2 * arch/sparc64/mm/init.c
4 * Copyright (C) 1996-1999 David S. Miller (davem@caip.rutgers.edu)
5 * Copyright (C) 1997-1999 Jakub Jelinek (jj@sunsite.mff.cuni.cz)
8 #include <linux/extable.h>
9 #include <linux/kernel.h>
10 #include <linux/sched.h>
11 #include <linux/string.h>
12 #include <linux/init.h>
13 #include <linux/bootmem.h>
15 #include <linux/hugetlb.h>
16 #include <linux/initrd.h>
17 #include <linux/swap.h>
18 #include <linux/pagemap.h>
19 #include <linux/poison.h>
21 #include <linux/seq_file.h>
22 #include <linux/kprobes.h>
23 #include <linux/cache.h>
24 #include <linux/sort.h>
25 #include <linux/ioport.h>
26 #include <linux/percpu.h>
27 #include <linux/memblock.h>
28 #include <linux/mmzone.h>
29 #include <linux/gfp.h>
33 #include <asm/pgalloc.h>
34 #include <asm/pgtable.h>
35 #include <asm/oplib.h>
36 #include <asm/iommu.h>
38 #include <linux/uaccess.h>
39 #include <asm/mmu_context.h>
40 #include <asm/tlbflush.h>
42 #include <asm/starfire.h>
44 #include <asm/spitfire.h>
45 #include <asm/sections.h>
47 #include <asm/hypervisor.h>
49 #include <asm/mdesc.h>
50 #include <asm/cpudata.h>
51 #include <asm/setup.h>
56 unsigned long kern_linear_pte_xor[4] __read_mostly;
57 static unsigned long page_cache4v_flag;
59 /* A bitmap, two bits for every 256MB of physical memory. These two
60 * bits determine what page size we use for kernel linear
61 * translations. They form an index into kern_linear_pte_xor[]. The
62 * value in the indexed slot is XOR'd with the TLB miss virtual
63 * address to form the resulting TTE. The mapping is:
70 * All sun4v chips support 256MB pages. Only SPARC-T4 and later
71 * support 2GB pages, and hopefully future cpus will support the 16GB
72 * pages as well. For slots 2 and 3, we encode a 256MB TTE xor there
73 * if these larger page sizes are not supported by the cpu.
75 * It would be nice to determine this from the machine description
76 * 'cpu' properties, but we need to have this table setup before the
77 * MDESC is initialized.
80 #ifndef CONFIG_DEBUG_PAGEALLOC
81 /* A special kernel TSB for 4MB, 256MB, 2GB and 16GB linear mappings.
82 * Space is allocated for this right after the trap table in
83 * arch/sparc64/kernel/head.S
85 extern struct tsb swapper_4m_tsb[KERNEL_TSB4M_NENTRIES];
87 extern struct tsb swapper_tsb[KERNEL_TSB_NENTRIES];
89 static unsigned long cpu_pgsz_mask;
91 #define MAX_BANKS 1024
93 static struct linux_prom64_registers pavail[MAX_BANKS];
94 static int pavail_ents;
96 u64 numa_latency[MAX_NUMNODES][MAX_NUMNODES];
98 static int cmp_p64(const void *a, const void *b)
100 const struct linux_prom64_registers *x = a, *y = b;
102 if (x->phys_addr > y->phys_addr)
104 if (x->phys_addr < y->phys_addr)
109 static void __init read_obp_memory(const char *property,
110 struct linux_prom64_registers *regs,
113 phandle node = prom_finddevice("/memory");
114 int prop_size = prom_getproplen(node, property);
117 ents = prop_size / sizeof(struct linux_prom64_registers);
118 if (ents > MAX_BANKS) {
119 prom_printf("The machine has more %s property entries than "
120 "this kernel can support (%d).\n",
121 property, MAX_BANKS);
125 ret = prom_getproperty(node, property, (char *) regs, prop_size);
127 prom_printf("Couldn't get %s property from /memory.\n",
132 /* Sanitize what we got from the firmware, by page aligning
135 for (i = 0; i < ents; i++) {
136 unsigned long base, size;
138 base = regs[i].phys_addr;
139 size = regs[i].reg_size;
142 if (base & ~PAGE_MASK) {
143 unsigned long new_base = PAGE_ALIGN(base);
145 size -= new_base - base;
146 if ((long) size < 0L)
151 /* If it is empty, simply get rid of it.
152 * This simplifies the logic of the other
153 * functions that process these arrays.
155 memmove(®s[i], ®s[i + 1],
156 (ents - i - 1) * sizeof(regs[0]));
161 regs[i].phys_addr = base;
162 regs[i].reg_size = size;
167 sort(regs, ents, sizeof(struct linux_prom64_registers),
171 /* Kernel physical address base and size in bytes. */
172 unsigned long kern_base __read_mostly;
173 unsigned long kern_size __read_mostly;
175 /* Initial ramdisk setup */
176 extern unsigned long sparc_ramdisk_image64;
177 extern unsigned int sparc_ramdisk_image;
178 extern unsigned int sparc_ramdisk_size;
180 struct page *mem_map_zero __read_mostly;
181 EXPORT_SYMBOL(mem_map_zero);
183 unsigned int sparc64_highest_unlocked_tlb_ent __read_mostly;
185 unsigned long sparc64_kern_pri_context __read_mostly;
186 unsigned long sparc64_kern_pri_nuc_bits __read_mostly;
187 unsigned long sparc64_kern_sec_context __read_mostly;
189 int num_kernel_image_mappings;
191 #ifdef CONFIG_DEBUG_DCFLUSH
192 atomic_t dcpage_flushes = ATOMIC_INIT(0);
194 atomic_t dcpage_flushes_xcall = ATOMIC_INIT(0);
198 inline void flush_dcache_page_impl(struct page *page)
200 BUG_ON(tlb_type == hypervisor);
201 #ifdef CONFIG_DEBUG_DCFLUSH
202 atomic_inc(&dcpage_flushes);
205 #ifdef DCACHE_ALIASING_POSSIBLE
206 __flush_dcache_page(page_address(page),
207 ((tlb_type == spitfire) &&
208 page_mapping(page) != NULL));
210 if (page_mapping(page) != NULL &&
211 tlb_type == spitfire)
212 __flush_icache_page(__pa(page_address(page)));
216 #define PG_dcache_dirty PG_arch_1
217 #define PG_dcache_cpu_shift 32UL
218 #define PG_dcache_cpu_mask \
219 ((1UL<<ilog2(roundup_pow_of_two(NR_CPUS)))-1UL)
221 #define dcache_dirty_cpu(page) \
222 (((page)->flags >> PG_dcache_cpu_shift) & PG_dcache_cpu_mask)
224 static inline void set_dcache_dirty(struct page *page, int this_cpu)
226 unsigned long mask = this_cpu;
227 unsigned long non_cpu_bits;
229 non_cpu_bits = ~(PG_dcache_cpu_mask << PG_dcache_cpu_shift);
230 mask = (mask << PG_dcache_cpu_shift) | (1UL << PG_dcache_dirty);
232 __asm__ __volatile__("1:\n\t"
234 "and %%g7, %1, %%g1\n\t"
235 "or %%g1, %0, %%g1\n\t"
236 "casx [%2], %%g7, %%g1\n\t"
238 "bne,pn %%xcc, 1b\n\t"
241 : "r" (mask), "r" (non_cpu_bits), "r" (&page->flags)
245 static inline void clear_dcache_dirty_cpu(struct page *page, unsigned long cpu)
247 unsigned long mask = (1UL << PG_dcache_dirty);
249 __asm__ __volatile__("! test_and_clear_dcache_dirty\n"
252 "srlx %%g7, %4, %%g1\n\t"
253 "and %%g1, %3, %%g1\n\t"
255 "bne,pn %%icc, 2f\n\t"
256 " andn %%g7, %1, %%g1\n\t"
257 "casx [%2], %%g7, %%g1\n\t"
259 "bne,pn %%xcc, 1b\n\t"
263 : "r" (cpu), "r" (mask), "r" (&page->flags),
264 "i" (PG_dcache_cpu_mask),
265 "i" (PG_dcache_cpu_shift)
269 static inline void tsb_insert(struct tsb *ent, unsigned long tag, unsigned long pte)
271 unsigned long tsb_addr = (unsigned long) ent;
273 if (tlb_type == cheetah_plus || tlb_type == hypervisor)
274 tsb_addr = __pa(tsb_addr);
276 __tsb_insert(tsb_addr, tag, pte);
279 unsigned long _PAGE_ALL_SZ_BITS __read_mostly;
281 static void flush_dcache(unsigned long pfn)
285 page = pfn_to_page(pfn);
287 unsigned long pg_flags;
289 pg_flags = page->flags;
290 if (pg_flags & (1UL << PG_dcache_dirty)) {
291 int cpu = ((pg_flags >> PG_dcache_cpu_shift) &
293 int this_cpu = get_cpu();
295 /* This is just to optimize away some function calls
299 flush_dcache_page_impl(page);
301 smp_flush_dcache_page_impl(page, cpu);
303 clear_dcache_dirty_cpu(page, cpu);
310 /* mm->context.lock must be held */
311 static void __update_mmu_tsb_insert(struct mm_struct *mm, unsigned long tsb_index,
312 unsigned long tsb_hash_shift, unsigned long address,
315 struct tsb *tsb = mm->context.tsb_block[tsb_index].tsb;
321 tsb += ((address >> tsb_hash_shift) &
322 (mm->context.tsb_block[tsb_index].tsb_nentries - 1UL));
323 tag = (address >> 22UL);
324 tsb_insert(tsb, tag, tte);
327 #ifdef CONFIG_HUGETLB_PAGE
328 static int __init setup_hugepagesz(char *string)
330 unsigned long long hugepage_size;
331 unsigned int hugepage_shift;
332 unsigned short hv_pgsz_idx;
333 unsigned int hv_pgsz_mask;
336 hugepage_size = memparse(string, &string);
337 hugepage_shift = ilog2(hugepage_size);
339 switch (hugepage_shift) {
340 case HPAGE_2GB_SHIFT:
341 hv_pgsz_mask = HV_PGSZ_MASK_2GB;
342 hv_pgsz_idx = HV_PGSZ_IDX_2GB;
344 case HPAGE_256MB_SHIFT:
345 hv_pgsz_mask = HV_PGSZ_MASK_256MB;
346 hv_pgsz_idx = HV_PGSZ_IDX_256MB;
349 hv_pgsz_mask = HV_PGSZ_MASK_4MB;
350 hv_pgsz_idx = HV_PGSZ_IDX_4MB;
352 case HPAGE_64K_SHIFT:
353 hv_pgsz_mask = HV_PGSZ_MASK_64K;
354 hv_pgsz_idx = HV_PGSZ_IDX_64K;
360 if ((hv_pgsz_mask & cpu_pgsz_mask) == 0U) {
362 pr_err("hugepagesz=%llu not supported by MMU.\n",
367 hugetlb_add_hstate(hugepage_shift - PAGE_SHIFT);
373 __setup("hugepagesz=", setup_hugepagesz);
374 #endif /* CONFIG_HUGETLB_PAGE */
376 void update_mmu_cache(struct vm_area_struct *vma, unsigned long address, pte_t *ptep)
378 struct mm_struct *mm;
382 if (tlb_type != hypervisor) {
383 unsigned long pfn = pte_pfn(pte);
391 /* Don't insert a non-valid PTE into the TSB, we'll deadlock. */
392 if (!pte_accessible(mm, pte))
395 spin_lock_irqsave(&mm->context.lock, flags);
397 #if defined(CONFIG_HUGETLB_PAGE) || defined(CONFIG_TRANSPARENT_HUGEPAGE)
398 if ((mm->context.hugetlb_pte_count || mm->context.thp_pte_count) &&
399 is_hugetlb_pmd(__pmd(pte_val(pte)))) {
400 /* We are fabricating 8MB pages using 4MB real hw pages. */
401 pte_val(pte) |= (address & (1UL << REAL_HPAGE_SHIFT));
402 __update_mmu_tsb_insert(mm, MM_TSB_HUGE, REAL_HPAGE_SHIFT,
403 address, pte_val(pte));
406 __update_mmu_tsb_insert(mm, MM_TSB_BASE, PAGE_SHIFT,
407 address, pte_val(pte));
409 spin_unlock_irqrestore(&mm->context.lock, flags);
412 void flush_dcache_page(struct page *page)
414 struct address_space *mapping;
417 if (tlb_type == hypervisor)
420 /* Do not bother with the expensive D-cache flush if it
421 * is merely the zero page. The 'bigcore' testcase in GDB
422 * causes this case to run millions of times.
424 if (page == ZERO_PAGE(0))
427 this_cpu = get_cpu();
429 mapping = page_mapping(page);
430 if (mapping && !mapping_mapped(mapping)) {
431 int dirty = test_bit(PG_dcache_dirty, &page->flags);
433 int dirty_cpu = dcache_dirty_cpu(page);
435 if (dirty_cpu == this_cpu)
437 smp_flush_dcache_page_impl(page, dirty_cpu);
439 set_dcache_dirty(page, this_cpu);
441 /* We could delay the flush for the !page_mapping
442 * case too. But that case is for exec env/arg
443 * pages and those are %99 certainly going to get
444 * faulted into the tlb (and thus flushed) anyways.
446 flush_dcache_page_impl(page);
452 EXPORT_SYMBOL(flush_dcache_page);
454 void __kprobes flush_icache_range(unsigned long start, unsigned long end)
456 /* Cheetah and Hypervisor platform cpus have coherent I-cache. */
457 if (tlb_type == spitfire) {
460 /* This code only runs on Spitfire cpus so this is
461 * why we can assume _PAGE_PADDR_4U.
463 for (kaddr = start; kaddr < end; kaddr += PAGE_SIZE) {
464 unsigned long paddr, mask = _PAGE_PADDR_4U;
466 if (kaddr >= PAGE_OFFSET)
467 paddr = kaddr & mask;
469 pgd_t *pgdp = pgd_offset_k(kaddr);
470 pud_t *pudp = pud_offset(pgdp, kaddr);
471 pmd_t *pmdp = pmd_offset(pudp, kaddr);
472 pte_t *ptep = pte_offset_kernel(pmdp, kaddr);
474 paddr = pte_val(*ptep) & mask;
476 __flush_icache_page(paddr);
480 EXPORT_SYMBOL(flush_icache_range);
482 void mmu_info(struct seq_file *m)
484 static const char *pgsz_strings[] = {
485 "8K", "64K", "512K", "4MB", "32MB",
486 "256MB", "2GB", "16GB",
490 if (tlb_type == cheetah)
491 seq_printf(m, "MMU Type\t: Cheetah\n");
492 else if (tlb_type == cheetah_plus)
493 seq_printf(m, "MMU Type\t: Cheetah+\n");
494 else if (tlb_type == spitfire)
495 seq_printf(m, "MMU Type\t: Spitfire\n");
496 else if (tlb_type == hypervisor)
497 seq_printf(m, "MMU Type\t: Hypervisor (sun4v)\n");
499 seq_printf(m, "MMU Type\t: ???\n");
501 seq_printf(m, "MMU PGSZs\t: ");
503 for (i = 0; i < ARRAY_SIZE(pgsz_strings); i++) {
504 if (cpu_pgsz_mask & (1UL << i)) {
505 seq_printf(m, "%s%s",
506 printed ? "," : "", pgsz_strings[i]);
512 #ifdef CONFIG_DEBUG_DCFLUSH
513 seq_printf(m, "DCPageFlushes\t: %d\n",
514 atomic_read(&dcpage_flushes));
516 seq_printf(m, "DCPageFlushesXC\t: %d\n",
517 atomic_read(&dcpage_flushes_xcall));
518 #endif /* CONFIG_SMP */
519 #endif /* CONFIG_DEBUG_DCFLUSH */
522 struct linux_prom_translation prom_trans[512] __read_mostly;
523 unsigned int prom_trans_ents __read_mostly;
525 unsigned long kern_locked_tte_data;
527 /* The obp translations are saved based on 8k pagesize, since obp can
528 * use a mixture of pagesizes. Misses to the LOW_OBP_ADDRESS ->
529 * HI_OBP_ADDRESS range are handled in ktlb.S.
531 static inline int in_obp_range(unsigned long vaddr)
533 return (vaddr >= LOW_OBP_ADDRESS &&
534 vaddr < HI_OBP_ADDRESS);
537 static int cmp_ptrans(const void *a, const void *b)
539 const struct linux_prom_translation *x = a, *y = b;
541 if (x->virt > y->virt)
543 if (x->virt < y->virt)
548 /* Read OBP translations property into 'prom_trans[]'. */
549 static void __init read_obp_translations(void)
551 int n, node, ents, first, last, i;
553 node = prom_finddevice("/virtual-memory");
554 n = prom_getproplen(node, "translations");
555 if (unlikely(n == 0 || n == -1)) {
556 prom_printf("prom_mappings: Couldn't get size.\n");
559 if (unlikely(n > sizeof(prom_trans))) {
560 prom_printf("prom_mappings: Size %d is too big.\n", n);
564 if ((n = prom_getproperty(node, "translations",
565 (char *)&prom_trans[0],
566 sizeof(prom_trans))) == -1) {
567 prom_printf("prom_mappings: Couldn't get property.\n");
571 n = n / sizeof(struct linux_prom_translation);
575 sort(prom_trans, ents, sizeof(struct linux_prom_translation),
578 /* Now kick out all the non-OBP entries. */
579 for (i = 0; i < ents; i++) {
580 if (in_obp_range(prom_trans[i].virt))
584 for (; i < ents; i++) {
585 if (!in_obp_range(prom_trans[i].virt))
590 for (i = 0; i < (last - first); i++) {
591 struct linux_prom_translation *src = &prom_trans[i + first];
592 struct linux_prom_translation *dest = &prom_trans[i];
596 for (; i < ents; i++) {
597 struct linux_prom_translation *dest = &prom_trans[i];
598 dest->virt = dest->size = dest->data = 0x0UL;
601 prom_trans_ents = last - first;
603 if (tlb_type == spitfire) {
604 /* Clear diag TTE bits. */
605 for (i = 0; i < prom_trans_ents; i++)
606 prom_trans[i].data &= ~0x0003fe0000000000UL;
609 /* Force execute bit on. */
610 for (i = 0; i < prom_trans_ents; i++)
611 prom_trans[i].data |= (tlb_type == hypervisor ?
612 _PAGE_EXEC_4V : _PAGE_EXEC_4U);
615 static void __init hypervisor_tlb_lock(unsigned long vaddr,
619 unsigned long ret = sun4v_mmu_map_perm_addr(vaddr, 0, pte, mmu);
622 prom_printf("hypervisor_tlb_lock[%lx:%x:%lx:%lx]: "
623 "errors with %lx\n", vaddr, 0, pte, mmu, ret);
628 static unsigned long kern_large_tte(unsigned long paddr);
630 static void __init remap_kernel(void)
632 unsigned long phys_page, tte_vaddr, tte_data;
633 int i, tlb_ent = sparc64_highest_locked_tlbent();
635 tte_vaddr = (unsigned long) KERNBASE;
636 phys_page = (prom_boot_mapping_phys_low >> ILOG2_4MB) << ILOG2_4MB;
637 tte_data = kern_large_tte(phys_page);
639 kern_locked_tte_data = tte_data;
641 /* Now lock us into the TLBs via Hypervisor or OBP. */
642 if (tlb_type == hypervisor) {
643 for (i = 0; i < num_kernel_image_mappings; i++) {
644 hypervisor_tlb_lock(tte_vaddr, tte_data, HV_MMU_DMMU);
645 hypervisor_tlb_lock(tte_vaddr, tte_data, HV_MMU_IMMU);
646 tte_vaddr += 0x400000;
647 tte_data += 0x400000;
650 for (i = 0; i < num_kernel_image_mappings; i++) {
651 prom_dtlb_load(tlb_ent - i, tte_data, tte_vaddr);
652 prom_itlb_load(tlb_ent - i, tte_data, tte_vaddr);
653 tte_vaddr += 0x400000;
654 tte_data += 0x400000;
656 sparc64_highest_unlocked_tlb_ent = tlb_ent - i;
658 if (tlb_type == cheetah_plus) {
659 sparc64_kern_pri_context = (CTX_CHEETAH_PLUS_CTX0 |
660 CTX_CHEETAH_PLUS_NUC);
661 sparc64_kern_pri_nuc_bits = CTX_CHEETAH_PLUS_NUC;
662 sparc64_kern_sec_context = CTX_CHEETAH_PLUS_CTX0;
667 static void __init inherit_prom_mappings(void)
669 /* Now fixup OBP's idea about where we really are mapped. */
670 printk("Remapping the kernel... ");
675 void prom_world(int enter)
680 __asm__ __volatile__("flushw");
683 void __flush_dcache_range(unsigned long start, unsigned long end)
687 if (tlb_type == spitfire) {
690 for (va = start; va < end; va += 32) {
691 spitfire_put_dcache_tag(va & 0x3fe0, 0x0);
695 } else if (tlb_type == cheetah || tlb_type == cheetah_plus) {
698 for (va = start; va < end; va += 32)
699 __asm__ __volatile__("stxa %%g0, [%0] %1\n\t"
703 "i" (ASI_DCACHE_INVALIDATE));
706 EXPORT_SYMBOL(__flush_dcache_range);
708 /* get_new_mmu_context() uses "cache + 1". */
709 DEFINE_SPINLOCK(ctx_alloc_lock);
710 unsigned long tlb_context_cache = CTX_FIRST_VERSION;
711 #define MAX_CTX_NR (1UL << CTX_NR_BITS)
712 #define CTX_BMAP_SLOTS BITS_TO_LONGS(MAX_CTX_NR)
713 DECLARE_BITMAP(mmu_context_bmap, MAX_CTX_NR);
714 DEFINE_PER_CPU(struct mm_struct *, per_cpu_secondary_mm) = {0};
716 static void mmu_context_wrap(void)
718 unsigned long old_ver = tlb_context_cache & CTX_VERSION_MASK;
719 unsigned long new_ver, new_ctx, old_ctx;
720 struct mm_struct *mm;
723 bitmap_zero(mmu_context_bmap, 1 << CTX_NR_BITS);
725 /* Reserve kernel context */
726 set_bit(0, mmu_context_bmap);
728 new_ver = (tlb_context_cache & CTX_VERSION_MASK) + CTX_FIRST_VERSION;
729 if (unlikely(new_ver == 0))
730 new_ver = CTX_FIRST_VERSION;
731 tlb_context_cache = new_ver;
734 * Make sure that any new mm that are added into per_cpu_secondary_mm,
735 * are going to go through get_new_mmu_context() path.
740 * Updated versions to current on those CPUs that had valid secondary
743 for_each_online_cpu(cpu) {
745 * If a new mm is stored after we took this mm from the array,
746 * it will go into get_new_mmu_context() path, because we
747 * already bumped the version in tlb_context_cache.
749 mm = per_cpu(per_cpu_secondary_mm, cpu);
751 if (unlikely(!mm || mm == &init_mm))
754 old_ctx = mm->context.sparc64_ctx_val;
755 if (likely((old_ctx & CTX_VERSION_MASK) == old_ver)) {
756 new_ctx = (old_ctx & ~CTX_VERSION_MASK) | new_ver;
757 set_bit(new_ctx & CTX_NR_MASK, mmu_context_bmap);
758 mm->context.sparc64_ctx_val = new_ctx;
763 /* Caller does TLB context flushing on local CPU if necessary.
764 * The caller also ensures that CTX_VALID(mm->context) is false.
766 * We must be careful about boundary cases so that we never
767 * let the user have CTX 0 (nucleus) or we ever use a CTX
768 * version of zero (and thus NO_CONTEXT would not be caught
769 * by version mis-match tests in mmu_context.h).
771 * Always invoked with interrupts disabled.
773 void get_new_mmu_context(struct mm_struct *mm)
775 unsigned long ctx, new_ctx;
776 unsigned long orig_pgsz_bits;
778 spin_lock(&ctx_alloc_lock);
780 /* wrap might have happened, test again if our context became valid */
781 if (unlikely(CTX_VALID(mm->context)))
783 orig_pgsz_bits = (mm->context.sparc64_ctx_val & CTX_PGSZ_MASK);
784 ctx = (tlb_context_cache + 1) & CTX_NR_MASK;
785 new_ctx = find_next_zero_bit(mmu_context_bmap, 1 << CTX_NR_BITS, ctx);
786 if (new_ctx >= (1 << CTX_NR_BITS)) {
787 new_ctx = find_next_zero_bit(mmu_context_bmap, ctx, 1);
788 if (new_ctx >= ctx) {
793 if (mm->context.sparc64_ctx_val)
794 cpumask_clear(mm_cpumask(mm));
795 mmu_context_bmap[new_ctx>>6] |= (1UL << (new_ctx & 63));
796 new_ctx |= (tlb_context_cache & CTX_VERSION_MASK);
797 tlb_context_cache = new_ctx;
798 mm->context.sparc64_ctx_val = new_ctx | orig_pgsz_bits;
800 spin_unlock(&ctx_alloc_lock);
803 static int numa_enabled = 1;
804 static int numa_debug;
806 static int __init early_numa(char *p)
811 if (strstr(p, "off"))
814 if (strstr(p, "debug"))
819 early_param("numa", early_numa);
821 #define numadbg(f, a...) \
822 do { if (numa_debug) \
823 printk(KERN_INFO f, ## a); \
826 static void __init find_ramdisk(unsigned long phys_base)
828 #ifdef CONFIG_BLK_DEV_INITRD
829 if (sparc_ramdisk_image || sparc_ramdisk_image64) {
830 unsigned long ramdisk_image;
832 /* Older versions of the bootloader only supported a
833 * 32-bit physical address for the ramdisk image
834 * location, stored at sparc_ramdisk_image. Newer
835 * SILO versions set sparc_ramdisk_image to zero and
836 * provide a full 64-bit physical address at
837 * sparc_ramdisk_image64.
839 ramdisk_image = sparc_ramdisk_image;
841 ramdisk_image = sparc_ramdisk_image64;
843 /* Another bootloader quirk. The bootloader normalizes
844 * the physical address to KERNBASE, so we have to
845 * factor that back out and add in the lowest valid
846 * physical page address to get the true physical address.
848 ramdisk_image -= KERNBASE;
849 ramdisk_image += phys_base;
851 numadbg("Found ramdisk at physical address 0x%lx, size %u\n",
852 ramdisk_image, sparc_ramdisk_size);
854 initrd_start = ramdisk_image;
855 initrd_end = ramdisk_image + sparc_ramdisk_size;
857 memblock_reserve(initrd_start, sparc_ramdisk_size);
859 initrd_start += PAGE_OFFSET;
860 initrd_end += PAGE_OFFSET;
865 struct node_mem_mask {
869 static struct node_mem_mask node_masks[MAX_NUMNODES];
870 static int num_node_masks;
872 #ifdef CONFIG_NEED_MULTIPLE_NODES
874 struct mdesc_mlgroup {
881 static struct mdesc_mlgroup *mlgroups;
882 static int num_mlgroups;
884 int numa_cpu_lookup_table[NR_CPUS];
885 cpumask_t numa_cpumask_lookup_table[MAX_NUMNODES];
887 struct mdesc_mblock {
890 u64 offset; /* RA-to-PA */
892 static struct mdesc_mblock *mblocks;
893 static int num_mblocks;
895 static struct mdesc_mblock * __init addr_to_mblock(unsigned long addr)
897 struct mdesc_mblock *m = NULL;
900 for (i = 0; i < num_mblocks; i++) {
903 if (addr >= m->base &&
904 addr < (m->base + m->size)) {
912 static u64 __init memblock_nid_range_sun4u(u64 start, u64 end, int *nid)
914 int prev_nid, new_nid;
917 for ( ; start < end; start += PAGE_SIZE) {
918 for (new_nid = 0; new_nid < num_node_masks; new_nid++) {
919 struct node_mem_mask *p = &node_masks[new_nid];
921 if ((start & p->mask) == p->match) {
928 if (new_nid == num_node_masks) {
930 WARN_ONCE(1, "addr[%Lx] doesn't match a NUMA node rule. Some memory will be owned by node 0.",
935 if (prev_nid != new_nid)
940 return start > end ? end : start;
943 static u64 __init memblock_nid_range(u64 start, u64 end, int *nid)
945 u64 ret_end, pa_start, m_mask, m_match, m_end;
946 struct mdesc_mblock *mblock;
949 if (tlb_type != hypervisor)
950 return memblock_nid_range_sun4u(start, end, nid);
952 mblock = addr_to_mblock(start);
954 WARN_ONCE(1, "memblock_nid_range: Can't find mblock addr[%Lx]",
962 pa_start = start + mblock->offset;
966 for (_nid = 0; _nid < num_node_masks; _nid++) {
967 struct node_mem_mask *const m = &node_masks[_nid];
969 if ((pa_start & m->mask) == m->match) {
976 if (num_node_masks == _nid) {
977 /* We could not find NUMA group, so default to 0, but lets
978 * search for latency group, so we could calculate the correct
979 * end address that we return
983 for (i = 0; i < num_mlgroups; i++) {
984 struct mdesc_mlgroup *const m = &mlgroups[i];
986 if ((pa_start & m->mask) == m->match) {
993 if (i == num_mlgroups) {
994 WARN_ONCE(1, "memblock_nid_range: Can't find latency group addr[%Lx]",
1003 * Each latency group has match and mask, and each memory block has an
1004 * offset. An address belongs to a latency group if its address matches
1005 * the following formula: ((addr + offset) & mask) == match
1006 * It is, however, slow to check every single page if it matches a
1007 * particular latency group. As optimization we calculate end value by
1008 * using bit arithmetics.
1010 m_end = m_match + (1ul << __ffs(m_mask)) - mblock->offset;
1011 m_end += pa_start & ~((1ul << fls64(m_mask)) - 1);
1012 ret_end = m_end > end ? end : m_end;
1020 /* This must be invoked after performing all of the necessary
1021 * memblock_set_node() calls for 'nid'. We need to be able to get
1022 * correct data from get_pfn_range_for_nid().
1024 static void __init allocate_node_data(int nid)
1026 struct pglist_data *p;
1027 unsigned long start_pfn, end_pfn;
1028 #ifdef CONFIG_NEED_MULTIPLE_NODES
1029 unsigned long paddr;
1031 paddr = memblock_alloc_try_nid(sizeof(struct pglist_data), SMP_CACHE_BYTES, nid);
1033 prom_printf("Cannot allocate pglist_data for nid[%d]\n", nid);
1036 NODE_DATA(nid) = __va(paddr);
1037 memset(NODE_DATA(nid), 0, sizeof(struct pglist_data));
1039 NODE_DATA(nid)->node_id = nid;
1044 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
1045 p->node_start_pfn = start_pfn;
1046 p->node_spanned_pages = end_pfn - start_pfn;
1049 static void init_node_masks_nonnuma(void)
1051 #ifdef CONFIG_NEED_MULTIPLE_NODES
1055 numadbg("Initializing tables for non-numa.\n");
1057 node_masks[0].mask = 0;
1058 node_masks[0].match = 0;
1061 #ifdef CONFIG_NEED_MULTIPLE_NODES
1062 for (i = 0; i < NR_CPUS; i++)
1063 numa_cpu_lookup_table[i] = 0;
1065 cpumask_setall(&numa_cpumask_lookup_table[0]);
1069 #ifdef CONFIG_NEED_MULTIPLE_NODES
1070 struct pglist_data *node_data[MAX_NUMNODES];
1072 EXPORT_SYMBOL(numa_cpu_lookup_table);
1073 EXPORT_SYMBOL(numa_cpumask_lookup_table);
1074 EXPORT_SYMBOL(node_data);
1076 static int scan_pio_for_cfg_handle(struct mdesc_handle *md, u64 pio,
1081 mdesc_for_each_arc(arc, md, pio, MDESC_ARC_TYPE_FWD) {
1082 u64 target = mdesc_arc_target(md, arc);
1085 val = mdesc_get_property(md, target,
1086 "cfg-handle", NULL);
1087 if (val && *val == cfg_handle)
1093 static int scan_arcs_for_cfg_handle(struct mdesc_handle *md, u64 grp,
1096 u64 arc, candidate, best_latency = ~(u64)0;
1098 candidate = MDESC_NODE_NULL;
1099 mdesc_for_each_arc(arc, md, grp, MDESC_ARC_TYPE_FWD) {
1100 u64 target = mdesc_arc_target(md, arc);
1101 const char *name = mdesc_node_name(md, target);
1104 if (strcmp(name, "pio-latency-group"))
1107 val = mdesc_get_property(md, target, "latency", NULL);
1111 if (*val < best_latency) {
1113 best_latency = *val;
1117 if (candidate == MDESC_NODE_NULL)
1120 return scan_pio_for_cfg_handle(md, candidate, cfg_handle);
1123 int of_node_to_nid(struct device_node *dp)
1125 const struct linux_prom64_registers *regs;
1126 struct mdesc_handle *md;
1131 /* This is the right thing to do on currently supported
1132 * SUN4U NUMA platforms as well, as the PCI controller does
1133 * not sit behind any particular memory controller.
1138 regs = of_get_property(dp, "reg", NULL);
1142 cfg_handle = (regs->phys_addr >> 32UL) & 0x0fffffff;
1148 mdesc_for_each_node_by_name(md, grp, "group") {
1149 if (!scan_arcs_for_cfg_handle(md, grp, cfg_handle)) {
1161 static void __init add_node_ranges(void)
1163 struct memblock_region *reg;
1164 unsigned long prev_max;
1167 prev_max = memblock.memory.max;
1169 for_each_memblock(memory, reg) {
1170 unsigned long size = reg->size;
1171 unsigned long start, end;
1175 while (start < end) {
1176 unsigned long this_end;
1179 this_end = memblock_nid_range(start, end, &nid);
1181 numadbg("Setting memblock NUMA node nid[%d] "
1182 "start[%lx] end[%lx]\n",
1183 nid, start, this_end);
1185 memblock_set_node(start, this_end - start,
1186 &memblock.memory, nid);
1187 if (memblock.memory.max != prev_max)
1188 goto memblock_resized;
1194 static int __init grab_mlgroups(struct mdesc_handle *md)
1196 unsigned long paddr;
1200 mdesc_for_each_node_by_name(md, node, "memory-latency-group")
1205 paddr = memblock_alloc(count * sizeof(struct mdesc_mlgroup),
1210 mlgroups = __va(paddr);
1211 num_mlgroups = count;
1214 mdesc_for_each_node_by_name(md, node, "memory-latency-group") {
1215 struct mdesc_mlgroup *m = &mlgroups[count++];
1220 val = mdesc_get_property(md, node, "latency", NULL);
1222 val = mdesc_get_property(md, node, "address-match", NULL);
1224 val = mdesc_get_property(md, node, "address-mask", NULL);
1227 numadbg("MLGROUP[%d]: node[%llx] latency[%llx] "
1228 "match[%llx] mask[%llx]\n",
1229 count - 1, m->node, m->latency, m->match, m->mask);
1235 static int __init grab_mblocks(struct mdesc_handle *md)
1237 unsigned long paddr;
1241 mdesc_for_each_node_by_name(md, node, "mblock")
1246 paddr = memblock_alloc(count * sizeof(struct mdesc_mblock),
1251 mblocks = __va(paddr);
1252 num_mblocks = count;
1255 mdesc_for_each_node_by_name(md, node, "mblock") {
1256 struct mdesc_mblock *m = &mblocks[count++];
1259 val = mdesc_get_property(md, node, "base", NULL);
1261 val = mdesc_get_property(md, node, "size", NULL);
1263 val = mdesc_get_property(md, node,
1264 "address-congruence-offset", NULL);
1266 /* The address-congruence-offset property is optional.
1267 * Explicity zero it be identifty this.
1274 numadbg("MBLOCK[%d]: base[%llx] size[%llx] offset[%llx]\n",
1275 count - 1, m->base, m->size, m->offset);
1281 static void __init numa_parse_mdesc_group_cpus(struct mdesc_handle *md,
1282 u64 grp, cpumask_t *mask)
1286 cpumask_clear(mask);
1288 mdesc_for_each_arc(arc, md, grp, MDESC_ARC_TYPE_BACK) {
1289 u64 target = mdesc_arc_target(md, arc);
1290 const char *name = mdesc_node_name(md, target);
1293 if (strcmp(name, "cpu"))
1295 id = mdesc_get_property(md, target, "id", NULL);
1296 if (*id < nr_cpu_ids)
1297 cpumask_set_cpu(*id, mask);
1301 static struct mdesc_mlgroup * __init find_mlgroup(u64 node)
1305 for (i = 0; i < num_mlgroups; i++) {
1306 struct mdesc_mlgroup *m = &mlgroups[i];
1307 if (m->node == node)
1313 int __node_distance(int from, int to)
1315 if ((from >= MAX_NUMNODES) || (to >= MAX_NUMNODES)) {
1316 pr_warn("Returning default NUMA distance value for %d->%d\n",
1318 return (from == to) ? LOCAL_DISTANCE : REMOTE_DISTANCE;
1320 return numa_latency[from][to];
1323 static int __init find_best_numa_node_for_mlgroup(struct mdesc_mlgroup *grp)
1327 for (i = 0; i < MAX_NUMNODES; i++) {
1328 struct node_mem_mask *n = &node_masks[i];
1330 if ((grp->mask == n->mask) && (grp->match == n->match))
1336 static void __init find_numa_latencies_for_group(struct mdesc_handle *md,
1341 mdesc_for_each_arc(arc, md, grp, MDESC_ARC_TYPE_FWD) {
1343 u64 target = mdesc_arc_target(md, arc);
1344 struct mdesc_mlgroup *m = find_mlgroup(target);
1348 tnode = find_best_numa_node_for_mlgroup(m);
1349 if (tnode == MAX_NUMNODES)
1351 numa_latency[index][tnode] = m->latency;
1355 static int __init numa_attach_mlgroup(struct mdesc_handle *md, u64 grp,
1358 struct mdesc_mlgroup *candidate = NULL;
1359 u64 arc, best_latency = ~(u64)0;
1360 struct node_mem_mask *n;
1362 mdesc_for_each_arc(arc, md, grp, MDESC_ARC_TYPE_FWD) {
1363 u64 target = mdesc_arc_target(md, arc);
1364 struct mdesc_mlgroup *m = find_mlgroup(target);
1367 if (m->latency < best_latency) {
1369 best_latency = m->latency;
1375 if (num_node_masks != index) {
1376 printk(KERN_ERR "Inconsistent NUMA state, "
1377 "index[%d] != num_node_masks[%d]\n",
1378 index, num_node_masks);
1382 n = &node_masks[num_node_masks++];
1384 n->mask = candidate->mask;
1385 n->match = candidate->match;
1387 numadbg("NUMA NODE[%d]: mask[%lx] match[%lx] (latency[%llx])\n",
1388 index, n->mask, n->match, candidate->latency);
1393 static int __init numa_parse_mdesc_group(struct mdesc_handle *md, u64 grp,
1399 numa_parse_mdesc_group_cpus(md, grp, &mask);
1401 for_each_cpu(cpu, &mask)
1402 numa_cpu_lookup_table[cpu] = index;
1403 cpumask_copy(&numa_cpumask_lookup_table[index], &mask);
1406 printk(KERN_INFO "NUMA GROUP[%d]: cpus [ ", index);
1407 for_each_cpu(cpu, &mask)
1412 return numa_attach_mlgroup(md, grp, index);
1415 static int __init numa_parse_mdesc(void)
1417 struct mdesc_handle *md = mdesc_grab();
1418 int i, j, err, count;
1421 node = mdesc_node_by_name(md, MDESC_NODE_NULL, "latency-groups");
1422 if (node == MDESC_NODE_NULL) {
1427 err = grab_mblocks(md);
1431 err = grab_mlgroups(md);
1436 mdesc_for_each_node_by_name(md, node, "group") {
1437 err = numa_parse_mdesc_group(md, node, count);
1444 mdesc_for_each_node_by_name(md, node, "group") {
1445 find_numa_latencies_for_group(md, node, count);
1449 /* Normalize numa latency matrix according to ACPI SLIT spec. */
1450 for (i = 0; i < MAX_NUMNODES; i++) {
1451 u64 self_latency = numa_latency[i][i];
1453 for (j = 0; j < MAX_NUMNODES; j++) {
1454 numa_latency[i][j] =
1455 (numa_latency[i][j] * LOCAL_DISTANCE) /
1462 for (i = 0; i < num_node_masks; i++) {
1463 allocate_node_data(i);
1473 static int __init numa_parse_jbus(void)
1475 unsigned long cpu, index;
1477 /* NUMA node id is encoded in bits 36 and higher, and there is
1478 * a 1-to-1 mapping from CPU ID to NUMA node ID.
1481 for_each_present_cpu(cpu) {
1482 numa_cpu_lookup_table[cpu] = index;
1483 cpumask_copy(&numa_cpumask_lookup_table[index], cpumask_of(cpu));
1484 node_masks[index].mask = ~((1UL << 36UL) - 1UL);
1485 node_masks[index].match = cpu << 36UL;
1489 num_node_masks = index;
1493 for (index = 0; index < num_node_masks; index++) {
1494 allocate_node_data(index);
1495 node_set_online(index);
1501 static int __init numa_parse_sun4u(void)
1503 if (tlb_type == cheetah || tlb_type == cheetah_plus) {
1506 __asm__ ("rdpr %%ver, %0" : "=r" (ver));
1507 if ((ver >> 32UL) == __JALAPENO_ID ||
1508 (ver >> 32UL) == __SERRANO_ID)
1509 return numa_parse_jbus();
1514 static int __init bootmem_init_numa(void)
1519 numadbg("bootmem_init_numa()\n");
1521 /* Some sane defaults for numa latency values */
1522 for (i = 0; i < MAX_NUMNODES; i++) {
1523 for (j = 0; j < MAX_NUMNODES; j++)
1524 numa_latency[i][j] = (i == j) ?
1525 LOCAL_DISTANCE : REMOTE_DISTANCE;
1529 if (tlb_type == hypervisor)
1530 err = numa_parse_mdesc();
1532 err = numa_parse_sun4u();
1539 static int bootmem_init_numa(void)
1546 static void __init bootmem_init_nonnuma(void)
1548 unsigned long top_of_ram = memblock_end_of_DRAM();
1549 unsigned long total_ram = memblock_phys_mem_size();
1551 numadbg("bootmem_init_nonnuma()\n");
1553 printk(KERN_INFO "Top of RAM: 0x%lx, Total RAM: 0x%lx\n",
1554 top_of_ram, total_ram);
1555 printk(KERN_INFO "Memory hole size: %ldMB\n",
1556 (top_of_ram - total_ram) >> 20);
1558 init_node_masks_nonnuma();
1559 memblock_set_node(0, (phys_addr_t)ULLONG_MAX, &memblock.memory, 0);
1560 allocate_node_data(0);
1564 static unsigned long __init bootmem_init(unsigned long phys_base)
1566 unsigned long end_pfn;
1568 end_pfn = memblock_end_of_DRAM() >> PAGE_SHIFT;
1569 max_pfn = max_low_pfn = end_pfn;
1570 min_low_pfn = (phys_base >> PAGE_SHIFT);
1572 if (bootmem_init_numa() < 0)
1573 bootmem_init_nonnuma();
1575 /* Dump memblock with node info. */
1576 memblock_dump_all();
1578 /* XXX cpu notifier XXX */
1580 sparse_memory_present_with_active_regions(MAX_NUMNODES);
1586 static struct linux_prom64_registers pall[MAX_BANKS] __initdata;
1587 static int pall_ents __initdata;
1589 static unsigned long max_phys_bits = 40;
1591 bool kern_addr_valid(unsigned long addr)
1598 if ((long)addr < 0L) {
1599 unsigned long pa = __pa(addr);
1601 if ((pa >> max_phys_bits) != 0UL)
1604 return pfn_valid(pa >> PAGE_SHIFT);
1607 if (addr >= (unsigned long) KERNBASE &&
1608 addr < (unsigned long)&_end)
1611 pgd = pgd_offset_k(addr);
1615 pud = pud_offset(pgd, addr);
1619 if (pud_large(*pud))
1620 return pfn_valid(pud_pfn(*pud));
1622 pmd = pmd_offset(pud, addr);
1626 if (pmd_large(*pmd))
1627 return pfn_valid(pmd_pfn(*pmd));
1629 pte = pte_offset_kernel(pmd, addr);
1633 return pfn_valid(pte_pfn(*pte));
1635 EXPORT_SYMBOL(kern_addr_valid);
1637 static unsigned long __ref kernel_map_hugepud(unsigned long vstart,
1641 const unsigned long mask16gb = (1UL << 34) - 1UL;
1642 u64 pte_val = vstart;
1644 /* Each PUD is 8GB */
1645 if ((vstart & mask16gb) ||
1646 (vend - vstart <= mask16gb)) {
1647 pte_val ^= kern_linear_pte_xor[2];
1648 pud_val(*pud) = pte_val | _PAGE_PUD_HUGE;
1650 return vstart + PUD_SIZE;
1653 pte_val ^= kern_linear_pte_xor[3];
1654 pte_val |= _PAGE_PUD_HUGE;
1656 vend = vstart + mask16gb + 1UL;
1657 while (vstart < vend) {
1658 pud_val(*pud) = pte_val;
1660 pte_val += PUD_SIZE;
1667 static bool kernel_can_map_hugepud(unsigned long vstart, unsigned long vend,
1670 if (guard && !(vstart & ~PUD_MASK) && (vend - vstart) >= PUD_SIZE)
1676 static unsigned long __ref kernel_map_hugepmd(unsigned long vstart,
1680 const unsigned long mask256mb = (1UL << 28) - 1UL;
1681 const unsigned long mask2gb = (1UL << 31) - 1UL;
1682 u64 pte_val = vstart;
1684 /* Each PMD is 8MB */
1685 if ((vstart & mask256mb) ||
1686 (vend - vstart <= mask256mb)) {
1687 pte_val ^= kern_linear_pte_xor[0];
1688 pmd_val(*pmd) = pte_val | _PAGE_PMD_HUGE;
1690 return vstart + PMD_SIZE;
1693 if ((vstart & mask2gb) ||
1694 (vend - vstart <= mask2gb)) {
1695 pte_val ^= kern_linear_pte_xor[1];
1696 pte_val |= _PAGE_PMD_HUGE;
1697 vend = vstart + mask256mb + 1UL;
1699 pte_val ^= kern_linear_pte_xor[2];
1700 pte_val |= _PAGE_PMD_HUGE;
1701 vend = vstart + mask2gb + 1UL;
1704 while (vstart < vend) {
1705 pmd_val(*pmd) = pte_val;
1707 pte_val += PMD_SIZE;
1715 static bool kernel_can_map_hugepmd(unsigned long vstart, unsigned long vend,
1718 if (guard && !(vstart & ~PMD_MASK) && (vend - vstart) >= PMD_SIZE)
1724 static unsigned long __ref kernel_map_range(unsigned long pstart,
1725 unsigned long pend, pgprot_t prot,
1728 unsigned long vstart = PAGE_OFFSET + pstart;
1729 unsigned long vend = PAGE_OFFSET + pend;
1730 unsigned long alloc_bytes = 0UL;
1732 if ((vstart & ~PAGE_MASK) || (vend & ~PAGE_MASK)) {
1733 prom_printf("kernel_map: Unaligned physmem[%lx:%lx]\n",
1738 while (vstart < vend) {
1739 unsigned long this_end, paddr = __pa(vstart);
1740 pgd_t *pgd = pgd_offset_k(vstart);
1745 if (pgd_none(*pgd)) {
1748 new = __alloc_bootmem(PAGE_SIZE, PAGE_SIZE, PAGE_SIZE);
1749 alloc_bytes += PAGE_SIZE;
1750 pgd_populate(&init_mm, pgd, new);
1752 pud = pud_offset(pgd, vstart);
1753 if (pud_none(*pud)) {
1756 if (kernel_can_map_hugepud(vstart, vend, use_huge)) {
1757 vstart = kernel_map_hugepud(vstart, vend, pud);
1760 new = __alloc_bootmem(PAGE_SIZE, PAGE_SIZE, PAGE_SIZE);
1761 alloc_bytes += PAGE_SIZE;
1762 pud_populate(&init_mm, pud, new);
1765 pmd = pmd_offset(pud, vstart);
1766 if (pmd_none(*pmd)) {
1769 if (kernel_can_map_hugepmd(vstart, vend, use_huge)) {
1770 vstart = kernel_map_hugepmd(vstart, vend, pmd);
1773 new = __alloc_bootmem(PAGE_SIZE, PAGE_SIZE, PAGE_SIZE);
1774 alloc_bytes += PAGE_SIZE;
1775 pmd_populate_kernel(&init_mm, pmd, new);
1778 pte = pte_offset_kernel(pmd, vstart);
1779 this_end = (vstart + PMD_SIZE) & PMD_MASK;
1780 if (this_end > vend)
1783 while (vstart < this_end) {
1784 pte_val(*pte) = (paddr | pgprot_val(prot));
1786 vstart += PAGE_SIZE;
1795 static void __init flush_all_kernel_tsbs(void)
1799 for (i = 0; i < KERNEL_TSB_NENTRIES; i++) {
1800 struct tsb *ent = &swapper_tsb[i];
1802 ent->tag = (1UL << TSB_TAG_INVALID_BIT);
1804 #ifndef CONFIG_DEBUG_PAGEALLOC
1805 for (i = 0; i < KERNEL_TSB4M_NENTRIES; i++) {
1806 struct tsb *ent = &swapper_4m_tsb[i];
1808 ent->tag = (1UL << TSB_TAG_INVALID_BIT);
1813 extern unsigned int kvmap_linear_patch[1];
1815 static void __init kernel_physical_mapping_init(void)
1817 unsigned long i, mem_alloced = 0UL;
1818 bool use_huge = true;
1820 #ifdef CONFIG_DEBUG_PAGEALLOC
1823 for (i = 0; i < pall_ents; i++) {
1824 unsigned long phys_start, phys_end;
1826 phys_start = pall[i].phys_addr;
1827 phys_end = phys_start + pall[i].reg_size;
1829 mem_alloced += kernel_map_range(phys_start, phys_end,
1830 PAGE_KERNEL, use_huge);
1833 printk("Allocated %ld bytes for kernel page tables.\n",
1836 kvmap_linear_patch[0] = 0x01000000; /* nop */
1837 flushi(&kvmap_linear_patch[0]);
1839 flush_all_kernel_tsbs();
1844 #ifdef CONFIG_DEBUG_PAGEALLOC
1845 void __kernel_map_pages(struct page *page, int numpages, int enable)
1847 unsigned long phys_start = page_to_pfn(page) << PAGE_SHIFT;
1848 unsigned long phys_end = phys_start + (numpages * PAGE_SIZE);
1850 kernel_map_range(phys_start, phys_end,
1851 (enable ? PAGE_KERNEL : __pgprot(0)), false);
1853 flush_tsb_kernel_range(PAGE_OFFSET + phys_start,
1854 PAGE_OFFSET + phys_end);
1856 /* we should perform an IPI and flush all tlbs,
1857 * but that can deadlock->flush only current cpu.
1859 __flush_tlb_kernel_range(PAGE_OFFSET + phys_start,
1860 PAGE_OFFSET + phys_end);
1864 unsigned long __init find_ecache_flush_span(unsigned long size)
1868 for (i = 0; i < pavail_ents; i++) {
1869 if (pavail[i].reg_size >= size)
1870 return pavail[i].phys_addr;
1876 unsigned long PAGE_OFFSET;
1877 EXPORT_SYMBOL(PAGE_OFFSET);
1879 unsigned long VMALLOC_END = 0x0000010000000000UL;
1880 EXPORT_SYMBOL(VMALLOC_END);
1882 unsigned long sparc64_va_hole_top = 0xfffff80000000000UL;
1883 unsigned long sparc64_va_hole_bottom = 0x0000080000000000UL;
1885 static void __init setup_page_offset(void)
1887 if (tlb_type == cheetah || tlb_type == cheetah_plus) {
1888 /* Cheetah/Panther support a full 64-bit virtual
1889 * address, so we can use all that our page tables
1892 sparc64_va_hole_top = 0xfff0000000000000UL;
1893 sparc64_va_hole_bottom = 0x0010000000000000UL;
1896 } else if (tlb_type == hypervisor) {
1897 switch (sun4v_chip_type) {
1898 case SUN4V_CHIP_NIAGARA1:
1899 case SUN4V_CHIP_NIAGARA2:
1900 /* T1 and T2 support 48-bit virtual addresses. */
1901 sparc64_va_hole_top = 0xffff800000000000UL;
1902 sparc64_va_hole_bottom = 0x0000800000000000UL;
1906 case SUN4V_CHIP_NIAGARA3:
1907 /* T3 supports 48-bit virtual addresses. */
1908 sparc64_va_hole_top = 0xffff800000000000UL;
1909 sparc64_va_hole_bottom = 0x0000800000000000UL;
1913 case SUN4V_CHIP_NIAGARA4:
1914 case SUN4V_CHIP_NIAGARA5:
1915 case SUN4V_CHIP_SPARC64X:
1916 case SUN4V_CHIP_SPARC_M6:
1917 /* T4 and later support 52-bit virtual addresses. */
1918 sparc64_va_hole_top = 0xfff8000000000000UL;
1919 sparc64_va_hole_bottom = 0x0008000000000000UL;
1922 case SUN4V_CHIP_SPARC_M7:
1923 case SUN4V_CHIP_SPARC_SN:
1925 /* M7 and later support 52-bit virtual addresses. */
1926 sparc64_va_hole_top = 0xfff8000000000000UL;
1927 sparc64_va_hole_bottom = 0x0008000000000000UL;
1933 if (max_phys_bits > MAX_PHYS_ADDRESS_BITS) {
1934 prom_printf("MAX_PHYS_ADDRESS_BITS is too small, need %lu\n",
1939 PAGE_OFFSET = sparc64_va_hole_top;
1940 VMALLOC_END = ((sparc64_va_hole_bottom >> 1) +
1941 (sparc64_va_hole_bottom >> 2));
1943 pr_info("MM: PAGE_OFFSET is 0x%016lx (max_phys_bits == %lu)\n",
1944 PAGE_OFFSET, max_phys_bits);
1945 pr_info("MM: VMALLOC [0x%016lx --> 0x%016lx]\n",
1946 VMALLOC_START, VMALLOC_END);
1947 pr_info("MM: VMEMMAP [0x%016lx --> 0x%016lx]\n",
1948 VMEMMAP_BASE, VMEMMAP_BASE << 1);
1951 static void __init tsb_phys_patch(void)
1953 struct tsb_ldquad_phys_patch_entry *pquad;
1954 struct tsb_phys_patch_entry *p;
1956 pquad = &__tsb_ldquad_phys_patch;
1957 while (pquad < &__tsb_ldquad_phys_patch_end) {
1958 unsigned long addr = pquad->addr;
1960 if (tlb_type == hypervisor)
1961 *(unsigned int *) addr = pquad->sun4v_insn;
1963 *(unsigned int *) addr = pquad->sun4u_insn;
1965 __asm__ __volatile__("flush %0"
1972 p = &__tsb_phys_patch;
1973 while (p < &__tsb_phys_patch_end) {
1974 unsigned long addr = p->addr;
1976 *(unsigned int *) addr = p->insn;
1978 __asm__ __volatile__("flush %0"
1986 /* Don't mark as init, we give this to the Hypervisor. */
1987 #ifndef CONFIG_DEBUG_PAGEALLOC
1988 #define NUM_KTSB_DESCR 2
1990 #define NUM_KTSB_DESCR 1
1992 static struct hv_tsb_descr ktsb_descr[NUM_KTSB_DESCR];
1994 /* The swapper TSBs are loaded with a base sequence of:
1996 * sethi %uhi(SYMBOL), REG1
1997 * sethi %hi(SYMBOL), REG2
1998 * or REG1, %ulo(SYMBOL), REG1
1999 * or REG2, %lo(SYMBOL), REG2
2000 * sllx REG1, 32, REG1
2001 * or REG1, REG2, REG1
2003 * When we use physical addressing for the TSB accesses, we patch the
2004 * first four instructions in the above sequence.
2007 static void patch_one_ktsb_phys(unsigned int *start, unsigned int *end, unsigned long pa)
2009 unsigned long high_bits, low_bits;
2011 high_bits = (pa >> 32) & 0xffffffff;
2012 low_bits = (pa >> 0) & 0xffffffff;
2014 while (start < end) {
2015 unsigned int *ia = (unsigned int *)(unsigned long)*start;
2017 ia[0] = (ia[0] & ~0x3fffff) | (high_bits >> 10);
2018 __asm__ __volatile__("flush %0" : : "r" (ia));
2020 ia[1] = (ia[1] & ~0x3fffff) | (low_bits >> 10);
2021 __asm__ __volatile__("flush %0" : : "r" (ia + 1));
2023 ia[2] = (ia[2] & ~0x1fff) | (high_bits & 0x3ff);
2024 __asm__ __volatile__("flush %0" : : "r" (ia + 2));
2026 ia[3] = (ia[3] & ~0x1fff) | (low_bits & 0x3ff);
2027 __asm__ __volatile__("flush %0" : : "r" (ia + 3));
2033 static void ktsb_phys_patch(void)
2035 extern unsigned int __swapper_tsb_phys_patch;
2036 extern unsigned int __swapper_tsb_phys_patch_end;
2037 unsigned long ktsb_pa;
2039 ktsb_pa = kern_base + ((unsigned long)&swapper_tsb[0] - KERNBASE);
2040 patch_one_ktsb_phys(&__swapper_tsb_phys_patch,
2041 &__swapper_tsb_phys_patch_end, ktsb_pa);
2042 #ifndef CONFIG_DEBUG_PAGEALLOC
2044 extern unsigned int __swapper_4m_tsb_phys_patch;
2045 extern unsigned int __swapper_4m_tsb_phys_patch_end;
2046 ktsb_pa = (kern_base +
2047 ((unsigned long)&swapper_4m_tsb[0] - KERNBASE));
2048 patch_one_ktsb_phys(&__swapper_4m_tsb_phys_patch,
2049 &__swapper_4m_tsb_phys_patch_end, ktsb_pa);
2054 static void __init sun4v_ktsb_init(void)
2056 unsigned long ktsb_pa;
2058 /* First KTSB for PAGE_SIZE mappings. */
2059 ktsb_pa = kern_base + ((unsigned long)&swapper_tsb[0] - KERNBASE);
2061 switch (PAGE_SIZE) {
2064 ktsb_descr[0].pgsz_idx = HV_PGSZ_IDX_8K;
2065 ktsb_descr[0].pgsz_mask = HV_PGSZ_MASK_8K;
2069 ktsb_descr[0].pgsz_idx = HV_PGSZ_IDX_64K;
2070 ktsb_descr[0].pgsz_mask = HV_PGSZ_MASK_64K;
2074 ktsb_descr[0].pgsz_idx = HV_PGSZ_IDX_512K;
2075 ktsb_descr[0].pgsz_mask = HV_PGSZ_MASK_512K;
2078 case 4 * 1024 * 1024:
2079 ktsb_descr[0].pgsz_idx = HV_PGSZ_IDX_4MB;
2080 ktsb_descr[0].pgsz_mask = HV_PGSZ_MASK_4MB;
2084 ktsb_descr[0].assoc = 1;
2085 ktsb_descr[0].num_ttes = KERNEL_TSB_NENTRIES;
2086 ktsb_descr[0].ctx_idx = 0;
2087 ktsb_descr[0].tsb_base = ktsb_pa;
2088 ktsb_descr[0].resv = 0;
2090 #ifndef CONFIG_DEBUG_PAGEALLOC
2091 /* Second KTSB for 4MB/256MB/2GB/16GB mappings. */
2092 ktsb_pa = (kern_base +
2093 ((unsigned long)&swapper_4m_tsb[0] - KERNBASE));
2095 ktsb_descr[1].pgsz_idx = HV_PGSZ_IDX_4MB;
2096 ktsb_descr[1].pgsz_mask = ((HV_PGSZ_MASK_4MB |
2097 HV_PGSZ_MASK_256MB |
2099 HV_PGSZ_MASK_16GB) &
2101 ktsb_descr[1].assoc = 1;
2102 ktsb_descr[1].num_ttes = KERNEL_TSB4M_NENTRIES;
2103 ktsb_descr[1].ctx_idx = 0;
2104 ktsb_descr[1].tsb_base = ktsb_pa;
2105 ktsb_descr[1].resv = 0;
2109 void sun4v_ktsb_register(void)
2111 unsigned long pa, ret;
2113 pa = kern_base + ((unsigned long)&ktsb_descr[0] - KERNBASE);
2115 ret = sun4v_mmu_tsb_ctx0(NUM_KTSB_DESCR, pa);
2117 prom_printf("hypervisor_mmu_tsb_ctx0[%lx]: "
2118 "errors with %lx\n", pa, ret);
2123 static void __init sun4u_linear_pte_xor_finalize(void)
2125 #ifndef CONFIG_DEBUG_PAGEALLOC
2126 /* This is where we would add Panther support for
2127 * 32MB and 256MB pages.
2132 static void __init sun4v_linear_pte_xor_finalize(void)
2134 unsigned long pagecv_flag;
2136 /* Bit 9 of TTE is no longer CV bit on M7 processor and it instead
2137 * enables MCD error. Do not set bit 9 on M7 processor.
2139 switch (sun4v_chip_type) {
2140 case SUN4V_CHIP_SPARC_M7:
2141 case SUN4V_CHIP_SPARC_SN:
2145 pagecv_flag = _PAGE_CV_4V;
2148 #ifndef CONFIG_DEBUG_PAGEALLOC
2149 if (cpu_pgsz_mask & HV_PGSZ_MASK_256MB) {
2150 kern_linear_pte_xor[1] = (_PAGE_VALID | _PAGE_SZ256MB_4V) ^
2152 kern_linear_pte_xor[1] |= (_PAGE_CP_4V | pagecv_flag |
2153 _PAGE_P_4V | _PAGE_W_4V);
2155 kern_linear_pte_xor[1] = kern_linear_pte_xor[0];
2158 if (cpu_pgsz_mask & HV_PGSZ_MASK_2GB) {
2159 kern_linear_pte_xor[2] = (_PAGE_VALID | _PAGE_SZ2GB_4V) ^
2161 kern_linear_pte_xor[2] |= (_PAGE_CP_4V | pagecv_flag |
2162 _PAGE_P_4V | _PAGE_W_4V);
2164 kern_linear_pte_xor[2] = kern_linear_pte_xor[1];
2167 if (cpu_pgsz_mask & HV_PGSZ_MASK_16GB) {
2168 kern_linear_pte_xor[3] = (_PAGE_VALID | _PAGE_SZ16GB_4V) ^
2170 kern_linear_pte_xor[3] |= (_PAGE_CP_4V | pagecv_flag |
2171 _PAGE_P_4V | _PAGE_W_4V);
2173 kern_linear_pte_xor[3] = kern_linear_pte_xor[2];
2178 /* paging_init() sets up the page tables */
2180 static unsigned long last_valid_pfn;
2182 static void sun4u_pgprot_init(void);
2183 static void sun4v_pgprot_init(void);
2185 static phys_addr_t __init available_memory(void)
2187 phys_addr_t available = 0ULL;
2188 phys_addr_t pa_start, pa_end;
2191 for_each_free_mem_range(i, NUMA_NO_NODE, MEMBLOCK_NONE, &pa_start,
2193 available = available + (pa_end - pa_start);
2198 #define _PAGE_CACHE_4U (_PAGE_CP_4U | _PAGE_CV_4U)
2199 #define _PAGE_CACHE_4V (_PAGE_CP_4V | _PAGE_CV_4V)
2200 #define __DIRTY_BITS_4U (_PAGE_MODIFIED_4U | _PAGE_WRITE_4U | _PAGE_W_4U)
2201 #define __DIRTY_BITS_4V (_PAGE_MODIFIED_4V | _PAGE_WRITE_4V | _PAGE_W_4V)
2202 #define __ACCESS_BITS_4U (_PAGE_ACCESSED_4U | _PAGE_READ_4U | _PAGE_R)
2203 #define __ACCESS_BITS_4V (_PAGE_ACCESSED_4V | _PAGE_READ_4V | _PAGE_R)
2205 /* We need to exclude reserved regions. This exclusion will include
2206 * vmlinux and initrd. To be more precise the initrd size could be used to
2207 * compute a new lower limit because it is freed later during initialization.
2209 static void __init reduce_memory(phys_addr_t limit_ram)
2211 phys_addr_t avail_ram = available_memory();
2212 phys_addr_t pa_start, pa_end;
2215 if (limit_ram >= avail_ram)
2218 for_each_free_mem_range(i, NUMA_NO_NODE, MEMBLOCK_NONE, &pa_start,
2220 phys_addr_t region_size = pa_end - pa_start;
2221 phys_addr_t clip_start = pa_start;
2223 avail_ram = avail_ram - region_size;
2224 /* Are we consuming too much? */
2225 if (avail_ram < limit_ram) {
2226 phys_addr_t give_back = limit_ram - avail_ram;
2228 region_size = region_size - give_back;
2229 clip_start = clip_start + give_back;
2232 memblock_remove(clip_start, region_size);
2234 if (avail_ram <= limit_ram)
2240 void __init paging_init(void)
2242 unsigned long end_pfn, shift, phys_base;
2243 unsigned long real_end, i;
2245 setup_page_offset();
2247 /* These build time checkes make sure that the dcache_dirty_cpu()
2248 * page->flags usage will work.
2250 * When a page gets marked as dcache-dirty, we store the
2251 * cpu number starting at bit 32 in the page->flags. Also,
2252 * functions like clear_dcache_dirty_cpu use the cpu mask
2253 * in 13-bit signed-immediate instruction fields.
2257 * Page flags must not reach into upper 32 bits that are used
2258 * for the cpu number
2260 BUILD_BUG_ON(NR_PAGEFLAGS > 32);
2263 * The bit fields placed in the high range must not reach below
2264 * the 32 bit boundary. Otherwise we cannot place the cpu field
2265 * at the 32 bit boundary.
2267 BUILD_BUG_ON(SECTIONS_WIDTH + NODES_WIDTH + ZONES_WIDTH +
2268 ilog2(roundup_pow_of_two(NR_CPUS)) > 32);
2270 BUILD_BUG_ON(NR_CPUS > 4096);
2272 kern_base = (prom_boot_mapping_phys_low >> ILOG2_4MB) << ILOG2_4MB;
2273 kern_size = (unsigned long)&_end - (unsigned long)KERNBASE;
2275 /* Invalidate both kernel TSBs. */
2276 memset(swapper_tsb, 0x40, sizeof(swapper_tsb));
2277 #ifndef CONFIG_DEBUG_PAGEALLOC
2278 memset(swapper_4m_tsb, 0x40, sizeof(swapper_4m_tsb));
2281 /* TTE.cv bit on sparc v9 occupies the same position as TTE.mcde
2282 * bit on M7 processor. This is a conflicting usage of the same
2283 * bit. Enabling TTE.cv on M7 would turn on Memory Corruption
2284 * Detection error on all pages and this will lead to problems
2285 * later. Kernel does not run with MCD enabled and hence rest
2286 * of the required steps to fully configure memory corruption
2287 * detection are not taken. We need to ensure TTE.mcde is not
2288 * set on M7 processor. Compute the value of cacheability
2289 * flag for use later taking this into consideration.
2291 switch (sun4v_chip_type) {
2292 case SUN4V_CHIP_SPARC_M7:
2293 case SUN4V_CHIP_SPARC_SN:
2294 page_cache4v_flag = _PAGE_CP_4V;
2297 page_cache4v_flag = _PAGE_CACHE_4V;
2301 if (tlb_type == hypervisor)
2302 sun4v_pgprot_init();
2304 sun4u_pgprot_init();
2306 if (tlb_type == cheetah_plus ||
2307 tlb_type == hypervisor) {
2312 if (tlb_type == hypervisor)
2313 sun4v_patch_tlb_handlers();
2315 /* Find available physical memory...
2317 * Read it twice in order to work around a bug in openfirmware.
2318 * The call to grab this table itself can cause openfirmware to
2319 * allocate memory, which in turn can take away some space from
2320 * the list of available memory. Reading it twice makes sure
2321 * we really do get the final value.
2323 read_obp_translations();
2324 read_obp_memory("reg", &pall[0], &pall_ents);
2325 read_obp_memory("available", &pavail[0], &pavail_ents);
2326 read_obp_memory("available", &pavail[0], &pavail_ents);
2328 phys_base = 0xffffffffffffffffUL;
2329 for (i = 0; i < pavail_ents; i++) {
2330 phys_base = min(phys_base, pavail[i].phys_addr);
2331 memblock_add(pavail[i].phys_addr, pavail[i].reg_size);
2334 memblock_reserve(kern_base, kern_size);
2336 find_ramdisk(phys_base);
2338 if (cmdline_memory_size)
2339 reduce_memory(cmdline_memory_size);
2341 memblock_allow_resize();
2342 memblock_dump_all();
2344 set_bit(0, mmu_context_bmap);
2346 shift = kern_base + PAGE_OFFSET - ((unsigned long)KERNBASE);
2348 real_end = (unsigned long)_end;
2349 num_kernel_image_mappings = DIV_ROUND_UP(real_end - KERNBASE, 1 << ILOG2_4MB);
2350 printk("Kernel: Using %d locked TLB entries for main kernel image.\n",
2351 num_kernel_image_mappings);
2353 /* Set kernel pgd to upper alias so physical page computations
2356 init_mm.pgd += ((shift) / (sizeof(pgd_t)));
2358 memset(swapper_pg_dir, 0, sizeof(swapper_pg_dir));
2360 inherit_prom_mappings();
2362 /* Ok, we can use our TLB miss and window trap handlers safely. */
2367 prom_build_devicetree();
2368 of_populate_present_mask();
2370 of_fill_in_cpu_data();
2373 if (tlb_type == hypervisor) {
2375 mdesc_populate_present_mask(cpu_all_mask);
2377 mdesc_fill_in_cpu_data(cpu_all_mask);
2379 mdesc_get_page_sizes(cpu_all_mask, &cpu_pgsz_mask);
2381 sun4v_linear_pte_xor_finalize();
2384 sun4v_ktsb_register();
2386 unsigned long impl, ver;
2388 cpu_pgsz_mask = (HV_PGSZ_MASK_8K | HV_PGSZ_MASK_64K |
2389 HV_PGSZ_MASK_512K | HV_PGSZ_MASK_4MB);
2391 __asm__ __volatile__("rdpr %%ver, %0" : "=r" (ver));
2392 impl = ((ver >> 32) & 0xffff);
2393 if (impl == PANTHER_IMPL)
2394 cpu_pgsz_mask |= (HV_PGSZ_MASK_32MB |
2395 HV_PGSZ_MASK_256MB);
2397 sun4u_linear_pte_xor_finalize();
2400 /* Flush the TLBs and the 4M TSB so that the updated linear
2401 * pte XOR settings are realized for all mappings.
2404 #ifndef CONFIG_DEBUG_PAGEALLOC
2405 memset(swapper_4m_tsb, 0x40, sizeof(swapper_4m_tsb));
2409 /* Setup bootmem... */
2410 last_valid_pfn = end_pfn = bootmem_init(phys_base);
2412 kernel_physical_mapping_init();
2415 unsigned long max_zone_pfns[MAX_NR_ZONES];
2417 memset(max_zone_pfns, 0, sizeof(max_zone_pfns));
2419 max_zone_pfns[ZONE_NORMAL] = end_pfn;
2421 free_area_init_nodes(max_zone_pfns);
2424 printk("Booting Linux...\n");
2427 int page_in_phys_avail(unsigned long paddr)
2433 for (i = 0; i < pavail_ents; i++) {
2434 unsigned long start, end;
2436 start = pavail[i].phys_addr;
2437 end = start + pavail[i].reg_size;
2439 if (paddr >= start && paddr < end)
2442 if (paddr >= kern_base && paddr < (kern_base + kern_size))
2444 #ifdef CONFIG_BLK_DEV_INITRD
2445 if (paddr >= __pa(initrd_start) &&
2446 paddr < __pa(PAGE_ALIGN(initrd_end)))
2453 static void __init register_page_bootmem_info(void)
2455 #ifdef CONFIG_NEED_MULTIPLE_NODES
2458 for_each_online_node(i)
2459 if (NODE_DATA(i)->node_spanned_pages)
2460 register_page_bootmem_info_node(NODE_DATA(i));
2463 void __init mem_init(void)
2465 high_memory = __va(last_valid_pfn << PAGE_SHIFT);
2467 register_page_bootmem_info();
2471 * Set up the zero page, mark it reserved, so that page count
2472 * is not manipulated when freeing the page from user ptes.
2474 mem_map_zero = alloc_pages(GFP_KERNEL|__GFP_ZERO, 0);
2475 if (mem_map_zero == NULL) {
2476 prom_printf("paging_init: Cannot alloc zero page.\n");
2479 mark_page_reserved(mem_map_zero);
2481 mem_init_print_info(NULL);
2483 if (tlb_type == cheetah || tlb_type == cheetah_plus)
2484 cheetah_ecache_flush_init();
2487 void free_initmem(void)
2489 unsigned long addr, initend;
2492 /* If the physical memory maps were trimmed by kernel command
2493 * line options, don't even try freeing this initmem stuff up.
2494 * The kernel image could have been in the trimmed out region
2495 * and if so the freeing below will free invalid page structs.
2497 if (cmdline_memory_size)
2501 * The init section is aligned to 8k in vmlinux.lds. Page align for >8k pagesizes.
2503 addr = PAGE_ALIGN((unsigned long)(__init_begin));
2504 initend = (unsigned long)(__init_end) & PAGE_MASK;
2505 for (; addr < initend; addr += PAGE_SIZE) {
2509 ((unsigned long) __va(kern_base)) -
2510 ((unsigned long) KERNBASE));
2511 memset((void *)addr, POISON_FREE_INITMEM, PAGE_SIZE);
2514 free_reserved_page(virt_to_page(page));
2518 #ifdef CONFIG_BLK_DEV_INITRD
2519 void free_initrd_mem(unsigned long start, unsigned long end)
2521 free_reserved_area((void *)start, (void *)end, POISON_FREE_INITMEM,
2526 pgprot_t PAGE_KERNEL __read_mostly;
2527 EXPORT_SYMBOL(PAGE_KERNEL);
2529 pgprot_t PAGE_KERNEL_LOCKED __read_mostly;
2530 pgprot_t PAGE_COPY __read_mostly;
2532 pgprot_t PAGE_SHARED __read_mostly;
2533 EXPORT_SYMBOL(PAGE_SHARED);
2535 unsigned long pg_iobits __read_mostly;
2537 unsigned long _PAGE_IE __read_mostly;
2538 EXPORT_SYMBOL(_PAGE_IE);
2540 unsigned long _PAGE_E __read_mostly;
2541 EXPORT_SYMBOL(_PAGE_E);
2543 unsigned long _PAGE_CACHE __read_mostly;
2544 EXPORT_SYMBOL(_PAGE_CACHE);
2546 #ifdef CONFIG_SPARSEMEM_VMEMMAP
2547 int __meminit vmemmap_populate(unsigned long vstart, unsigned long vend,
2550 unsigned long pte_base;
2552 pte_base = (_PAGE_VALID | _PAGE_SZ4MB_4U |
2553 _PAGE_CP_4U | _PAGE_CV_4U |
2554 _PAGE_P_4U | _PAGE_W_4U);
2555 if (tlb_type == hypervisor)
2556 pte_base = (_PAGE_VALID | _PAGE_SZ4MB_4V |
2557 page_cache4v_flag | _PAGE_P_4V | _PAGE_W_4V);
2559 pte_base |= _PAGE_PMD_HUGE;
2561 vstart = vstart & PMD_MASK;
2562 vend = ALIGN(vend, PMD_SIZE);
2563 for (; vstart < vend; vstart += PMD_SIZE) {
2564 pgd_t *pgd = pgd_offset_k(vstart);
2569 if (pgd_none(*pgd)) {
2570 pud_t *new = vmemmap_alloc_block(PAGE_SIZE, node);
2574 pgd_populate(&init_mm, pgd, new);
2577 pud = pud_offset(pgd, vstart);
2578 if (pud_none(*pud)) {
2579 pmd_t *new = vmemmap_alloc_block(PAGE_SIZE, node);
2583 pud_populate(&init_mm, pud, new);
2586 pmd = pmd_offset(pud, vstart);
2588 pte = pmd_val(*pmd);
2589 if (!(pte & _PAGE_VALID)) {
2590 void *block = vmemmap_alloc_block(PMD_SIZE, node);
2595 pmd_val(*pmd) = pte_base | __pa(block);
2602 void vmemmap_free(unsigned long start, unsigned long end)
2605 #endif /* CONFIG_SPARSEMEM_VMEMMAP */
2607 static void prot_init_common(unsigned long page_none,
2608 unsigned long page_shared,
2609 unsigned long page_copy,
2610 unsigned long page_readonly,
2611 unsigned long page_exec_bit)
2613 PAGE_COPY = __pgprot(page_copy);
2614 PAGE_SHARED = __pgprot(page_shared);
2616 protection_map[0x0] = __pgprot(page_none);
2617 protection_map[0x1] = __pgprot(page_readonly & ~page_exec_bit);
2618 protection_map[0x2] = __pgprot(page_copy & ~page_exec_bit);
2619 protection_map[0x3] = __pgprot(page_copy & ~page_exec_bit);
2620 protection_map[0x4] = __pgprot(page_readonly);
2621 protection_map[0x5] = __pgprot(page_readonly);
2622 protection_map[0x6] = __pgprot(page_copy);
2623 protection_map[0x7] = __pgprot(page_copy);
2624 protection_map[0x8] = __pgprot(page_none);
2625 protection_map[0x9] = __pgprot(page_readonly & ~page_exec_bit);
2626 protection_map[0xa] = __pgprot(page_shared & ~page_exec_bit);
2627 protection_map[0xb] = __pgprot(page_shared & ~page_exec_bit);
2628 protection_map[0xc] = __pgprot(page_readonly);
2629 protection_map[0xd] = __pgprot(page_readonly);
2630 protection_map[0xe] = __pgprot(page_shared);
2631 protection_map[0xf] = __pgprot(page_shared);
2634 static void __init sun4u_pgprot_init(void)
2636 unsigned long page_none, page_shared, page_copy, page_readonly;
2637 unsigned long page_exec_bit;
2640 PAGE_KERNEL = __pgprot (_PAGE_PRESENT_4U | _PAGE_VALID |
2641 _PAGE_CACHE_4U | _PAGE_P_4U |
2642 __ACCESS_BITS_4U | __DIRTY_BITS_4U |
2644 PAGE_KERNEL_LOCKED = __pgprot (_PAGE_PRESENT_4U | _PAGE_VALID |
2645 _PAGE_CACHE_4U | _PAGE_P_4U |
2646 __ACCESS_BITS_4U | __DIRTY_BITS_4U |
2647 _PAGE_EXEC_4U | _PAGE_L_4U);
2649 _PAGE_IE = _PAGE_IE_4U;
2650 _PAGE_E = _PAGE_E_4U;
2651 _PAGE_CACHE = _PAGE_CACHE_4U;
2653 pg_iobits = (_PAGE_VALID | _PAGE_PRESENT_4U | __DIRTY_BITS_4U |
2654 __ACCESS_BITS_4U | _PAGE_E_4U);
2656 #ifdef CONFIG_DEBUG_PAGEALLOC
2657 kern_linear_pte_xor[0] = _PAGE_VALID ^ PAGE_OFFSET;
2659 kern_linear_pte_xor[0] = (_PAGE_VALID | _PAGE_SZ4MB_4U) ^
2662 kern_linear_pte_xor[0] |= (_PAGE_CP_4U | _PAGE_CV_4U |
2663 _PAGE_P_4U | _PAGE_W_4U);
2665 for (i = 1; i < 4; i++)
2666 kern_linear_pte_xor[i] = kern_linear_pte_xor[0];
2668 _PAGE_ALL_SZ_BITS = (_PAGE_SZ4MB_4U | _PAGE_SZ512K_4U |
2669 _PAGE_SZ64K_4U | _PAGE_SZ8K_4U |
2670 _PAGE_SZ32MB_4U | _PAGE_SZ256MB_4U);
2673 page_none = _PAGE_PRESENT_4U | _PAGE_ACCESSED_4U | _PAGE_CACHE_4U;
2674 page_shared = (_PAGE_VALID | _PAGE_PRESENT_4U | _PAGE_CACHE_4U |
2675 __ACCESS_BITS_4U | _PAGE_WRITE_4U | _PAGE_EXEC_4U);
2676 page_copy = (_PAGE_VALID | _PAGE_PRESENT_4U | _PAGE_CACHE_4U |
2677 __ACCESS_BITS_4U | _PAGE_EXEC_4U);
2678 page_readonly = (_PAGE_VALID | _PAGE_PRESENT_4U | _PAGE_CACHE_4U |
2679 __ACCESS_BITS_4U | _PAGE_EXEC_4U);
2681 page_exec_bit = _PAGE_EXEC_4U;
2683 prot_init_common(page_none, page_shared, page_copy, page_readonly,
2687 static void __init sun4v_pgprot_init(void)
2689 unsigned long page_none, page_shared, page_copy, page_readonly;
2690 unsigned long page_exec_bit;
2693 PAGE_KERNEL = __pgprot (_PAGE_PRESENT_4V | _PAGE_VALID |
2694 page_cache4v_flag | _PAGE_P_4V |
2695 __ACCESS_BITS_4V | __DIRTY_BITS_4V |
2697 PAGE_KERNEL_LOCKED = PAGE_KERNEL;
2699 _PAGE_IE = _PAGE_IE_4V;
2700 _PAGE_E = _PAGE_E_4V;
2701 _PAGE_CACHE = page_cache4v_flag;
2703 #ifdef CONFIG_DEBUG_PAGEALLOC
2704 kern_linear_pte_xor[0] = _PAGE_VALID ^ PAGE_OFFSET;
2706 kern_linear_pte_xor[0] = (_PAGE_VALID | _PAGE_SZ4MB_4V) ^
2709 kern_linear_pte_xor[0] |= (page_cache4v_flag | _PAGE_P_4V |
2712 for (i = 1; i < 4; i++)
2713 kern_linear_pte_xor[i] = kern_linear_pte_xor[0];
2715 pg_iobits = (_PAGE_VALID | _PAGE_PRESENT_4V | __DIRTY_BITS_4V |
2716 __ACCESS_BITS_4V | _PAGE_E_4V);
2718 _PAGE_ALL_SZ_BITS = (_PAGE_SZ16GB_4V | _PAGE_SZ2GB_4V |
2719 _PAGE_SZ256MB_4V | _PAGE_SZ32MB_4V |
2720 _PAGE_SZ4MB_4V | _PAGE_SZ512K_4V |
2721 _PAGE_SZ64K_4V | _PAGE_SZ8K_4V);
2723 page_none = _PAGE_PRESENT_4V | _PAGE_ACCESSED_4V | page_cache4v_flag;
2724 page_shared = (_PAGE_VALID | _PAGE_PRESENT_4V | page_cache4v_flag |
2725 __ACCESS_BITS_4V | _PAGE_WRITE_4V | _PAGE_EXEC_4V);
2726 page_copy = (_PAGE_VALID | _PAGE_PRESENT_4V | page_cache4v_flag |
2727 __ACCESS_BITS_4V | _PAGE_EXEC_4V);
2728 page_readonly = (_PAGE_VALID | _PAGE_PRESENT_4V | page_cache4v_flag |
2729 __ACCESS_BITS_4V | _PAGE_EXEC_4V);
2731 page_exec_bit = _PAGE_EXEC_4V;
2733 prot_init_common(page_none, page_shared, page_copy, page_readonly,
2737 unsigned long pte_sz_bits(unsigned long sz)
2739 if (tlb_type == hypervisor) {
2743 return _PAGE_SZ8K_4V;
2745 return _PAGE_SZ64K_4V;
2747 return _PAGE_SZ512K_4V;
2748 case 4 * 1024 * 1024:
2749 return _PAGE_SZ4MB_4V;
2755 return _PAGE_SZ8K_4U;
2757 return _PAGE_SZ64K_4U;
2759 return _PAGE_SZ512K_4U;
2760 case 4 * 1024 * 1024:
2761 return _PAGE_SZ4MB_4U;
2766 pte_t mk_pte_io(unsigned long page, pgprot_t prot, int space, unsigned long page_size)
2770 pte_val(pte) = page | pgprot_val(pgprot_noncached(prot));
2771 pte_val(pte) |= (((unsigned long)space) << 32);
2772 pte_val(pte) |= pte_sz_bits(page_size);
2777 static unsigned long kern_large_tte(unsigned long paddr)
2781 val = (_PAGE_VALID | _PAGE_SZ4MB_4U |
2782 _PAGE_CP_4U | _PAGE_CV_4U | _PAGE_P_4U |
2783 _PAGE_EXEC_4U | _PAGE_L_4U | _PAGE_W_4U);
2784 if (tlb_type == hypervisor)
2785 val = (_PAGE_VALID | _PAGE_SZ4MB_4V |
2786 page_cache4v_flag | _PAGE_P_4V |
2787 _PAGE_EXEC_4V | _PAGE_W_4V);
2792 /* If not locked, zap it. */
2793 void __flush_tlb_all(void)
2795 unsigned long pstate;
2798 __asm__ __volatile__("flushw\n\t"
2799 "rdpr %%pstate, %0\n\t"
2800 "wrpr %0, %1, %%pstate"
2803 if (tlb_type == hypervisor) {
2804 sun4v_mmu_demap_all();
2805 } else if (tlb_type == spitfire) {
2806 for (i = 0; i < 64; i++) {
2807 /* Spitfire Errata #32 workaround */
2808 /* NOTE: Always runs on spitfire, so no
2809 * cheetah+ page size encodings.
2811 __asm__ __volatile__("stxa %0, [%1] %2\n\t"
2815 "r" (PRIMARY_CONTEXT), "i" (ASI_DMMU));
2817 if (!(spitfire_get_dtlb_data(i) & _PAGE_L_4U)) {
2818 __asm__ __volatile__("stxa %%g0, [%0] %1\n\t"
2821 : "r" (TLB_TAG_ACCESS), "i" (ASI_DMMU));
2822 spitfire_put_dtlb_data(i, 0x0UL);
2825 /* Spitfire Errata #32 workaround */
2826 /* NOTE: Always runs on spitfire, so no
2827 * cheetah+ page size encodings.
2829 __asm__ __volatile__("stxa %0, [%1] %2\n\t"
2833 "r" (PRIMARY_CONTEXT), "i" (ASI_DMMU));
2835 if (!(spitfire_get_itlb_data(i) & _PAGE_L_4U)) {
2836 __asm__ __volatile__("stxa %%g0, [%0] %1\n\t"
2839 : "r" (TLB_TAG_ACCESS), "i" (ASI_IMMU));
2840 spitfire_put_itlb_data(i, 0x0UL);
2843 } else if (tlb_type == cheetah || tlb_type == cheetah_plus) {
2844 cheetah_flush_dtlb_all();
2845 cheetah_flush_itlb_all();
2847 __asm__ __volatile__("wrpr %0, 0, %%pstate"
2851 pte_t *pte_alloc_one_kernel(struct mm_struct *mm,
2852 unsigned long address)
2854 struct page *page = alloc_page(GFP_KERNEL | __GFP_NOTRACK | __GFP_ZERO);
2858 pte = (pte_t *) page_address(page);
2863 pgtable_t pte_alloc_one(struct mm_struct *mm,
2864 unsigned long address)
2866 struct page *page = alloc_page(GFP_KERNEL | __GFP_NOTRACK | __GFP_ZERO);
2869 if (!pgtable_page_ctor(page)) {
2870 free_hot_cold_page(page, 0);
2873 return (pte_t *) page_address(page);
2876 void pte_free_kernel(struct mm_struct *mm, pte_t *pte)
2878 free_page((unsigned long)pte);
2881 static void __pte_free(pgtable_t pte)
2883 struct page *page = virt_to_page(pte);
2885 pgtable_page_dtor(page);
2889 void pte_free(struct mm_struct *mm, pgtable_t pte)
2894 void pgtable_free(void *table, bool is_page)
2899 kmem_cache_free(pgtable_cache, table);
2902 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
2903 void update_mmu_cache_pmd(struct vm_area_struct *vma, unsigned long addr,
2906 unsigned long pte, flags;
2907 struct mm_struct *mm;
2910 if (!pmd_large(entry) || !pmd_young(entry))
2913 pte = pmd_val(entry);
2915 /* Don't insert a non-valid PMD into the TSB, we'll deadlock. */
2916 if (!(pte & _PAGE_VALID))
2919 /* We are fabricating 8MB pages using 4MB real hw pages. */
2920 pte |= (addr & (1UL << REAL_HPAGE_SHIFT));
2924 spin_lock_irqsave(&mm->context.lock, flags);
2926 if (mm->context.tsb_block[MM_TSB_HUGE].tsb != NULL)
2927 __update_mmu_tsb_insert(mm, MM_TSB_HUGE, REAL_HPAGE_SHIFT,
2930 spin_unlock_irqrestore(&mm->context.lock, flags);
2932 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
2934 #if defined(CONFIG_HUGETLB_PAGE) || defined(CONFIG_TRANSPARENT_HUGEPAGE)
2935 static void context_reload(void *__data)
2937 struct mm_struct *mm = __data;
2939 if (mm == current->mm)
2940 load_secondary_context(mm);
2943 void hugetlb_setup(struct pt_regs *regs)
2945 struct mm_struct *mm = current->mm;
2946 struct tsb_config *tp;
2948 if (faulthandler_disabled() || !mm) {
2949 const struct exception_table_entry *entry;
2951 entry = search_exception_tables(regs->tpc);
2953 regs->tpc = entry->fixup;
2954 regs->tnpc = regs->tpc + 4;
2957 pr_alert("Unexpected HugeTLB setup in atomic context.\n");
2958 die_if_kernel("HugeTSB in atomic", regs);
2961 tp = &mm->context.tsb_block[MM_TSB_HUGE];
2962 if (likely(tp->tsb == NULL))
2963 tsb_grow(mm, MM_TSB_HUGE, 0);
2965 tsb_context_switch(mm);
2968 /* On UltraSPARC-III+ and later, configure the second half of
2969 * the Data-TLB for huge pages.
2971 if (tlb_type == cheetah_plus) {
2972 bool need_context_reload = false;
2975 spin_lock_irq(&ctx_alloc_lock);
2976 ctx = mm->context.sparc64_ctx_val;
2977 ctx &= ~CTX_PGSZ_MASK;
2978 ctx |= CTX_PGSZ_BASE << CTX_PGSZ0_SHIFT;
2979 ctx |= CTX_PGSZ_HUGE << CTX_PGSZ1_SHIFT;
2981 if (ctx != mm->context.sparc64_ctx_val) {
2982 /* When changing the page size fields, we
2983 * must perform a context flush so that no
2984 * stale entries match. This flush must
2985 * occur with the original context register
2988 do_flush_tlb_mm(mm);
2990 /* Reload the context register of all processors
2991 * also executing in this address space.
2993 mm->context.sparc64_ctx_val = ctx;
2994 need_context_reload = true;
2996 spin_unlock_irq(&ctx_alloc_lock);
2998 if (need_context_reload)
2999 on_each_cpu(context_reload, mm, 0);
3004 static struct resource code_resource = {
3005 .name = "Kernel code",
3006 .flags = IORESOURCE_BUSY | IORESOURCE_SYSTEM_RAM
3009 static struct resource data_resource = {
3010 .name = "Kernel data",
3011 .flags = IORESOURCE_BUSY | IORESOURCE_SYSTEM_RAM
3014 static struct resource bss_resource = {
3015 .name = "Kernel bss",
3016 .flags = IORESOURCE_BUSY | IORESOURCE_SYSTEM_RAM
3019 static inline resource_size_t compute_kern_paddr(void *addr)
3021 return (resource_size_t) (addr - KERNBASE + kern_base);
3024 static void __init kernel_lds_init(void)
3026 code_resource.start = compute_kern_paddr(_text);
3027 code_resource.end = compute_kern_paddr(_etext - 1);
3028 data_resource.start = compute_kern_paddr(_etext);
3029 data_resource.end = compute_kern_paddr(_edata - 1);
3030 bss_resource.start = compute_kern_paddr(__bss_start);
3031 bss_resource.end = compute_kern_paddr(_end - 1);
3034 static int __init report_memory(void)
3037 struct resource *res;
3041 for (i = 0; i < pavail_ents; i++) {
3042 res = kzalloc(sizeof(struct resource), GFP_KERNEL);
3045 pr_warn("Failed to allocate source.\n");
3049 res->name = "System RAM";
3050 res->start = pavail[i].phys_addr;
3051 res->end = pavail[i].phys_addr + pavail[i].reg_size - 1;
3052 res->flags = IORESOURCE_BUSY | IORESOURCE_SYSTEM_RAM;
3054 if (insert_resource(&iomem_resource, res) < 0) {
3055 pr_warn("Resource insertion failed.\n");
3059 insert_resource(res, &code_resource);
3060 insert_resource(res, &data_resource);
3061 insert_resource(res, &bss_resource);
3066 arch_initcall(report_memory);
3069 #define do_flush_tlb_kernel_range smp_flush_tlb_kernel_range
3071 #define do_flush_tlb_kernel_range __flush_tlb_kernel_range
3074 void flush_tlb_kernel_range(unsigned long start, unsigned long end)
3076 if (start < HI_OBP_ADDRESS && end > LOW_OBP_ADDRESS) {
3077 if (start < LOW_OBP_ADDRESS) {
3078 flush_tsb_kernel_range(start, LOW_OBP_ADDRESS);
3079 do_flush_tlb_kernel_range(start, LOW_OBP_ADDRESS);
3081 if (end > HI_OBP_ADDRESS) {
3082 flush_tsb_kernel_range(HI_OBP_ADDRESS, end);
3083 do_flush_tlb_kernel_range(HI_OBP_ADDRESS, end);
3086 flush_tsb_kernel_range(start, end);
3087 do_flush_tlb_kernel_range(start, end);