cpu = cpumask_first(&ktext_mask);
528 prot = ktext_set_nocache(prot);
531 BUG_ON(address != (unsigned long)_text);
533 for (; address < (unsigned long)_einittext;
534 pfn++, address += PAGE_SIZE) {
535 pte_ofs = pte_index(address);
538 assign_pte(pmd++, pte);
542 prot = set_remote_cache_cpu(prot, cpu);
543 cpu = cpumask_next(cpu, &ktext_mask);
545 cpu = cpumask_first(&ktext_mask);
547 pte[pte_ofs] = pfn_pte(pfn, prot);
550 assign_pte(pmd, pte);
552 pte_t pteval = pfn_pte(0, PAGE_KERNEL_EXEC);
553 pteval = pte_mkhuge(pteval);
555 pteval = hv_pte_set_mode(pteval,
556 HV_PTE_MODE_CACHE_HASH_L3);
557 pteval = ktext_set_nocache(pteval);
559 if (cpumask_weight(&ktext_mask) == 1) {
560 pteval = set_remote_cache_cpu(pteval,
561 cpumask_first(&ktext_mask));
562 pteval = hv_pte_set_mode(pteval,
563 HV_PTE_MODE_CACHE_TILE_L3);
564 pteval = ktext_set_nocache(pteval);
565 } else if (ktext_nocache)
566 pteval = hv_pte_set_mode(pteval,
567 HV_PTE_MODE_UNCACHED);
569 pteval = hv_pte_set_mode(pteval,
570 HV_PTE_MODE_CACHE_NO_L3);
571 for (; address < (unsigned long)_einittext;
572 pfn += PFN_DOWN(HPAGE_SIZE), address += HPAGE_SIZE)
573 *(pte_t *)(pmd++) = pfn_pte(pfn, pteval);
576 /* Set swapper_pgprot here so it is flushed to memory right away. */
577 swapper_pgprot = init_pgprot((unsigned long)swapper_pg_dir);
580 * Since we may be changing the caching of the stack and page
581 * table itself, we invoke an assembly helper to do the
584 * - flush the cache so we start with an empty slate
585 * - install pgtables[] as the real page table
586 * - flush the TLB so the new page table takes effect
588 irqmask = interrupt_mask_save_mask();
589 interrupt_mask_set_mask(-1ULL);
590 rc = flush_and_install_context(__pa(pgtables),
591 init_pgprot((unsigned long)pgtables),
592 __this_cpu_read(current_asid),
593 cpumask_bits(my_cpu_mask));
594 interrupt_mask_restore_mask(irqmask);
597 /* Copy the page table back to the normal swapper_pg_dir. */
598 memcpy(pgd_base, pgtables, sizeof(pgtables));
599 __install_page_table(pgd_base, __this_cpu_read(current_asid),
603 * We just read swapper_pgprot and thus brought it into the cache,
604 * with its new home & caching mode. When we start the other CPUs,
605 * they're going to reference swapper_pgprot via their initial fake
606 * VA-is-PA mappings, which cache everything locally. At that
607 * time, if it's in our cache with a conflicting home, the
608 * simulator's coherence checker will complain. So, flush it out
609 * of our cache; we're not going to ever use it again anyway.
611 __insn_finv(&swapper_pgprot);
615 * devmem_is_allowed() checks to see if /dev/mem access to a certain address
616 * is valid. The argument is a physical page number.
618 * On Tile, the only valid things for which we can just hand out unchecked
619 * PTEs are the kernel code and data. Anything else might change its
620 * homing with time, and we wouldn't know to adjust the /dev/mem PTEs.
621 * Note that init_thread_union is released to heap soon after boot,
622 * so we include it in the init data.
624 * For TILE-Gx, we might want to consider allowing access to PA
625 * regions corresponding to PCI space, etc.
627 int devmem_is_allowed(unsigned long pagenr)
629 return pagenr < kaddr_to_pfn(_end) &&
630 !(pagenr >= kaddr_to_pfn(&init_thread_union) ||
631 pagenr < kaddr_to_pfn(__init_end)) &&
632 !(pagenr >= kaddr_to_pfn(_sinittext) ||
633 pagenr <= kaddr_to_pfn(_einittext-1));
636 #ifdef CONFIG_HIGHMEM
637 static void __init permanent_kmaps_init(pgd_t *pgd_base)
646 page_table_range_init(vaddr, vaddr + PAGE_SIZE*LAST_PKMAP, pgd_base);
648 pgd = swapper_pg_dir + pgd_index(vaddr);
649 pud = pud_offset(pgd, vaddr);
650 pmd = pmd_offset(pud, vaddr);
651 pte = pte_offset_kernel(pmd, vaddr);
652 pkmap_page_table = pte;
654 #endif /* CONFIG_HIGHMEM */
658 static void __init init_free_pfn_range(unsigned long start, unsigned long end)
661 struct page *page = pfn_to_page(start);
663 for (pfn = start; pfn < end; ) {
664 /* Optimize by freeing pages in large batches */
665 int order = __ffs(pfn);
669 if (order >= MAX_ORDER)
672 while (pfn + count > end) {
676 for (p = page, i = 0; i < count; ++i, ++p) {
677 __ClearPageReserved(p);
679 * Hacky direct set to avoid unnecessary
680 * lock take/release for EVERY page here.
682 p->_refcount.counter = 0;
683 p->_mapcount.counter = -1;
685 init_page_count(page);
686 __free_pages(page, order);
687 adjust_managed_page_count(page, count);
694 static void __init set_non_bootmem_pages_init(void)
698 unsigned long start, end;
699 int nid = z->zone_pgdat->node_id;
700 #ifdef CONFIG_HIGHMEM
701 int idx = zone_idx(z);
704 start = z->zone_start_pfn;
705 end = start + z->spanned_pages;
706 start = max(start, node_free_pfn[nid]);
707 start = max(start, max_low_pfn);
709 #ifdef CONFIG_HIGHMEM
710 if (idx == ZONE_HIGHMEM)
711 totalhigh_pages += z->spanned_pages;
714 unsigned long percpu_pfn = node_percpu_pfn[nid];
715 if (start < percpu_pfn && end > percpu_pfn)
719 if (start <= pci_reserve_start_pfn &&
720 end > pci_reserve_start_pfn) {
721 if (end > pci_reserve_end_pfn)
722 init_free_pfn_range(pci_reserve_end_pfn, end);
723 end = pci_reserve_start_pfn;
726 init_free_pfn_range(start, end);
732 * paging_init() sets up the page tables - note that all of lowmem is
733 * already mapped by head.S.
735 void __init paging_init(void)
740 pgd_t *pgd_base = swapper_pg_dir;
742 kernel_physical_mapping_init(pgd_base);
744 /* Fixed mappings, only the page table structure has to be created. */
745 page_table_range_init(fix_to_virt(__end_of_fixed_addresses - 1),
746 FIXADDR_TOP, pgd_base);
748 #ifdef CONFIG_HIGHMEM
749 permanent_kmaps_init(pgd_base);
754 * Since GX allocates just one pmd_t array worth of vmalloc space,
755 * we go ahead and allocate it statically here, then share it
756 * globally. As a result we don't have to worry about any task
757 * changing init_mm once we get up and running, and there's no
758 * need for e.g. vmalloc_sync_all().
760 BUILD_BUG_ON(pgd_index(VMALLOC_START) != pgd_index(VMALLOC_END - 1));
761 pud = pud_offset(pgd_base + pgd_index(VMALLOC_START), VMALLOC_START);
762 assign_pmd(pud, alloc_pmd());
768 * Walk the kernel page tables and derive the page_home() from
769 * the PTEs, so that set_pte() can properly validate the caching
770 * of all PTEs it sees.
772 void __init set_page_homes(void)
776 static void __init set_max_mapnr_init(void)
778 #ifdef CONFIG_FLATMEM
779 max_mapnr = max_low_pfn;
783 void __init mem_init(void)
790 #ifdef CONFIG_FLATMEM
794 #ifdef CONFIG_HIGHMEM
795 /* check that fixmap and pkmap do not overlap */
796 if (PKMAP_ADDR(LAST_PKMAP-1) >= FIXADDR_START) {
797 pr_err("fixmap and kmap areas overlap - this will crash\n");
798 pr_err("pkstart: %lxh pkend: %lxh fixstart %lxh\n",
799 PKMAP_BASE, PKMAP_ADDR(LAST_PKMAP-1), FIXADDR_START);
804 set_max_mapnr_init();
806 /* this will put all bootmem onto the freelists */
810 /* count all remaining LOWMEM and give all HIGHMEM to page allocator */
811 set_non_bootmem_pages_init();
814 mem_init_print_info(NULL);
817 * In debug mode, dump some interesting memory mappings.
819 #ifdef CONFIG_HIGHMEM
820 printk(KERN_DEBUG " KMAP %#lx - %#lx\n",
821 FIXADDR_START, FIXADDR_TOP + PAGE_SIZE - 1);
822 printk(KERN_DEBUG " PKMAP %#lx - %#lx\n",
823 PKMAP_BASE, PKMAP_ADDR(LAST_PKMAP) - 1);
825 printk(KERN_DEBUG " VMALLOC %#lx - %#lx\n",
826 _VMALLOC_START, _VMALLOC_END - 1);
828 for (i = MAX_NUMNODES-1; i >= 0; --i) {
829 struct pglist_data *node = &node_data[i];
830 if (node->node_present_pages) {
831 unsigned long start = (unsigned long)
832 pfn_to_kaddr(node->node_start_pfn);
833 unsigned long end = start +
834 (node->node_present_pages << PAGE_SHIFT);
835 printk(KERN_DEBUG " MEM%d %#lx - %#lx\n",
841 for (i = MAX_NUMNODES-1; i >= 0; --i) {
842 if ((unsigned long)vbase_map[i] != -1UL) {
843 printk(KERN_DEBUG " LOWMEM%d %#lx - %#lx\n",
844 i, (unsigned long) (vbase_map[i]),
845 (unsigned long) (last-1));
853 * Convert from using one lock for all atomic operations to
856 __init_atomic_per_cpu();
861 * this is for the non-NUMA, single node SMP system case.
862 * Specifically, in the case of x86, we will always add
863 * memory to the highmem for now.
865 #ifndef CONFIG_NEED_MULTIPLE_NODES
866 int arch_add_memory(u64 start, u64 size, bool for_device)
868 struct pglist_data *pgdata = &contig_page_data;
869 struct zone *zone = pgdata->node_zones + MAX_NR_ZONES-1;
870 unsigned long start_pfn = start >> PAGE_SHIFT;
871 unsigned long nr_pages = size >> PAGE_SHIFT;
873 return __add_pages(zone, start_pfn, nr_pages);
876 int remove_memory(u64 start, u64 size)
881 #ifdef CONFIG_MEMORY_HOTREMOVE
882 int arch_remove_memory(u64 start, u64 size)
890 struct kmem_cache *pgd_cache;
892 void __init pgtable_cache_init(void)
894 pgd_cache = kmem_cache_create("pgd", SIZEOF_PGD, SIZEOF_PGD, 0, NULL);
896 panic("pgtable_cache_init(): Cannot create pgd cache");
899 static long __write_once initfree = 1;
900 static bool __write_once set_initfree_done;
902 /* Select whether to free (1) or mark unusable (0) the __init pages. */
903 static int __init set_initfree(char *str)
906 if (kstrtol(str, 0, &val) == 0) {
907 set_initfree_done = true;
909 pr_info("initfree: %s free init pages\n",
910 initfree ? "will" : "won't");
914 __setup("initfree=", set_initfree);
916 static void free_init_pages(char *what, unsigned long begin, unsigned long end)
918 unsigned long addr = (unsigned long) begin;
920 /* Prefer user request first */
921 if (!set_initfree_done) {
922 if (debug_pagealloc_enabled())
925 if (kdata_huge && !initfree) {
926 pr_warn("Warning: ignoring initfree=0: incompatible with kdata=huge\n");
929 end = (end + PAGE_SIZE - 1) & PAGE_MASK;
930 local_flush_tlb_pages(NULL, begin, PAGE_SIZE, end - begin);
931 for (addr = begin; addr < end; addr += PAGE_SIZE) {
933 * Note we just reset the home here directly in the
934 * page table. We know this is safe because our caller
935 * just flushed the caches on all the other cpus,
936 * and they won't be touching any of these pages.
938 int pfn = kaddr_to_pfn((void *)addr);
939 struct page *page = pfn_to_page(pfn);
940 pte_t *ptep = virt_to_kpte(addr);
943 * If debugging page accesses then do not free
944 * this memory but mark them not present - any
945 * buggy init-section access will create a
948 pte_clear(&init_mm, addr, ptep);
954 set_pte_at(&init_mm, addr, ptep,
955 pfn_pte(pfn, PAGE_KERNEL));
956 memset((void *)addr, POISON_FREE_INITMEM, PAGE_SIZE);
957 free_reserved_page(page);
959 pr_info("Freeing %s: %ldk freed\n", what, (end - begin) >> 10);
962 void free_initmem(void)
964 const unsigned long text_delta = MEM_SV_START - PAGE_OFFSET;
967 * Evict the cache on all cores to avoid incoherence.
968 * We are guaranteed that no one will touch the init pages any more.
970 homecache_evict(&cpu_cacheable_map);
972 /* Free the data pages that we won't use again after init. */
973 free_init_pages("unused kernel data",
974 (unsigned long)__init_begin,
975 (unsigned long)__init_end);
978 * Free the pages mapped from 0xc0000000 that correspond to code
979 * pages from MEM_SV_START that we won't use again after init.
981 free_init_pages("unused kernel text",
982 (unsigned long)_sinittext - text_delta,
983 (unsigned long)_einittext - text_delta);
984 /* Do a global TLB flush so everyone sees the changes. */