Merge branch 'for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/viro/vfs
[linux/fpc-iii.git] / arch / sparc / mm / init_64.c
blobeafbc65c9c47f63772162d384ef55fd549d1e149
1 /*
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)
6 */
8 #include <linux/module.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>
14 #include <linux/mm.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>
20 #include <linux/fs.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/percpu.h>
26 #include <linux/memblock.h>
27 #include <linux/mmzone.h>
28 #include <linux/gfp.h>
30 #include <asm/head.h>
31 #include <asm/page.h>
32 #include <asm/pgalloc.h>
33 #include <asm/pgtable.h>
34 #include <asm/oplib.h>
35 #include <asm/iommu.h>
36 #include <asm/io.h>
37 #include <asm/uaccess.h>
38 #include <asm/mmu_context.h>
39 #include <asm/tlbflush.h>
40 #include <asm/dma.h>
41 #include <asm/starfire.h>
42 #include <asm/tlb.h>
43 #include <asm/spitfire.h>
44 #include <asm/sections.h>
45 #include <asm/tsb.h>
46 #include <asm/hypervisor.h>
47 #include <asm/prom.h>
48 #include <asm/mdesc.h>
49 #include <asm/cpudata.h>
50 #include <asm/irq.h>
52 #include "init_64.h"
54 unsigned long kern_linear_pte_xor[4] __read_mostly;
56 /* A bitmap, two bits for every 256MB of physical memory. These two
57 * bits determine what page size we use for kernel linear
58 * translations. They form an index into kern_linear_pte_xor[]. The
59 * value in the indexed slot is XOR'd with the TLB miss virtual
60 * address to form the resulting TTE. The mapping is:
62 * 0 ==> 4MB
63 * 1 ==> 256MB
64 * 2 ==> 2GB
65 * 3 ==> 16GB
67 * All sun4v chips support 256MB pages. Only SPARC-T4 and later
68 * support 2GB pages, and hopefully future cpus will support the 16GB
69 * pages as well. For slots 2 and 3, we encode a 256MB TTE xor there
70 * if these larger page sizes are not supported by the cpu.
72 * It would be nice to determine this from the machine description
73 * 'cpu' properties, but we need to have this table setup before the
74 * MDESC is initialized.
76 unsigned long kpte_linear_bitmap[KPTE_BITMAP_BYTES / sizeof(unsigned long)];
78 #ifndef CONFIG_DEBUG_PAGEALLOC
79 /* A special kernel TSB for 4MB, 256MB, 2GB and 16GB linear mappings.
80 * Space is allocated for this right after the trap table in
81 * arch/sparc64/kernel/head.S
83 extern struct tsb swapper_4m_tsb[KERNEL_TSB4M_NENTRIES];
84 #endif
86 static unsigned long cpu_pgsz_mask;
88 #define MAX_BANKS 32
90 static struct linux_prom64_registers pavail[MAX_BANKS];
91 static int pavail_ents;
93 static int cmp_p64(const void *a, const void *b)
95 const struct linux_prom64_registers *x = a, *y = b;
97 if (x->phys_addr > y->phys_addr)
98 return 1;
99 if (x->phys_addr < y->phys_addr)
100 return -1;
101 return 0;
104 static void __init read_obp_memory(const char *property,
105 struct linux_prom64_registers *regs,
106 int *num_ents)
108 phandle node = prom_finddevice("/memory");
109 int prop_size = prom_getproplen(node, property);
110 int ents, ret, i;
112 ents = prop_size / sizeof(struct linux_prom64_registers);
113 if (ents > MAX_BANKS) {
114 prom_printf("The machine has more %s property entries than "
115 "this kernel can support (%d).\n",
116 property, MAX_BANKS);
117 prom_halt();
120 ret = prom_getproperty(node, property, (char *) regs, prop_size);
121 if (ret == -1) {
122 prom_printf("Couldn't get %s property from /memory.\n",
123 property);
124 prom_halt();
127 /* Sanitize what we got from the firmware, by page aligning
128 * everything.
130 for (i = 0; i < ents; i++) {
131 unsigned long base, size;
133 base = regs[i].phys_addr;
134 size = regs[i].reg_size;
136 size &= PAGE_MASK;
137 if (base & ~PAGE_MASK) {
138 unsigned long new_base = PAGE_ALIGN(base);
140 size -= new_base - base;
141 if ((long) size < 0L)
142 size = 0UL;
143 base = new_base;
145 if (size == 0UL) {
146 /* If it is empty, simply get rid of it.
147 * This simplifies the logic of the other
148 * functions that process these arrays.
150 memmove(&regs[i], &regs[i + 1],
151 (ents - i - 1) * sizeof(regs[0]));
152 i--;
153 ents--;
154 continue;
156 regs[i].phys_addr = base;
157 regs[i].reg_size = size;
160 *num_ents = ents;
162 sort(regs, ents, sizeof(struct linux_prom64_registers),
163 cmp_p64, NULL);
166 unsigned long sparc64_valid_addr_bitmap[VALID_ADDR_BITMAP_BYTES /
167 sizeof(unsigned long)];
168 EXPORT_SYMBOL(sparc64_valid_addr_bitmap);
170 /* Kernel physical address base and size in bytes. */
171 unsigned long kern_base __read_mostly;
172 unsigned long kern_size __read_mostly;
174 /* Initial ramdisk setup */
175 extern unsigned long sparc_ramdisk_image64;
176 extern unsigned int sparc_ramdisk_image;
177 extern unsigned int sparc_ramdisk_size;
179 struct page *mem_map_zero __read_mostly;
180 EXPORT_SYMBOL(mem_map_zero);
182 unsigned int sparc64_highest_unlocked_tlb_ent __read_mostly;
184 unsigned long sparc64_kern_pri_context __read_mostly;
185 unsigned long sparc64_kern_pri_nuc_bits __read_mostly;
186 unsigned long sparc64_kern_sec_context __read_mostly;
188 int num_kernel_image_mappings;
190 #ifdef CONFIG_DEBUG_DCFLUSH
191 atomic_t dcpage_flushes = ATOMIC_INIT(0);
192 #ifdef CONFIG_SMP
193 atomic_t dcpage_flushes_xcall = ATOMIC_INIT(0);
194 #endif
195 #endif
197 inline void flush_dcache_page_impl(struct page *page)
199 BUG_ON(tlb_type == hypervisor);
200 #ifdef CONFIG_DEBUG_DCFLUSH
201 atomic_inc(&dcpage_flushes);
202 #endif
204 #ifdef DCACHE_ALIASING_POSSIBLE
205 __flush_dcache_page(page_address(page),
206 ((tlb_type == spitfire) &&
207 page_mapping(page) != NULL));
208 #else
209 if (page_mapping(page) != NULL &&
210 tlb_type == spitfire)
211 __flush_icache_page(__pa(page_address(page)));
212 #endif
215 #define PG_dcache_dirty PG_arch_1
216 #define PG_dcache_cpu_shift 32UL
217 #define PG_dcache_cpu_mask \
218 ((1UL<<ilog2(roundup_pow_of_two(NR_CPUS)))-1UL)
220 #define dcache_dirty_cpu(page) \
221 (((page)->flags >> PG_dcache_cpu_shift) & PG_dcache_cpu_mask)
223 static inline void set_dcache_dirty(struct page *page, int this_cpu)
225 unsigned long mask = this_cpu;
226 unsigned long non_cpu_bits;
228 non_cpu_bits = ~(PG_dcache_cpu_mask << PG_dcache_cpu_shift);
229 mask = (mask << PG_dcache_cpu_shift) | (1UL << PG_dcache_dirty);
231 __asm__ __volatile__("1:\n\t"
232 "ldx [%2], %%g7\n\t"
233 "and %%g7, %1, %%g1\n\t"
234 "or %%g1, %0, %%g1\n\t"
235 "casx [%2], %%g7, %%g1\n\t"
236 "cmp %%g7, %%g1\n\t"
237 "bne,pn %%xcc, 1b\n\t"
238 " nop"
239 : /* no outputs */
240 : "r" (mask), "r" (non_cpu_bits), "r" (&page->flags)
241 : "g1", "g7");
244 static inline void clear_dcache_dirty_cpu(struct page *page, unsigned long cpu)
246 unsigned long mask = (1UL << PG_dcache_dirty);
248 __asm__ __volatile__("! test_and_clear_dcache_dirty\n"
249 "1:\n\t"
250 "ldx [%2], %%g7\n\t"
251 "srlx %%g7, %4, %%g1\n\t"
252 "and %%g1, %3, %%g1\n\t"
253 "cmp %%g1, %0\n\t"
254 "bne,pn %%icc, 2f\n\t"
255 " andn %%g7, %1, %%g1\n\t"
256 "casx [%2], %%g7, %%g1\n\t"
257 "cmp %%g7, %%g1\n\t"
258 "bne,pn %%xcc, 1b\n\t"
259 " nop\n"
260 "2:"
261 : /* no outputs */
262 : "r" (cpu), "r" (mask), "r" (&page->flags),
263 "i" (PG_dcache_cpu_mask),
264 "i" (PG_dcache_cpu_shift)
265 : "g1", "g7");
268 static inline void tsb_insert(struct tsb *ent, unsigned long tag, unsigned long pte)
270 unsigned long tsb_addr = (unsigned long) ent;
272 if (tlb_type == cheetah_plus || tlb_type == hypervisor)
273 tsb_addr = __pa(tsb_addr);
275 __tsb_insert(tsb_addr, tag, pte);
278 unsigned long _PAGE_ALL_SZ_BITS __read_mostly;
280 static void flush_dcache(unsigned long pfn)
282 struct page *page;
284 page = pfn_to_page(pfn);
285 if (page) {
286 unsigned long pg_flags;
288 pg_flags = page->flags;
289 if (pg_flags & (1UL << PG_dcache_dirty)) {
290 int cpu = ((pg_flags >> PG_dcache_cpu_shift) &
291 PG_dcache_cpu_mask);
292 int this_cpu = get_cpu();
294 /* This is just to optimize away some function calls
295 * in the SMP case.
297 if (cpu == this_cpu)
298 flush_dcache_page_impl(page);
299 else
300 smp_flush_dcache_page_impl(page, cpu);
302 clear_dcache_dirty_cpu(page, cpu);
304 put_cpu();
309 /* mm->context.lock must be held */
310 static void __update_mmu_tsb_insert(struct mm_struct *mm, unsigned long tsb_index,
311 unsigned long tsb_hash_shift, unsigned long address,
312 unsigned long tte)
314 struct tsb *tsb = mm->context.tsb_block[tsb_index].tsb;
315 unsigned long tag;
317 if (unlikely(!tsb))
318 return;
320 tsb += ((address >> tsb_hash_shift) &
321 (mm->context.tsb_block[tsb_index].tsb_nentries - 1UL));
322 tag = (address >> 22UL);
323 tsb_insert(tsb, tag, tte);
326 #if defined(CONFIG_HUGETLB_PAGE) || defined(CONFIG_TRANSPARENT_HUGEPAGE)
327 static inline bool is_hugetlb_pte(pte_t pte)
329 if ((tlb_type == hypervisor &&
330 (pte_val(pte) & _PAGE_SZALL_4V) == _PAGE_SZHUGE_4V) ||
331 (tlb_type != hypervisor &&
332 (pte_val(pte) & _PAGE_SZALL_4U) == _PAGE_SZHUGE_4U))
333 return true;
334 return false;
336 #endif
338 void update_mmu_cache(struct vm_area_struct *vma, unsigned long address, pte_t *ptep)
340 struct mm_struct *mm;
341 unsigned long flags;
342 pte_t pte = *ptep;
344 if (tlb_type != hypervisor) {
345 unsigned long pfn = pte_pfn(pte);
347 if (pfn_valid(pfn))
348 flush_dcache(pfn);
351 mm = vma->vm_mm;
353 spin_lock_irqsave(&mm->context.lock, flags);
355 #if defined(CONFIG_HUGETLB_PAGE) || defined(CONFIG_TRANSPARENT_HUGEPAGE)
356 if (mm->context.huge_pte_count && is_hugetlb_pte(pte))
357 __update_mmu_tsb_insert(mm, MM_TSB_HUGE, REAL_HPAGE_SHIFT,
358 address, pte_val(pte));
359 else
360 #endif
361 __update_mmu_tsb_insert(mm, MM_TSB_BASE, PAGE_SHIFT,
362 address, pte_val(pte));
364 spin_unlock_irqrestore(&mm->context.lock, flags);
367 void flush_dcache_page(struct page *page)
369 struct address_space *mapping;
370 int this_cpu;
372 if (tlb_type == hypervisor)
373 return;
375 /* Do not bother with the expensive D-cache flush if it
376 * is merely the zero page. The 'bigcore' testcase in GDB
377 * causes this case to run millions of times.
379 if (page == ZERO_PAGE(0))
380 return;
382 this_cpu = get_cpu();
384 mapping = page_mapping(page);
385 if (mapping && !mapping_mapped(mapping)) {
386 int dirty = test_bit(PG_dcache_dirty, &page->flags);
387 if (dirty) {
388 int dirty_cpu = dcache_dirty_cpu(page);
390 if (dirty_cpu == this_cpu)
391 goto out;
392 smp_flush_dcache_page_impl(page, dirty_cpu);
394 set_dcache_dirty(page, this_cpu);
395 } else {
396 /* We could delay the flush for the !page_mapping
397 * case too. But that case is for exec env/arg
398 * pages and those are %99 certainly going to get
399 * faulted into the tlb (and thus flushed) anyways.
401 flush_dcache_page_impl(page);
404 out:
405 put_cpu();
407 EXPORT_SYMBOL(flush_dcache_page);
409 void __kprobes flush_icache_range(unsigned long start, unsigned long end)
411 /* Cheetah and Hypervisor platform cpus have coherent I-cache. */
412 if (tlb_type == spitfire) {
413 unsigned long kaddr;
415 /* This code only runs on Spitfire cpus so this is
416 * why we can assume _PAGE_PADDR_4U.
418 for (kaddr = start; kaddr < end; kaddr += PAGE_SIZE) {
419 unsigned long paddr, mask = _PAGE_PADDR_4U;
421 if (kaddr >= PAGE_OFFSET)
422 paddr = kaddr & mask;
423 else {
424 pgd_t *pgdp = pgd_offset_k(kaddr);
425 pud_t *pudp = pud_offset(pgdp, kaddr);
426 pmd_t *pmdp = pmd_offset(pudp, kaddr);
427 pte_t *ptep = pte_offset_kernel(pmdp, kaddr);
429 paddr = pte_val(*ptep) & mask;
431 __flush_icache_page(paddr);
435 EXPORT_SYMBOL(flush_icache_range);
437 void mmu_info(struct seq_file *m)
439 static const char *pgsz_strings[] = {
440 "8K", "64K", "512K", "4MB", "32MB",
441 "256MB", "2GB", "16GB",
443 int i, printed;
445 if (tlb_type == cheetah)
446 seq_printf(m, "MMU Type\t: Cheetah\n");
447 else if (tlb_type == cheetah_plus)
448 seq_printf(m, "MMU Type\t: Cheetah+\n");
449 else if (tlb_type == spitfire)
450 seq_printf(m, "MMU Type\t: Spitfire\n");
451 else if (tlb_type == hypervisor)
452 seq_printf(m, "MMU Type\t: Hypervisor (sun4v)\n");
453 else
454 seq_printf(m, "MMU Type\t: ???\n");
456 seq_printf(m, "MMU PGSZs\t: ");
457 printed = 0;
458 for (i = 0; i < ARRAY_SIZE(pgsz_strings); i++) {
459 if (cpu_pgsz_mask & (1UL << i)) {
460 seq_printf(m, "%s%s",
461 printed ? "," : "", pgsz_strings[i]);
462 printed++;
465 seq_putc(m, '\n');
467 #ifdef CONFIG_DEBUG_DCFLUSH
468 seq_printf(m, "DCPageFlushes\t: %d\n",
469 atomic_read(&dcpage_flushes));
470 #ifdef CONFIG_SMP
471 seq_printf(m, "DCPageFlushesXC\t: %d\n",
472 atomic_read(&dcpage_flushes_xcall));
473 #endif /* CONFIG_SMP */
474 #endif /* CONFIG_DEBUG_DCFLUSH */
477 struct linux_prom_translation prom_trans[512] __read_mostly;
478 unsigned int prom_trans_ents __read_mostly;
480 unsigned long kern_locked_tte_data;
482 /* The obp translations are saved based on 8k pagesize, since obp can
483 * use a mixture of pagesizes. Misses to the LOW_OBP_ADDRESS ->
484 * HI_OBP_ADDRESS range are handled in ktlb.S.
486 static inline int in_obp_range(unsigned long vaddr)
488 return (vaddr >= LOW_OBP_ADDRESS &&
489 vaddr < HI_OBP_ADDRESS);
492 static int cmp_ptrans(const void *a, const void *b)
494 const struct linux_prom_translation *x = a, *y = b;
496 if (x->virt > y->virt)
497 return 1;
498 if (x->virt < y->virt)
499 return -1;
500 return 0;
503 /* Read OBP translations property into 'prom_trans[]'. */
504 static void __init read_obp_translations(void)
506 int n, node, ents, first, last, i;
508 node = prom_finddevice("/virtual-memory");
509 n = prom_getproplen(node, "translations");
510 if (unlikely(n == 0 || n == -1)) {
511 prom_printf("prom_mappings: Couldn't get size.\n");
512 prom_halt();
514 if (unlikely(n > sizeof(prom_trans))) {
515 prom_printf("prom_mappings: Size %d is too big.\n", n);
516 prom_halt();
519 if ((n = prom_getproperty(node, "translations",
520 (char *)&prom_trans[0],
521 sizeof(prom_trans))) == -1) {
522 prom_printf("prom_mappings: Couldn't get property.\n");
523 prom_halt();
526 n = n / sizeof(struct linux_prom_translation);
528 ents = n;
530 sort(prom_trans, ents, sizeof(struct linux_prom_translation),
531 cmp_ptrans, NULL);
533 /* Now kick out all the non-OBP entries. */
534 for (i = 0; i < ents; i++) {
535 if (in_obp_range(prom_trans[i].virt))
536 break;
538 first = i;
539 for (; i < ents; i++) {
540 if (!in_obp_range(prom_trans[i].virt))
541 break;
543 last = i;
545 for (i = 0; i < (last - first); i++) {
546 struct linux_prom_translation *src = &prom_trans[i + first];
547 struct linux_prom_translation *dest = &prom_trans[i];
549 *dest = *src;
551 for (; i < ents; i++) {
552 struct linux_prom_translation *dest = &prom_trans[i];
553 dest->virt = dest->size = dest->data = 0x0UL;
556 prom_trans_ents = last - first;
558 if (tlb_type == spitfire) {
559 /* Clear diag TTE bits. */
560 for (i = 0; i < prom_trans_ents; i++)
561 prom_trans[i].data &= ~0x0003fe0000000000UL;
564 /* Force execute bit on. */
565 for (i = 0; i < prom_trans_ents; i++)
566 prom_trans[i].data |= (tlb_type == hypervisor ?
567 _PAGE_EXEC_4V : _PAGE_EXEC_4U);
570 static void __init hypervisor_tlb_lock(unsigned long vaddr,
571 unsigned long pte,
572 unsigned long mmu)
574 unsigned long ret = sun4v_mmu_map_perm_addr(vaddr, 0, pte, mmu);
576 if (ret != 0) {
577 prom_printf("hypervisor_tlb_lock[%lx:%x:%lx:%lx]: "
578 "errors with %lx\n", vaddr, 0, pte, mmu, ret);
579 prom_halt();
583 static unsigned long kern_large_tte(unsigned long paddr);
585 static void __init remap_kernel(void)
587 unsigned long phys_page, tte_vaddr, tte_data;
588 int i, tlb_ent = sparc64_highest_locked_tlbent();
590 tte_vaddr = (unsigned long) KERNBASE;
591 phys_page = (prom_boot_mapping_phys_low >> 22UL) << 22UL;
592 tte_data = kern_large_tte(phys_page);
594 kern_locked_tte_data = tte_data;
596 /* Now lock us into the TLBs via Hypervisor or OBP. */
597 if (tlb_type == hypervisor) {
598 for (i = 0; i < num_kernel_image_mappings; i++) {
599 hypervisor_tlb_lock(tte_vaddr, tte_data, HV_MMU_DMMU);
600 hypervisor_tlb_lock(tte_vaddr, tte_data, HV_MMU_IMMU);
601 tte_vaddr += 0x400000;
602 tte_data += 0x400000;
604 } else {
605 for (i = 0; i < num_kernel_image_mappings; i++) {
606 prom_dtlb_load(tlb_ent - i, tte_data, tte_vaddr);
607 prom_itlb_load(tlb_ent - i, tte_data, tte_vaddr);
608 tte_vaddr += 0x400000;
609 tte_data += 0x400000;
611 sparc64_highest_unlocked_tlb_ent = tlb_ent - i;
613 if (tlb_type == cheetah_plus) {
614 sparc64_kern_pri_context = (CTX_CHEETAH_PLUS_CTX0 |
615 CTX_CHEETAH_PLUS_NUC);
616 sparc64_kern_pri_nuc_bits = CTX_CHEETAH_PLUS_NUC;
617 sparc64_kern_sec_context = CTX_CHEETAH_PLUS_CTX0;
622 static void __init inherit_prom_mappings(void)
624 /* Now fixup OBP's idea about where we really are mapped. */
625 printk("Remapping the kernel... ");
626 remap_kernel();
627 printk("done.\n");
630 void prom_world(int enter)
632 if (!enter)
633 set_fs(get_fs());
635 __asm__ __volatile__("flushw");
638 void __flush_dcache_range(unsigned long start, unsigned long end)
640 unsigned long va;
642 if (tlb_type == spitfire) {
643 int n = 0;
645 for (va = start; va < end; va += 32) {
646 spitfire_put_dcache_tag(va & 0x3fe0, 0x0);
647 if (++n >= 512)
648 break;
650 } else if (tlb_type == cheetah || tlb_type == cheetah_plus) {
651 start = __pa(start);
652 end = __pa(end);
653 for (va = start; va < end; va += 32)
654 __asm__ __volatile__("stxa %%g0, [%0] %1\n\t"
655 "membar #Sync"
656 : /* no outputs */
657 : "r" (va),
658 "i" (ASI_DCACHE_INVALIDATE));
661 EXPORT_SYMBOL(__flush_dcache_range);
663 /* get_new_mmu_context() uses "cache + 1". */
664 DEFINE_SPINLOCK(ctx_alloc_lock);
665 unsigned long tlb_context_cache = CTX_FIRST_VERSION - 1;
666 #define MAX_CTX_NR (1UL << CTX_NR_BITS)
667 #define CTX_BMAP_SLOTS BITS_TO_LONGS(MAX_CTX_NR)
668 DECLARE_BITMAP(mmu_context_bmap, MAX_CTX_NR);
670 /* Caller does TLB context flushing on local CPU if necessary.
671 * The caller also ensures that CTX_VALID(mm->context) is false.
673 * We must be careful about boundary cases so that we never
674 * let the user have CTX 0 (nucleus) or we ever use a CTX
675 * version of zero (and thus NO_CONTEXT would not be caught
676 * by version mis-match tests in mmu_context.h).
678 * Always invoked with interrupts disabled.
680 void get_new_mmu_context(struct mm_struct *mm)
682 unsigned long ctx, new_ctx;
683 unsigned long orig_pgsz_bits;
684 int new_version;
686 spin_lock(&ctx_alloc_lock);
687 orig_pgsz_bits = (mm->context.sparc64_ctx_val & CTX_PGSZ_MASK);
688 ctx = (tlb_context_cache + 1) & CTX_NR_MASK;
689 new_ctx = find_next_zero_bit(mmu_context_bmap, 1 << CTX_NR_BITS, ctx);
690 new_version = 0;
691 if (new_ctx >= (1 << CTX_NR_BITS)) {
692 new_ctx = find_next_zero_bit(mmu_context_bmap, ctx, 1);
693 if (new_ctx >= ctx) {
694 int i;
695 new_ctx = (tlb_context_cache & CTX_VERSION_MASK) +
696 CTX_FIRST_VERSION;
697 if (new_ctx == 1)
698 new_ctx = CTX_FIRST_VERSION;
700 /* Don't call memset, for 16 entries that's just
701 * plain silly...
703 mmu_context_bmap[0] = 3;
704 mmu_context_bmap[1] = 0;
705 mmu_context_bmap[2] = 0;
706 mmu_context_bmap[3] = 0;
707 for (i = 4; i < CTX_BMAP_SLOTS; i += 4) {
708 mmu_context_bmap[i + 0] = 0;
709 mmu_context_bmap[i + 1] = 0;
710 mmu_context_bmap[i + 2] = 0;
711 mmu_context_bmap[i + 3] = 0;
713 new_version = 1;
714 goto out;
717 mmu_context_bmap[new_ctx>>6] |= (1UL << (new_ctx & 63));
718 new_ctx |= (tlb_context_cache & CTX_VERSION_MASK);
719 out:
720 tlb_context_cache = new_ctx;
721 mm->context.sparc64_ctx_val = new_ctx | orig_pgsz_bits;
722 spin_unlock(&ctx_alloc_lock);
724 if (unlikely(new_version))
725 smp_new_mmu_context_version();
728 static int numa_enabled = 1;
729 static int numa_debug;
731 static int __init early_numa(char *p)
733 if (!p)
734 return 0;
736 if (strstr(p, "off"))
737 numa_enabled = 0;
739 if (strstr(p, "debug"))
740 numa_debug = 1;
742 return 0;
744 early_param("numa", early_numa);
746 #define numadbg(f, a...) \
747 do { if (numa_debug) \
748 printk(KERN_INFO f, ## a); \
749 } while (0)
751 static void __init find_ramdisk(unsigned long phys_base)
753 #ifdef CONFIG_BLK_DEV_INITRD
754 if (sparc_ramdisk_image || sparc_ramdisk_image64) {
755 unsigned long ramdisk_image;
757 /* Older versions of the bootloader only supported a
758 * 32-bit physical address for the ramdisk image
759 * location, stored at sparc_ramdisk_image. Newer
760 * SILO versions set sparc_ramdisk_image to zero and
761 * provide a full 64-bit physical address at
762 * sparc_ramdisk_image64.
764 ramdisk_image = sparc_ramdisk_image;
765 if (!ramdisk_image)
766 ramdisk_image = sparc_ramdisk_image64;
768 /* Another bootloader quirk. The bootloader normalizes
769 * the physical address to KERNBASE, so we have to
770 * factor that back out and add in the lowest valid
771 * physical page address to get the true physical address.
773 ramdisk_image -= KERNBASE;
774 ramdisk_image += phys_base;
776 numadbg("Found ramdisk at physical address 0x%lx, size %u\n",
777 ramdisk_image, sparc_ramdisk_size);
779 initrd_start = ramdisk_image;
780 initrd_end = ramdisk_image + sparc_ramdisk_size;
782 memblock_reserve(initrd_start, sparc_ramdisk_size);
784 initrd_start += PAGE_OFFSET;
785 initrd_end += PAGE_OFFSET;
787 #endif
790 struct node_mem_mask {
791 unsigned long mask;
792 unsigned long val;
794 static struct node_mem_mask node_masks[MAX_NUMNODES];
795 static int num_node_masks;
797 int numa_cpu_lookup_table[NR_CPUS];
798 cpumask_t numa_cpumask_lookup_table[MAX_NUMNODES];
800 #ifdef CONFIG_NEED_MULTIPLE_NODES
802 struct mdesc_mblock {
803 u64 base;
804 u64 size;
805 u64 offset; /* RA-to-PA */
807 static struct mdesc_mblock *mblocks;
808 static int num_mblocks;
810 static unsigned long ra_to_pa(unsigned long addr)
812 int i;
814 for (i = 0; i < num_mblocks; i++) {
815 struct mdesc_mblock *m = &mblocks[i];
817 if (addr >= m->base &&
818 addr < (m->base + m->size)) {
819 addr += m->offset;
820 break;
823 return addr;
826 static int find_node(unsigned long addr)
828 int i;
830 addr = ra_to_pa(addr);
831 for (i = 0; i < num_node_masks; i++) {
832 struct node_mem_mask *p = &node_masks[i];
834 if ((addr & p->mask) == p->val)
835 return i;
837 return -1;
840 static u64 memblock_nid_range(u64 start, u64 end, int *nid)
842 *nid = find_node(start);
843 start += PAGE_SIZE;
844 while (start < end) {
845 int n = find_node(start);
847 if (n != *nid)
848 break;
849 start += PAGE_SIZE;
852 if (start > end)
853 start = end;
855 return start;
857 #endif
859 /* This must be invoked after performing all of the necessary
860 * memblock_set_node() calls for 'nid'. We need to be able to get
861 * correct data from get_pfn_range_for_nid().
863 static void __init allocate_node_data(int nid)
865 struct pglist_data *p;
866 unsigned long start_pfn, end_pfn;
867 #ifdef CONFIG_NEED_MULTIPLE_NODES
868 unsigned long paddr;
870 paddr = memblock_alloc_try_nid(sizeof(struct pglist_data), SMP_CACHE_BYTES, nid);
871 if (!paddr) {
872 prom_printf("Cannot allocate pglist_data for nid[%d]\n", nid);
873 prom_halt();
875 NODE_DATA(nid) = __va(paddr);
876 memset(NODE_DATA(nid), 0, sizeof(struct pglist_data));
878 NODE_DATA(nid)->node_id = nid;
879 #endif
881 p = NODE_DATA(nid);
883 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
884 p->node_start_pfn = start_pfn;
885 p->node_spanned_pages = end_pfn - start_pfn;
888 static void init_node_masks_nonnuma(void)
890 int i;
892 numadbg("Initializing tables for non-numa.\n");
894 node_masks[0].mask = node_masks[0].val = 0;
895 num_node_masks = 1;
897 for (i = 0; i < NR_CPUS; i++)
898 numa_cpu_lookup_table[i] = 0;
900 cpumask_setall(&numa_cpumask_lookup_table[0]);
903 #ifdef CONFIG_NEED_MULTIPLE_NODES
904 struct pglist_data *node_data[MAX_NUMNODES];
906 EXPORT_SYMBOL(numa_cpu_lookup_table);
907 EXPORT_SYMBOL(numa_cpumask_lookup_table);
908 EXPORT_SYMBOL(node_data);
910 struct mdesc_mlgroup {
911 u64 node;
912 u64 latency;
913 u64 match;
914 u64 mask;
916 static struct mdesc_mlgroup *mlgroups;
917 static int num_mlgroups;
919 static int scan_pio_for_cfg_handle(struct mdesc_handle *md, u64 pio,
920 u32 cfg_handle)
922 u64 arc;
924 mdesc_for_each_arc(arc, md, pio, MDESC_ARC_TYPE_FWD) {
925 u64 target = mdesc_arc_target(md, arc);
926 const u64 *val;
928 val = mdesc_get_property(md, target,
929 "cfg-handle", NULL);
930 if (val && *val == cfg_handle)
931 return 0;
933 return -ENODEV;
936 static int scan_arcs_for_cfg_handle(struct mdesc_handle *md, u64 grp,
937 u32 cfg_handle)
939 u64 arc, candidate, best_latency = ~(u64)0;
941 candidate = MDESC_NODE_NULL;
942 mdesc_for_each_arc(arc, md, grp, MDESC_ARC_TYPE_FWD) {
943 u64 target = mdesc_arc_target(md, arc);
944 const char *name = mdesc_node_name(md, target);
945 const u64 *val;
947 if (strcmp(name, "pio-latency-group"))
948 continue;
950 val = mdesc_get_property(md, target, "latency", NULL);
951 if (!val)
952 continue;
954 if (*val < best_latency) {
955 candidate = target;
956 best_latency = *val;
960 if (candidate == MDESC_NODE_NULL)
961 return -ENODEV;
963 return scan_pio_for_cfg_handle(md, candidate, cfg_handle);
966 int of_node_to_nid(struct device_node *dp)
968 const struct linux_prom64_registers *regs;
969 struct mdesc_handle *md;
970 u32 cfg_handle;
971 int count, nid;
972 u64 grp;
974 /* This is the right thing to do on currently supported
975 * SUN4U NUMA platforms as well, as the PCI controller does
976 * not sit behind any particular memory controller.
978 if (!mlgroups)
979 return -1;
981 regs = of_get_property(dp, "reg", NULL);
982 if (!regs)
983 return -1;
985 cfg_handle = (regs->phys_addr >> 32UL) & 0x0fffffff;
987 md = mdesc_grab();
989 count = 0;
990 nid = -1;
991 mdesc_for_each_node_by_name(md, grp, "group") {
992 if (!scan_arcs_for_cfg_handle(md, grp, cfg_handle)) {
993 nid = count;
994 break;
996 count++;
999 mdesc_release(md);
1001 return nid;
1004 static void __init add_node_ranges(void)
1006 struct memblock_region *reg;
1008 for_each_memblock(memory, reg) {
1009 unsigned long size = reg->size;
1010 unsigned long start, end;
1012 start = reg->base;
1013 end = start + size;
1014 while (start < end) {
1015 unsigned long this_end;
1016 int nid;
1018 this_end = memblock_nid_range(start, end, &nid);
1020 numadbg("Setting memblock NUMA node nid[%d] "
1021 "start[%lx] end[%lx]\n",
1022 nid, start, this_end);
1024 memblock_set_node(start, this_end - start,
1025 &memblock.memory, nid);
1026 start = this_end;
1031 static int __init grab_mlgroups(struct mdesc_handle *md)
1033 unsigned long paddr;
1034 int count = 0;
1035 u64 node;
1037 mdesc_for_each_node_by_name(md, node, "memory-latency-group")
1038 count++;
1039 if (!count)
1040 return -ENOENT;
1042 paddr = memblock_alloc(count * sizeof(struct mdesc_mlgroup),
1043 SMP_CACHE_BYTES);
1044 if (!paddr)
1045 return -ENOMEM;
1047 mlgroups = __va(paddr);
1048 num_mlgroups = count;
1050 count = 0;
1051 mdesc_for_each_node_by_name(md, node, "memory-latency-group") {
1052 struct mdesc_mlgroup *m = &mlgroups[count++];
1053 const u64 *val;
1055 m->node = node;
1057 val = mdesc_get_property(md, node, "latency", NULL);
1058 m->latency = *val;
1059 val = mdesc_get_property(md, node, "address-match", NULL);
1060 m->match = *val;
1061 val = mdesc_get_property(md, node, "address-mask", NULL);
1062 m->mask = *val;
1064 numadbg("MLGROUP[%d]: node[%llx] latency[%llx] "
1065 "match[%llx] mask[%llx]\n",
1066 count - 1, m->node, m->latency, m->match, m->mask);
1069 return 0;
1072 static int __init grab_mblocks(struct mdesc_handle *md)
1074 unsigned long paddr;
1075 int count = 0;
1076 u64 node;
1078 mdesc_for_each_node_by_name(md, node, "mblock")
1079 count++;
1080 if (!count)
1081 return -ENOENT;
1083 paddr = memblock_alloc(count * sizeof(struct mdesc_mblock),
1084 SMP_CACHE_BYTES);
1085 if (!paddr)
1086 return -ENOMEM;
1088 mblocks = __va(paddr);
1089 num_mblocks = count;
1091 count = 0;
1092 mdesc_for_each_node_by_name(md, node, "mblock") {
1093 struct mdesc_mblock *m = &mblocks[count++];
1094 const u64 *val;
1096 val = mdesc_get_property(md, node, "base", NULL);
1097 m->base = *val;
1098 val = mdesc_get_property(md, node, "size", NULL);
1099 m->size = *val;
1100 val = mdesc_get_property(md, node,
1101 "address-congruence-offset", NULL);
1103 /* The address-congruence-offset property is optional.
1104 * Explicity zero it be identifty this.
1106 if (val)
1107 m->offset = *val;
1108 else
1109 m->offset = 0UL;
1111 numadbg("MBLOCK[%d]: base[%llx] size[%llx] offset[%llx]\n",
1112 count - 1, m->base, m->size, m->offset);
1115 return 0;
1118 static void __init numa_parse_mdesc_group_cpus(struct mdesc_handle *md,
1119 u64 grp, cpumask_t *mask)
1121 u64 arc;
1123 cpumask_clear(mask);
1125 mdesc_for_each_arc(arc, md, grp, MDESC_ARC_TYPE_BACK) {
1126 u64 target = mdesc_arc_target(md, arc);
1127 const char *name = mdesc_node_name(md, target);
1128 const u64 *id;
1130 if (strcmp(name, "cpu"))
1131 continue;
1132 id = mdesc_get_property(md, target, "id", NULL);
1133 if (*id < nr_cpu_ids)
1134 cpumask_set_cpu(*id, mask);
1138 static struct mdesc_mlgroup * __init find_mlgroup(u64 node)
1140 int i;
1142 for (i = 0; i < num_mlgroups; i++) {
1143 struct mdesc_mlgroup *m = &mlgroups[i];
1144 if (m->node == node)
1145 return m;
1147 return NULL;
1150 static int __init numa_attach_mlgroup(struct mdesc_handle *md, u64 grp,
1151 int index)
1153 struct mdesc_mlgroup *candidate = NULL;
1154 u64 arc, best_latency = ~(u64)0;
1155 struct node_mem_mask *n;
1157 mdesc_for_each_arc(arc, md, grp, MDESC_ARC_TYPE_FWD) {
1158 u64 target = mdesc_arc_target(md, arc);
1159 struct mdesc_mlgroup *m = find_mlgroup(target);
1160 if (!m)
1161 continue;
1162 if (m->latency < best_latency) {
1163 candidate = m;
1164 best_latency = m->latency;
1167 if (!candidate)
1168 return -ENOENT;
1170 if (num_node_masks != index) {
1171 printk(KERN_ERR "Inconsistent NUMA state, "
1172 "index[%d] != num_node_masks[%d]\n",
1173 index, num_node_masks);
1174 return -EINVAL;
1177 n = &node_masks[num_node_masks++];
1179 n->mask = candidate->mask;
1180 n->val = candidate->match;
1182 numadbg("NUMA NODE[%d]: mask[%lx] val[%lx] (latency[%llx])\n",
1183 index, n->mask, n->val, candidate->latency);
1185 return 0;
1188 static int __init numa_parse_mdesc_group(struct mdesc_handle *md, u64 grp,
1189 int index)
1191 cpumask_t mask;
1192 int cpu;
1194 numa_parse_mdesc_group_cpus(md, grp, &mask);
1196 for_each_cpu(cpu, &mask)
1197 numa_cpu_lookup_table[cpu] = index;
1198 cpumask_copy(&numa_cpumask_lookup_table[index], &mask);
1200 if (numa_debug) {
1201 printk(KERN_INFO "NUMA GROUP[%d]: cpus [ ", index);
1202 for_each_cpu(cpu, &mask)
1203 printk("%d ", cpu);
1204 printk("]\n");
1207 return numa_attach_mlgroup(md, grp, index);
1210 static int __init numa_parse_mdesc(void)
1212 struct mdesc_handle *md = mdesc_grab();
1213 int i, err, count;
1214 u64 node;
1216 node = mdesc_node_by_name(md, MDESC_NODE_NULL, "latency-groups");
1217 if (node == MDESC_NODE_NULL) {
1218 mdesc_release(md);
1219 return -ENOENT;
1222 err = grab_mblocks(md);
1223 if (err < 0)
1224 goto out;
1226 err = grab_mlgroups(md);
1227 if (err < 0)
1228 goto out;
1230 count = 0;
1231 mdesc_for_each_node_by_name(md, node, "group") {
1232 err = numa_parse_mdesc_group(md, node, count);
1233 if (err < 0)
1234 break;
1235 count++;
1238 add_node_ranges();
1240 for (i = 0; i < num_node_masks; i++) {
1241 allocate_node_data(i);
1242 node_set_online(i);
1245 err = 0;
1246 out:
1247 mdesc_release(md);
1248 return err;
1251 static int __init numa_parse_jbus(void)
1253 unsigned long cpu, index;
1255 /* NUMA node id is encoded in bits 36 and higher, and there is
1256 * a 1-to-1 mapping from CPU ID to NUMA node ID.
1258 index = 0;
1259 for_each_present_cpu(cpu) {
1260 numa_cpu_lookup_table[cpu] = index;
1261 cpumask_copy(&numa_cpumask_lookup_table[index], cpumask_of(cpu));
1262 node_masks[index].mask = ~((1UL << 36UL) - 1UL);
1263 node_masks[index].val = cpu << 36UL;
1265 index++;
1267 num_node_masks = index;
1269 add_node_ranges();
1271 for (index = 0; index < num_node_masks; index++) {
1272 allocate_node_data(index);
1273 node_set_online(index);
1276 return 0;
1279 static int __init numa_parse_sun4u(void)
1281 if (tlb_type == cheetah || tlb_type == cheetah_plus) {
1282 unsigned long ver;
1284 __asm__ ("rdpr %%ver, %0" : "=r" (ver));
1285 if ((ver >> 32UL) == __JALAPENO_ID ||
1286 (ver >> 32UL) == __SERRANO_ID)
1287 return numa_parse_jbus();
1289 return -1;
1292 static int __init bootmem_init_numa(void)
1294 int err = -1;
1296 numadbg("bootmem_init_numa()\n");
1298 if (numa_enabled) {
1299 if (tlb_type == hypervisor)
1300 err = numa_parse_mdesc();
1301 else
1302 err = numa_parse_sun4u();
1304 return err;
1307 #else
1309 static int bootmem_init_numa(void)
1311 return -1;
1314 #endif
1316 static void __init bootmem_init_nonnuma(void)
1318 unsigned long top_of_ram = memblock_end_of_DRAM();
1319 unsigned long total_ram = memblock_phys_mem_size();
1321 numadbg("bootmem_init_nonnuma()\n");
1323 printk(KERN_INFO "Top of RAM: 0x%lx, Total RAM: 0x%lx\n",
1324 top_of_ram, total_ram);
1325 printk(KERN_INFO "Memory hole size: %ldMB\n",
1326 (top_of_ram - total_ram) >> 20);
1328 init_node_masks_nonnuma();
1329 memblock_set_node(0, (phys_addr_t)ULLONG_MAX, &memblock.memory, 0);
1330 allocate_node_data(0);
1331 node_set_online(0);
1334 static unsigned long __init bootmem_init(unsigned long phys_base)
1336 unsigned long end_pfn;
1338 end_pfn = memblock_end_of_DRAM() >> PAGE_SHIFT;
1339 max_pfn = max_low_pfn = end_pfn;
1340 min_low_pfn = (phys_base >> PAGE_SHIFT);
1342 if (bootmem_init_numa() < 0)
1343 bootmem_init_nonnuma();
1345 /* Dump memblock with node info. */
1346 memblock_dump_all();
1348 /* XXX cpu notifier XXX */
1350 sparse_memory_present_with_active_regions(MAX_NUMNODES);
1351 sparse_init();
1353 return end_pfn;
1356 static struct linux_prom64_registers pall[MAX_BANKS] __initdata;
1357 static int pall_ents __initdata;
1359 #ifdef CONFIG_DEBUG_PAGEALLOC
1360 static unsigned long __ref kernel_map_range(unsigned long pstart,
1361 unsigned long pend, pgprot_t prot)
1363 unsigned long vstart = PAGE_OFFSET + pstart;
1364 unsigned long vend = PAGE_OFFSET + pend;
1365 unsigned long alloc_bytes = 0UL;
1367 if ((vstart & ~PAGE_MASK) || (vend & ~PAGE_MASK)) {
1368 prom_printf("kernel_map: Unaligned physmem[%lx:%lx]\n",
1369 vstart, vend);
1370 prom_halt();
1373 while (vstart < vend) {
1374 unsigned long this_end, paddr = __pa(vstart);
1375 pgd_t *pgd = pgd_offset_k(vstart);
1376 pud_t *pud;
1377 pmd_t *pmd;
1378 pte_t *pte;
1380 pud = pud_offset(pgd, vstart);
1381 if (pud_none(*pud)) {
1382 pmd_t *new;
1384 new = __alloc_bootmem(PAGE_SIZE, PAGE_SIZE, PAGE_SIZE);
1385 alloc_bytes += PAGE_SIZE;
1386 pud_populate(&init_mm, pud, new);
1389 pmd = pmd_offset(pud, vstart);
1390 if (!pmd_present(*pmd)) {
1391 pte_t *new;
1393 new = __alloc_bootmem(PAGE_SIZE, PAGE_SIZE, PAGE_SIZE);
1394 alloc_bytes += PAGE_SIZE;
1395 pmd_populate_kernel(&init_mm, pmd, new);
1398 pte = pte_offset_kernel(pmd, vstart);
1399 this_end = (vstart + PMD_SIZE) & PMD_MASK;
1400 if (this_end > vend)
1401 this_end = vend;
1403 while (vstart < this_end) {
1404 pte_val(*pte) = (paddr | pgprot_val(prot));
1406 vstart += PAGE_SIZE;
1407 paddr += PAGE_SIZE;
1408 pte++;
1412 return alloc_bytes;
1415 extern unsigned int kvmap_linear_patch[1];
1416 #endif /* CONFIG_DEBUG_PAGEALLOC */
1418 static void __init kpte_set_val(unsigned long index, unsigned long val)
1420 unsigned long *ptr = kpte_linear_bitmap;
1422 val <<= ((index % (BITS_PER_LONG / 2)) * 2);
1423 ptr += (index / (BITS_PER_LONG / 2));
1425 *ptr |= val;
1428 static const unsigned long kpte_shift_min = 28; /* 256MB */
1429 static const unsigned long kpte_shift_max = 34; /* 16GB */
1430 static const unsigned long kpte_shift_incr = 3;
1432 static unsigned long kpte_mark_using_shift(unsigned long start, unsigned long end,
1433 unsigned long shift)
1435 unsigned long size = (1UL << shift);
1436 unsigned long mask = (size - 1UL);
1437 unsigned long remains = end - start;
1438 unsigned long val;
1440 if (remains < size || (start & mask))
1441 return start;
1443 /* VAL maps:
1445 * shift 28 --> kern_linear_pte_xor index 1
1446 * shift 31 --> kern_linear_pte_xor index 2
1447 * shift 34 --> kern_linear_pte_xor index 3
1449 val = ((shift - kpte_shift_min) / kpte_shift_incr) + 1;
1451 remains &= ~mask;
1452 if (shift != kpte_shift_max)
1453 remains = size;
1455 while (remains) {
1456 unsigned long index = start >> kpte_shift_min;
1458 kpte_set_val(index, val);
1460 start += 1UL << kpte_shift_min;
1461 remains -= 1UL << kpte_shift_min;
1464 return start;
1467 static void __init mark_kpte_bitmap(unsigned long start, unsigned long end)
1469 unsigned long smallest_size, smallest_mask;
1470 unsigned long s;
1472 smallest_size = (1UL << kpte_shift_min);
1473 smallest_mask = (smallest_size - 1UL);
1475 while (start < end) {
1476 unsigned long orig_start = start;
1478 for (s = kpte_shift_max; s >= kpte_shift_min; s -= kpte_shift_incr) {
1479 start = kpte_mark_using_shift(start, end, s);
1481 if (start != orig_start)
1482 break;
1485 if (start == orig_start)
1486 start = (start + smallest_size) & ~smallest_mask;
1490 static void __init init_kpte_bitmap(void)
1492 unsigned long i;
1494 for (i = 0; i < pall_ents; i++) {
1495 unsigned long phys_start, phys_end;
1497 phys_start = pall[i].phys_addr;
1498 phys_end = phys_start + pall[i].reg_size;
1500 mark_kpte_bitmap(phys_start, phys_end);
1504 static void __init kernel_physical_mapping_init(void)
1506 #ifdef CONFIG_DEBUG_PAGEALLOC
1507 unsigned long i, mem_alloced = 0UL;
1509 for (i = 0; i < pall_ents; i++) {
1510 unsigned long phys_start, phys_end;
1512 phys_start = pall[i].phys_addr;
1513 phys_end = phys_start + pall[i].reg_size;
1515 mem_alloced += kernel_map_range(phys_start, phys_end,
1516 PAGE_KERNEL);
1519 printk("Allocated %ld bytes for kernel page tables.\n",
1520 mem_alloced);
1522 kvmap_linear_patch[0] = 0x01000000; /* nop */
1523 flushi(&kvmap_linear_patch[0]);
1525 __flush_tlb_all();
1526 #endif
1529 #ifdef CONFIG_DEBUG_PAGEALLOC
1530 void kernel_map_pages(struct page *page, int numpages, int enable)
1532 unsigned long phys_start = page_to_pfn(page) << PAGE_SHIFT;
1533 unsigned long phys_end = phys_start + (numpages * PAGE_SIZE);
1535 kernel_map_range(phys_start, phys_end,
1536 (enable ? PAGE_KERNEL : __pgprot(0)));
1538 flush_tsb_kernel_range(PAGE_OFFSET + phys_start,
1539 PAGE_OFFSET + phys_end);
1541 /* we should perform an IPI and flush all tlbs,
1542 * but that can deadlock->flush only current cpu.
1544 __flush_tlb_kernel_range(PAGE_OFFSET + phys_start,
1545 PAGE_OFFSET + phys_end);
1547 #endif
1549 unsigned long __init find_ecache_flush_span(unsigned long size)
1551 int i;
1553 for (i = 0; i < pavail_ents; i++) {
1554 if (pavail[i].reg_size >= size)
1555 return pavail[i].phys_addr;
1558 return ~0UL;
1561 unsigned long PAGE_OFFSET;
1562 EXPORT_SYMBOL(PAGE_OFFSET);
1564 static void __init page_offset_shift_patch_one(unsigned int *insn, unsigned long phys_bits)
1566 unsigned long final_shift;
1567 unsigned int val = *insn;
1568 unsigned int cnt;
1570 /* We are patching in ilog2(max_supported_phys_address), and
1571 * we are doing so in a manner similar to a relocation addend.
1572 * That is, we are adding the shift value to whatever value
1573 * is in the shift instruction count field already.
1575 cnt = (val & 0x3f);
1576 val &= ~0x3f;
1578 /* If we are trying to shift >= 64 bits, clear the destination
1579 * register. This can happen when phys_bits ends up being equal
1580 * to MAX_PHYS_ADDRESS_BITS.
1582 final_shift = (cnt + (64 - phys_bits));
1583 if (final_shift >= 64) {
1584 unsigned int rd = (val >> 25) & 0x1f;
1586 val = 0x80100000 | (rd << 25);
1587 } else {
1588 val |= final_shift;
1590 *insn = val;
1592 __asm__ __volatile__("flush %0"
1593 : /* no outputs */
1594 : "r" (insn));
1597 static void __init page_offset_shift_patch(unsigned long phys_bits)
1599 extern unsigned int __page_offset_shift_patch;
1600 extern unsigned int __page_offset_shift_patch_end;
1601 unsigned int *p;
1603 p = &__page_offset_shift_patch;
1604 while (p < &__page_offset_shift_patch_end) {
1605 unsigned int *insn = (unsigned int *)(unsigned long)*p;
1607 page_offset_shift_patch_one(insn, phys_bits);
1609 p++;
1613 static void __init setup_page_offset(void)
1615 unsigned long max_phys_bits = 40;
1617 if (tlb_type == cheetah || tlb_type == cheetah_plus) {
1618 max_phys_bits = 42;
1619 } else if (tlb_type == hypervisor) {
1620 switch (sun4v_chip_type) {
1621 case SUN4V_CHIP_NIAGARA1:
1622 case SUN4V_CHIP_NIAGARA2:
1623 max_phys_bits = 39;
1624 break;
1625 case SUN4V_CHIP_NIAGARA3:
1626 max_phys_bits = 43;
1627 break;
1628 case SUN4V_CHIP_NIAGARA4:
1629 case SUN4V_CHIP_NIAGARA5:
1630 case SUN4V_CHIP_SPARC64X:
1631 default:
1632 max_phys_bits = 47;
1633 break;
1637 if (max_phys_bits > MAX_PHYS_ADDRESS_BITS) {
1638 prom_printf("MAX_PHYS_ADDRESS_BITS is too small, need %lu\n",
1639 max_phys_bits);
1640 prom_halt();
1643 PAGE_OFFSET = PAGE_OFFSET_BY_BITS(max_phys_bits);
1645 pr_info("PAGE_OFFSET is 0x%016lx (max_phys_bits == %lu)\n",
1646 PAGE_OFFSET, max_phys_bits);
1648 page_offset_shift_patch(max_phys_bits);
1651 static void __init tsb_phys_patch(void)
1653 struct tsb_ldquad_phys_patch_entry *pquad;
1654 struct tsb_phys_patch_entry *p;
1656 pquad = &__tsb_ldquad_phys_patch;
1657 while (pquad < &__tsb_ldquad_phys_patch_end) {
1658 unsigned long addr = pquad->addr;
1660 if (tlb_type == hypervisor)
1661 *(unsigned int *) addr = pquad->sun4v_insn;
1662 else
1663 *(unsigned int *) addr = pquad->sun4u_insn;
1664 wmb();
1665 __asm__ __volatile__("flush %0"
1666 : /* no outputs */
1667 : "r" (addr));
1669 pquad++;
1672 p = &__tsb_phys_patch;
1673 while (p < &__tsb_phys_patch_end) {
1674 unsigned long addr = p->addr;
1676 *(unsigned int *) addr = p->insn;
1677 wmb();
1678 __asm__ __volatile__("flush %0"
1679 : /* no outputs */
1680 : "r" (addr));
1682 p++;
1686 /* Don't mark as init, we give this to the Hypervisor. */
1687 #ifndef CONFIG_DEBUG_PAGEALLOC
1688 #define NUM_KTSB_DESCR 2
1689 #else
1690 #define NUM_KTSB_DESCR 1
1691 #endif
1692 static struct hv_tsb_descr ktsb_descr[NUM_KTSB_DESCR];
1693 extern struct tsb swapper_tsb[KERNEL_TSB_NENTRIES];
1695 static void patch_one_ktsb_phys(unsigned int *start, unsigned int *end, unsigned long pa)
1697 pa >>= KTSB_PHYS_SHIFT;
1699 while (start < end) {
1700 unsigned int *ia = (unsigned int *)(unsigned long)*start;
1702 ia[0] = (ia[0] & ~0x3fffff) | (pa >> 10);
1703 __asm__ __volatile__("flush %0" : : "r" (ia));
1705 ia[1] = (ia[1] & ~0x3ff) | (pa & 0x3ff);
1706 __asm__ __volatile__("flush %0" : : "r" (ia + 1));
1708 start++;
1712 static void ktsb_phys_patch(void)
1714 extern unsigned int __swapper_tsb_phys_patch;
1715 extern unsigned int __swapper_tsb_phys_patch_end;
1716 unsigned long ktsb_pa;
1718 ktsb_pa = kern_base + ((unsigned long)&swapper_tsb[0] - KERNBASE);
1719 patch_one_ktsb_phys(&__swapper_tsb_phys_patch,
1720 &__swapper_tsb_phys_patch_end, ktsb_pa);
1721 #ifndef CONFIG_DEBUG_PAGEALLOC
1723 extern unsigned int __swapper_4m_tsb_phys_patch;
1724 extern unsigned int __swapper_4m_tsb_phys_patch_end;
1725 ktsb_pa = (kern_base +
1726 ((unsigned long)&swapper_4m_tsb[0] - KERNBASE));
1727 patch_one_ktsb_phys(&__swapper_4m_tsb_phys_patch,
1728 &__swapper_4m_tsb_phys_patch_end, ktsb_pa);
1730 #endif
1733 static void __init sun4v_ktsb_init(void)
1735 unsigned long ktsb_pa;
1737 /* First KTSB for PAGE_SIZE mappings. */
1738 ktsb_pa = kern_base + ((unsigned long)&swapper_tsb[0] - KERNBASE);
1740 switch (PAGE_SIZE) {
1741 case 8 * 1024:
1742 default:
1743 ktsb_descr[0].pgsz_idx = HV_PGSZ_IDX_8K;
1744 ktsb_descr[0].pgsz_mask = HV_PGSZ_MASK_8K;
1745 break;
1747 case 64 * 1024:
1748 ktsb_descr[0].pgsz_idx = HV_PGSZ_IDX_64K;
1749 ktsb_descr[0].pgsz_mask = HV_PGSZ_MASK_64K;
1750 break;
1752 case 512 * 1024:
1753 ktsb_descr[0].pgsz_idx = HV_PGSZ_IDX_512K;
1754 ktsb_descr[0].pgsz_mask = HV_PGSZ_MASK_512K;
1755 break;
1757 case 4 * 1024 * 1024:
1758 ktsb_descr[0].pgsz_idx = HV_PGSZ_IDX_4MB;
1759 ktsb_descr[0].pgsz_mask = HV_PGSZ_MASK_4MB;
1760 break;
1763 ktsb_descr[0].assoc = 1;
1764 ktsb_descr[0].num_ttes = KERNEL_TSB_NENTRIES;
1765 ktsb_descr[0].ctx_idx = 0;
1766 ktsb_descr[0].tsb_base = ktsb_pa;
1767 ktsb_descr[0].resv = 0;
1769 #ifndef CONFIG_DEBUG_PAGEALLOC
1770 /* Second KTSB for 4MB/256MB/2GB/16GB mappings. */
1771 ktsb_pa = (kern_base +
1772 ((unsigned long)&swapper_4m_tsb[0] - KERNBASE));
1774 ktsb_descr[1].pgsz_idx = HV_PGSZ_IDX_4MB;
1775 ktsb_descr[1].pgsz_mask = ((HV_PGSZ_MASK_4MB |
1776 HV_PGSZ_MASK_256MB |
1777 HV_PGSZ_MASK_2GB |
1778 HV_PGSZ_MASK_16GB) &
1779 cpu_pgsz_mask);
1780 ktsb_descr[1].assoc = 1;
1781 ktsb_descr[1].num_ttes = KERNEL_TSB4M_NENTRIES;
1782 ktsb_descr[1].ctx_idx = 0;
1783 ktsb_descr[1].tsb_base = ktsb_pa;
1784 ktsb_descr[1].resv = 0;
1785 #endif
1788 void sun4v_ktsb_register(void)
1790 unsigned long pa, ret;
1792 pa = kern_base + ((unsigned long)&ktsb_descr[0] - KERNBASE);
1794 ret = sun4v_mmu_tsb_ctx0(NUM_KTSB_DESCR, pa);
1795 if (ret != 0) {
1796 prom_printf("hypervisor_mmu_tsb_ctx0[%lx]: "
1797 "errors with %lx\n", pa, ret);
1798 prom_halt();
1802 static void __init sun4u_linear_pte_xor_finalize(void)
1804 #ifndef CONFIG_DEBUG_PAGEALLOC
1805 /* This is where we would add Panther support for
1806 * 32MB and 256MB pages.
1808 #endif
1811 static void __init sun4v_linear_pte_xor_finalize(void)
1813 #ifndef CONFIG_DEBUG_PAGEALLOC
1814 if (cpu_pgsz_mask & HV_PGSZ_MASK_256MB) {
1815 kern_linear_pte_xor[1] = (_PAGE_VALID | _PAGE_SZ256MB_4V) ^
1816 PAGE_OFFSET;
1817 kern_linear_pte_xor[1] |= (_PAGE_CP_4V | _PAGE_CV_4V |
1818 _PAGE_P_4V | _PAGE_W_4V);
1819 } else {
1820 kern_linear_pte_xor[1] = kern_linear_pte_xor[0];
1823 if (cpu_pgsz_mask & HV_PGSZ_MASK_2GB) {
1824 kern_linear_pte_xor[2] = (_PAGE_VALID | _PAGE_SZ2GB_4V) ^
1825 PAGE_OFFSET;
1826 kern_linear_pte_xor[2] |= (_PAGE_CP_4V | _PAGE_CV_4V |
1827 _PAGE_P_4V | _PAGE_W_4V);
1828 } else {
1829 kern_linear_pte_xor[2] = kern_linear_pte_xor[1];
1832 if (cpu_pgsz_mask & HV_PGSZ_MASK_16GB) {
1833 kern_linear_pte_xor[3] = (_PAGE_VALID | _PAGE_SZ16GB_4V) ^
1834 PAGE_OFFSET;
1835 kern_linear_pte_xor[3] |= (_PAGE_CP_4V | _PAGE_CV_4V |
1836 _PAGE_P_4V | _PAGE_W_4V);
1837 } else {
1838 kern_linear_pte_xor[3] = kern_linear_pte_xor[2];
1840 #endif
1843 /* paging_init() sets up the page tables */
1845 static unsigned long last_valid_pfn;
1846 pgd_t swapper_pg_dir[PTRS_PER_PGD];
1848 static void sun4u_pgprot_init(void);
1849 static void sun4v_pgprot_init(void);
1851 void __init paging_init(void)
1853 unsigned long end_pfn, shift, phys_base;
1854 unsigned long real_end, i;
1855 int node;
1857 setup_page_offset();
1859 /* These build time checkes make sure that the dcache_dirty_cpu()
1860 * page->flags usage will work.
1862 * When a page gets marked as dcache-dirty, we store the
1863 * cpu number starting at bit 32 in the page->flags. Also,
1864 * functions like clear_dcache_dirty_cpu use the cpu mask
1865 * in 13-bit signed-immediate instruction fields.
1869 * Page flags must not reach into upper 32 bits that are used
1870 * for the cpu number
1872 BUILD_BUG_ON(NR_PAGEFLAGS > 32);
1875 * The bit fields placed in the high range must not reach below
1876 * the 32 bit boundary. Otherwise we cannot place the cpu field
1877 * at the 32 bit boundary.
1879 BUILD_BUG_ON(SECTIONS_WIDTH + NODES_WIDTH + ZONES_WIDTH +
1880 ilog2(roundup_pow_of_two(NR_CPUS)) > 32);
1882 BUILD_BUG_ON(NR_CPUS > 4096);
1884 kern_base = (prom_boot_mapping_phys_low >> 22UL) << 22UL;
1885 kern_size = (unsigned long)&_end - (unsigned long)KERNBASE;
1887 /* Invalidate both kernel TSBs. */
1888 memset(swapper_tsb, 0x40, sizeof(swapper_tsb));
1889 #ifndef CONFIG_DEBUG_PAGEALLOC
1890 memset(swapper_4m_tsb, 0x40, sizeof(swapper_4m_tsb));
1891 #endif
1893 if (tlb_type == hypervisor)
1894 sun4v_pgprot_init();
1895 else
1896 sun4u_pgprot_init();
1898 if (tlb_type == cheetah_plus ||
1899 tlb_type == hypervisor) {
1900 tsb_phys_patch();
1901 ktsb_phys_patch();
1904 if (tlb_type == hypervisor)
1905 sun4v_patch_tlb_handlers();
1907 /* Find available physical memory...
1909 * Read it twice in order to work around a bug in openfirmware.
1910 * The call to grab this table itself can cause openfirmware to
1911 * allocate memory, which in turn can take away some space from
1912 * the list of available memory. Reading it twice makes sure
1913 * we really do get the final value.
1915 read_obp_translations();
1916 read_obp_memory("reg", &pall[0], &pall_ents);
1917 read_obp_memory("available", &pavail[0], &pavail_ents);
1918 read_obp_memory("available", &pavail[0], &pavail_ents);
1920 phys_base = 0xffffffffffffffffUL;
1921 for (i = 0; i < pavail_ents; i++) {
1922 phys_base = min(phys_base, pavail[i].phys_addr);
1923 memblock_add(pavail[i].phys_addr, pavail[i].reg_size);
1926 memblock_reserve(kern_base, kern_size);
1928 find_ramdisk(phys_base);
1930 memblock_enforce_memory_limit(cmdline_memory_size);
1932 memblock_allow_resize();
1933 memblock_dump_all();
1935 set_bit(0, mmu_context_bmap);
1937 shift = kern_base + PAGE_OFFSET - ((unsigned long)KERNBASE);
1939 real_end = (unsigned long)_end;
1940 num_kernel_image_mappings = DIV_ROUND_UP(real_end - KERNBASE, 1 << 22);
1941 printk("Kernel: Using %d locked TLB entries for main kernel image.\n",
1942 num_kernel_image_mappings);
1944 /* Set kernel pgd to upper alias so physical page computations
1945 * work.
1947 init_mm.pgd += ((shift) / (sizeof(pgd_t)));
1949 memset(swapper_low_pmd_dir, 0, sizeof(swapper_low_pmd_dir));
1951 /* Now can init the kernel/bad page tables. */
1952 pud_set(pud_offset(&swapper_pg_dir[0], 0),
1953 swapper_low_pmd_dir + (shift / sizeof(pgd_t)));
1955 inherit_prom_mappings();
1957 init_kpte_bitmap();
1959 /* Ok, we can use our TLB miss and window trap handlers safely. */
1960 setup_tba();
1962 __flush_tlb_all();
1964 prom_build_devicetree();
1965 of_populate_present_mask();
1966 #ifndef CONFIG_SMP
1967 of_fill_in_cpu_data();
1968 #endif
1970 if (tlb_type == hypervisor) {
1971 sun4v_mdesc_init();
1972 mdesc_populate_present_mask(cpu_all_mask);
1973 #ifndef CONFIG_SMP
1974 mdesc_fill_in_cpu_data(cpu_all_mask);
1975 #endif
1976 mdesc_get_page_sizes(cpu_all_mask, &cpu_pgsz_mask);
1978 sun4v_linear_pte_xor_finalize();
1980 sun4v_ktsb_init();
1981 sun4v_ktsb_register();
1982 } else {
1983 unsigned long impl, ver;
1985 cpu_pgsz_mask = (HV_PGSZ_MASK_8K | HV_PGSZ_MASK_64K |
1986 HV_PGSZ_MASK_512K | HV_PGSZ_MASK_4MB);
1988 __asm__ __volatile__("rdpr %%ver, %0" : "=r" (ver));
1989 impl = ((ver >> 32) & 0xffff);
1990 if (impl == PANTHER_IMPL)
1991 cpu_pgsz_mask |= (HV_PGSZ_MASK_32MB |
1992 HV_PGSZ_MASK_256MB);
1994 sun4u_linear_pte_xor_finalize();
1997 /* Flush the TLBs and the 4M TSB so that the updated linear
1998 * pte XOR settings are realized for all mappings.
2000 __flush_tlb_all();
2001 #ifndef CONFIG_DEBUG_PAGEALLOC
2002 memset(swapper_4m_tsb, 0x40, sizeof(swapper_4m_tsb));
2003 #endif
2004 __flush_tlb_all();
2006 /* Setup bootmem... */
2007 last_valid_pfn = end_pfn = bootmem_init(phys_base);
2009 /* Once the OF device tree and MDESC have been setup, we know
2010 * the list of possible cpus. Therefore we can allocate the
2011 * IRQ stacks.
2013 for_each_possible_cpu(i) {
2014 node = cpu_to_node(i);
2016 softirq_stack[i] = __alloc_bootmem_node(NODE_DATA(node),
2017 THREAD_SIZE,
2018 THREAD_SIZE, 0);
2019 hardirq_stack[i] = __alloc_bootmem_node(NODE_DATA(node),
2020 THREAD_SIZE,
2021 THREAD_SIZE, 0);
2024 kernel_physical_mapping_init();
2027 unsigned long max_zone_pfns[MAX_NR_ZONES];
2029 memset(max_zone_pfns, 0, sizeof(max_zone_pfns));
2031 max_zone_pfns[ZONE_NORMAL] = end_pfn;
2033 free_area_init_nodes(max_zone_pfns);
2036 printk("Booting Linux...\n");
2039 int page_in_phys_avail(unsigned long paddr)
2041 int i;
2043 paddr &= PAGE_MASK;
2045 for (i = 0; i < pavail_ents; i++) {
2046 unsigned long start, end;
2048 start = pavail[i].phys_addr;
2049 end = start + pavail[i].reg_size;
2051 if (paddr >= start && paddr < end)
2052 return 1;
2054 if (paddr >= kern_base && paddr < (kern_base + kern_size))
2055 return 1;
2056 #ifdef CONFIG_BLK_DEV_INITRD
2057 if (paddr >= __pa(initrd_start) &&
2058 paddr < __pa(PAGE_ALIGN(initrd_end)))
2059 return 1;
2060 #endif
2062 return 0;
2065 static struct linux_prom64_registers pavail_rescan[MAX_BANKS] __initdata;
2066 static int pavail_rescan_ents __initdata;
2068 /* Certain OBP calls, such as fetching "available" properties, can
2069 * claim physical memory. So, along with initializing the valid
2070 * address bitmap, what we do here is refetch the physical available
2071 * memory list again, and make sure it provides at least as much
2072 * memory as 'pavail' does.
2074 static void __init setup_valid_addr_bitmap_from_pavail(unsigned long *bitmap)
2076 int i;
2078 read_obp_memory("available", &pavail_rescan[0], &pavail_rescan_ents);
2080 for (i = 0; i < pavail_ents; i++) {
2081 unsigned long old_start, old_end;
2083 old_start = pavail[i].phys_addr;
2084 old_end = old_start + pavail[i].reg_size;
2085 while (old_start < old_end) {
2086 int n;
2088 for (n = 0; n < pavail_rescan_ents; n++) {
2089 unsigned long new_start, new_end;
2091 new_start = pavail_rescan[n].phys_addr;
2092 new_end = new_start +
2093 pavail_rescan[n].reg_size;
2095 if (new_start <= old_start &&
2096 new_end >= (old_start + PAGE_SIZE)) {
2097 set_bit(old_start >> 22, bitmap);
2098 goto do_next_page;
2102 prom_printf("mem_init: Lost memory in pavail\n");
2103 prom_printf("mem_init: OLD start[%lx] size[%lx]\n",
2104 pavail[i].phys_addr,
2105 pavail[i].reg_size);
2106 prom_printf("mem_init: NEW start[%lx] size[%lx]\n",
2107 pavail_rescan[i].phys_addr,
2108 pavail_rescan[i].reg_size);
2109 prom_printf("mem_init: Cannot continue, aborting.\n");
2110 prom_halt();
2112 do_next_page:
2113 old_start += PAGE_SIZE;
2118 static void __init patch_tlb_miss_handler_bitmap(void)
2120 extern unsigned int valid_addr_bitmap_insn[];
2121 extern unsigned int valid_addr_bitmap_patch[];
2123 valid_addr_bitmap_insn[1] = valid_addr_bitmap_patch[1];
2124 mb();
2125 valid_addr_bitmap_insn[0] = valid_addr_bitmap_patch[0];
2126 flushi(&valid_addr_bitmap_insn[0]);
2129 static void __init register_page_bootmem_info(void)
2131 #ifdef CONFIG_NEED_MULTIPLE_NODES
2132 int i;
2134 for_each_online_node(i)
2135 if (NODE_DATA(i)->node_spanned_pages)
2136 register_page_bootmem_info_node(NODE_DATA(i));
2137 #endif
2139 void __init mem_init(void)
2141 unsigned long addr, last;
2143 addr = PAGE_OFFSET + kern_base;
2144 last = PAGE_ALIGN(kern_size) + addr;
2145 while (addr < last) {
2146 set_bit(__pa(addr) >> 22, sparc64_valid_addr_bitmap);
2147 addr += PAGE_SIZE;
2150 setup_valid_addr_bitmap_from_pavail(sparc64_valid_addr_bitmap);
2151 patch_tlb_miss_handler_bitmap();
2153 high_memory = __va(last_valid_pfn << PAGE_SHIFT);
2155 register_page_bootmem_info();
2156 free_all_bootmem();
2159 * Set up the zero page, mark it reserved, so that page count
2160 * is not manipulated when freeing the page from user ptes.
2162 mem_map_zero = alloc_pages(GFP_KERNEL|__GFP_ZERO, 0);
2163 if (mem_map_zero == NULL) {
2164 prom_printf("paging_init: Cannot alloc zero page.\n");
2165 prom_halt();
2167 mark_page_reserved(mem_map_zero);
2169 mem_init_print_info(NULL);
2171 if (tlb_type == cheetah || tlb_type == cheetah_plus)
2172 cheetah_ecache_flush_init();
2175 void free_initmem(void)
2177 unsigned long addr, initend;
2178 int do_free = 1;
2180 /* If the physical memory maps were trimmed by kernel command
2181 * line options, don't even try freeing this initmem stuff up.
2182 * The kernel image could have been in the trimmed out region
2183 * and if so the freeing below will free invalid page structs.
2185 if (cmdline_memory_size)
2186 do_free = 0;
2189 * The init section is aligned to 8k in vmlinux.lds. Page align for >8k pagesizes.
2191 addr = PAGE_ALIGN((unsigned long)(__init_begin));
2192 initend = (unsigned long)(__init_end) & PAGE_MASK;
2193 for (; addr < initend; addr += PAGE_SIZE) {
2194 unsigned long page;
2196 page = (addr +
2197 ((unsigned long) __va(kern_base)) -
2198 ((unsigned long) KERNBASE));
2199 memset((void *)addr, POISON_FREE_INITMEM, PAGE_SIZE);
2201 if (do_free)
2202 free_reserved_page(virt_to_page(page));
2206 #ifdef CONFIG_BLK_DEV_INITRD
2207 void free_initrd_mem(unsigned long start, unsigned long end)
2209 free_reserved_area((void *)start, (void *)end, POISON_FREE_INITMEM,
2210 "initrd");
2212 #endif
2214 #define _PAGE_CACHE_4U (_PAGE_CP_4U | _PAGE_CV_4U)
2215 #define _PAGE_CACHE_4V (_PAGE_CP_4V | _PAGE_CV_4V)
2216 #define __DIRTY_BITS_4U (_PAGE_MODIFIED_4U | _PAGE_WRITE_4U | _PAGE_W_4U)
2217 #define __DIRTY_BITS_4V (_PAGE_MODIFIED_4V | _PAGE_WRITE_4V | _PAGE_W_4V)
2218 #define __ACCESS_BITS_4U (_PAGE_ACCESSED_4U | _PAGE_READ_4U | _PAGE_R)
2219 #define __ACCESS_BITS_4V (_PAGE_ACCESSED_4V | _PAGE_READ_4V | _PAGE_R)
2221 pgprot_t PAGE_KERNEL __read_mostly;
2222 EXPORT_SYMBOL(PAGE_KERNEL);
2224 pgprot_t PAGE_KERNEL_LOCKED __read_mostly;
2225 pgprot_t PAGE_COPY __read_mostly;
2227 pgprot_t PAGE_SHARED __read_mostly;
2228 EXPORT_SYMBOL(PAGE_SHARED);
2230 unsigned long pg_iobits __read_mostly;
2232 unsigned long _PAGE_IE __read_mostly;
2233 EXPORT_SYMBOL(_PAGE_IE);
2235 unsigned long _PAGE_E __read_mostly;
2236 EXPORT_SYMBOL(_PAGE_E);
2238 unsigned long _PAGE_CACHE __read_mostly;
2239 EXPORT_SYMBOL(_PAGE_CACHE);
2241 #ifdef CONFIG_SPARSEMEM_VMEMMAP
2242 unsigned long vmemmap_table[VMEMMAP_SIZE];
2244 static long __meminitdata addr_start, addr_end;
2245 static int __meminitdata node_start;
2247 int __meminit vmemmap_populate(unsigned long vstart, unsigned long vend,
2248 int node)
2250 unsigned long phys_start = (vstart - VMEMMAP_BASE);
2251 unsigned long phys_end = (vend - VMEMMAP_BASE);
2252 unsigned long addr = phys_start & VMEMMAP_CHUNK_MASK;
2253 unsigned long end = VMEMMAP_ALIGN(phys_end);
2254 unsigned long pte_base;
2256 pte_base = (_PAGE_VALID | _PAGE_SZ4MB_4U |
2257 _PAGE_CP_4U | _PAGE_CV_4U |
2258 _PAGE_P_4U | _PAGE_W_4U);
2259 if (tlb_type == hypervisor)
2260 pte_base = (_PAGE_VALID | _PAGE_SZ4MB_4V |
2261 _PAGE_CP_4V | _PAGE_CV_4V |
2262 _PAGE_P_4V | _PAGE_W_4V);
2264 for (; addr < end; addr += VMEMMAP_CHUNK) {
2265 unsigned long *vmem_pp =
2266 vmemmap_table + (addr >> VMEMMAP_CHUNK_SHIFT);
2267 void *block;
2269 if (!(*vmem_pp & _PAGE_VALID)) {
2270 block = vmemmap_alloc_block(1UL << 22, node);
2271 if (!block)
2272 return -ENOMEM;
2274 *vmem_pp = pte_base | __pa(block);
2276 /* check to see if we have contiguous blocks */
2277 if (addr_end != addr || node_start != node) {
2278 if (addr_start)
2279 printk(KERN_DEBUG " [%lx-%lx] on node %d\n",
2280 addr_start, addr_end-1, node_start);
2281 addr_start = addr;
2282 node_start = node;
2284 addr_end = addr + VMEMMAP_CHUNK;
2287 return 0;
2290 void __meminit vmemmap_populate_print_last(void)
2292 if (addr_start) {
2293 printk(KERN_DEBUG " [%lx-%lx] on node %d\n",
2294 addr_start, addr_end-1, node_start);
2295 addr_start = 0;
2296 addr_end = 0;
2297 node_start = 0;
2301 void vmemmap_free(unsigned long start, unsigned long end)
2305 #endif /* CONFIG_SPARSEMEM_VMEMMAP */
2307 static void prot_init_common(unsigned long page_none,
2308 unsigned long page_shared,
2309 unsigned long page_copy,
2310 unsigned long page_readonly,
2311 unsigned long page_exec_bit)
2313 PAGE_COPY = __pgprot(page_copy);
2314 PAGE_SHARED = __pgprot(page_shared);
2316 protection_map[0x0] = __pgprot(page_none);
2317 protection_map[0x1] = __pgprot(page_readonly & ~page_exec_bit);
2318 protection_map[0x2] = __pgprot(page_copy & ~page_exec_bit);
2319 protection_map[0x3] = __pgprot(page_copy & ~page_exec_bit);
2320 protection_map[0x4] = __pgprot(page_readonly);
2321 protection_map[0x5] = __pgprot(page_readonly);
2322 protection_map[0x6] = __pgprot(page_copy);
2323 protection_map[0x7] = __pgprot(page_copy);
2324 protection_map[0x8] = __pgprot(page_none);
2325 protection_map[0x9] = __pgprot(page_readonly & ~page_exec_bit);
2326 protection_map[0xa] = __pgprot(page_shared & ~page_exec_bit);
2327 protection_map[0xb] = __pgprot(page_shared & ~page_exec_bit);
2328 protection_map[0xc] = __pgprot(page_readonly);
2329 protection_map[0xd] = __pgprot(page_readonly);
2330 protection_map[0xe] = __pgprot(page_shared);
2331 protection_map[0xf] = __pgprot(page_shared);
2334 static void __init sun4u_pgprot_init(void)
2336 unsigned long page_none, page_shared, page_copy, page_readonly;
2337 unsigned long page_exec_bit;
2338 int i;
2340 PAGE_KERNEL = __pgprot (_PAGE_PRESENT_4U | _PAGE_VALID |
2341 _PAGE_CACHE_4U | _PAGE_P_4U |
2342 __ACCESS_BITS_4U | __DIRTY_BITS_4U |
2343 _PAGE_EXEC_4U);
2344 PAGE_KERNEL_LOCKED = __pgprot (_PAGE_PRESENT_4U | _PAGE_VALID |
2345 _PAGE_CACHE_4U | _PAGE_P_4U |
2346 __ACCESS_BITS_4U | __DIRTY_BITS_4U |
2347 _PAGE_EXEC_4U | _PAGE_L_4U);
2349 _PAGE_IE = _PAGE_IE_4U;
2350 _PAGE_E = _PAGE_E_4U;
2351 _PAGE_CACHE = _PAGE_CACHE_4U;
2353 pg_iobits = (_PAGE_VALID | _PAGE_PRESENT_4U | __DIRTY_BITS_4U |
2354 __ACCESS_BITS_4U | _PAGE_E_4U);
2356 #ifdef CONFIG_DEBUG_PAGEALLOC
2357 kern_linear_pte_xor[0] = _PAGE_VALID ^ PAGE_OFFSET;
2358 #else
2359 kern_linear_pte_xor[0] = (_PAGE_VALID | _PAGE_SZ4MB_4U) ^
2360 PAGE_OFFSET;
2361 #endif
2362 kern_linear_pte_xor[0] |= (_PAGE_CP_4U | _PAGE_CV_4U |
2363 _PAGE_P_4U | _PAGE_W_4U);
2365 for (i = 1; i < 4; i++)
2366 kern_linear_pte_xor[i] = kern_linear_pte_xor[0];
2368 _PAGE_ALL_SZ_BITS = (_PAGE_SZ4MB_4U | _PAGE_SZ512K_4U |
2369 _PAGE_SZ64K_4U | _PAGE_SZ8K_4U |
2370 _PAGE_SZ32MB_4U | _PAGE_SZ256MB_4U);
2373 page_none = _PAGE_PRESENT_4U | _PAGE_ACCESSED_4U | _PAGE_CACHE_4U;
2374 page_shared = (_PAGE_VALID | _PAGE_PRESENT_4U | _PAGE_CACHE_4U |
2375 __ACCESS_BITS_4U | _PAGE_WRITE_4U | _PAGE_EXEC_4U);
2376 page_copy = (_PAGE_VALID | _PAGE_PRESENT_4U | _PAGE_CACHE_4U |
2377 __ACCESS_BITS_4U | _PAGE_EXEC_4U);
2378 page_readonly = (_PAGE_VALID | _PAGE_PRESENT_4U | _PAGE_CACHE_4U |
2379 __ACCESS_BITS_4U | _PAGE_EXEC_4U);
2381 page_exec_bit = _PAGE_EXEC_4U;
2383 prot_init_common(page_none, page_shared, page_copy, page_readonly,
2384 page_exec_bit);
2387 static void __init sun4v_pgprot_init(void)
2389 unsigned long page_none, page_shared, page_copy, page_readonly;
2390 unsigned long page_exec_bit;
2391 int i;
2393 PAGE_KERNEL = __pgprot (_PAGE_PRESENT_4V | _PAGE_VALID |
2394 _PAGE_CACHE_4V | _PAGE_P_4V |
2395 __ACCESS_BITS_4V | __DIRTY_BITS_4V |
2396 _PAGE_EXEC_4V);
2397 PAGE_KERNEL_LOCKED = PAGE_KERNEL;
2399 _PAGE_IE = _PAGE_IE_4V;
2400 _PAGE_E = _PAGE_E_4V;
2401 _PAGE_CACHE = _PAGE_CACHE_4V;
2403 #ifdef CONFIG_DEBUG_PAGEALLOC
2404 kern_linear_pte_xor[0] = _PAGE_VALID ^ PAGE_OFFSET;
2405 #else
2406 kern_linear_pte_xor[0] = (_PAGE_VALID | _PAGE_SZ4MB_4V) ^
2407 PAGE_OFFSET;
2408 #endif
2409 kern_linear_pte_xor[0] |= (_PAGE_CP_4V | _PAGE_CV_4V |
2410 _PAGE_P_4V | _PAGE_W_4V);
2412 for (i = 1; i < 4; i++)
2413 kern_linear_pte_xor[i] = kern_linear_pte_xor[0];
2415 pg_iobits = (_PAGE_VALID | _PAGE_PRESENT_4V | __DIRTY_BITS_4V |
2416 __ACCESS_BITS_4V | _PAGE_E_4V);
2418 _PAGE_ALL_SZ_BITS = (_PAGE_SZ16GB_4V | _PAGE_SZ2GB_4V |
2419 _PAGE_SZ256MB_4V | _PAGE_SZ32MB_4V |
2420 _PAGE_SZ4MB_4V | _PAGE_SZ512K_4V |
2421 _PAGE_SZ64K_4V | _PAGE_SZ8K_4V);
2423 page_none = _PAGE_PRESENT_4V | _PAGE_ACCESSED_4V | _PAGE_CACHE_4V;
2424 page_shared = (_PAGE_VALID | _PAGE_PRESENT_4V | _PAGE_CACHE_4V |
2425 __ACCESS_BITS_4V | _PAGE_WRITE_4V | _PAGE_EXEC_4V);
2426 page_copy = (_PAGE_VALID | _PAGE_PRESENT_4V | _PAGE_CACHE_4V |
2427 __ACCESS_BITS_4V | _PAGE_EXEC_4V);
2428 page_readonly = (_PAGE_VALID | _PAGE_PRESENT_4V | _PAGE_CACHE_4V |
2429 __ACCESS_BITS_4V | _PAGE_EXEC_4V);
2431 page_exec_bit = _PAGE_EXEC_4V;
2433 prot_init_common(page_none, page_shared, page_copy, page_readonly,
2434 page_exec_bit);
2437 unsigned long pte_sz_bits(unsigned long sz)
2439 if (tlb_type == hypervisor) {
2440 switch (sz) {
2441 case 8 * 1024:
2442 default:
2443 return _PAGE_SZ8K_4V;
2444 case 64 * 1024:
2445 return _PAGE_SZ64K_4V;
2446 case 512 * 1024:
2447 return _PAGE_SZ512K_4V;
2448 case 4 * 1024 * 1024:
2449 return _PAGE_SZ4MB_4V;
2451 } else {
2452 switch (sz) {
2453 case 8 * 1024:
2454 default:
2455 return _PAGE_SZ8K_4U;
2456 case 64 * 1024:
2457 return _PAGE_SZ64K_4U;
2458 case 512 * 1024:
2459 return _PAGE_SZ512K_4U;
2460 case 4 * 1024 * 1024:
2461 return _PAGE_SZ4MB_4U;
2466 pte_t mk_pte_io(unsigned long page, pgprot_t prot, int space, unsigned long page_size)
2468 pte_t pte;
2470 pte_val(pte) = page | pgprot_val(pgprot_noncached(prot));
2471 pte_val(pte) |= (((unsigned long)space) << 32);
2472 pte_val(pte) |= pte_sz_bits(page_size);
2474 return pte;
2477 static unsigned long kern_large_tte(unsigned long paddr)
2479 unsigned long val;
2481 val = (_PAGE_VALID | _PAGE_SZ4MB_4U |
2482 _PAGE_CP_4U | _PAGE_CV_4U | _PAGE_P_4U |
2483 _PAGE_EXEC_4U | _PAGE_L_4U | _PAGE_W_4U);
2484 if (tlb_type == hypervisor)
2485 val = (_PAGE_VALID | _PAGE_SZ4MB_4V |
2486 _PAGE_CP_4V | _PAGE_CV_4V | _PAGE_P_4V |
2487 _PAGE_EXEC_4V | _PAGE_W_4V);
2489 return val | paddr;
2492 /* If not locked, zap it. */
2493 void __flush_tlb_all(void)
2495 unsigned long pstate;
2496 int i;
2498 __asm__ __volatile__("flushw\n\t"
2499 "rdpr %%pstate, %0\n\t"
2500 "wrpr %0, %1, %%pstate"
2501 : "=r" (pstate)
2502 : "i" (PSTATE_IE));
2503 if (tlb_type == hypervisor) {
2504 sun4v_mmu_demap_all();
2505 } else if (tlb_type == spitfire) {
2506 for (i = 0; i < 64; i++) {
2507 /* Spitfire Errata #32 workaround */
2508 /* NOTE: Always runs on spitfire, so no
2509 * cheetah+ page size encodings.
2511 __asm__ __volatile__("stxa %0, [%1] %2\n\t"
2512 "flush %%g6"
2513 : /* No outputs */
2514 : "r" (0),
2515 "r" (PRIMARY_CONTEXT), "i" (ASI_DMMU));
2517 if (!(spitfire_get_dtlb_data(i) & _PAGE_L_4U)) {
2518 __asm__ __volatile__("stxa %%g0, [%0] %1\n\t"
2519 "membar #Sync"
2520 : /* no outputs */
2521 : "r" (TLB_TAG_ACCESS), "i" (ASI_DMMU));
2522 spitfire_put_dtlb_data(i, 0x0UL);
2525 /* Spitfire Errata #32 workaround */
2526 /* NOTE: Always runs on spitfire, so no
2527 * cheetah+ page size encodings.
2529 __asm__ __volatile__("stxa %0, [%1] %2\n\t"
2530 "flush %%g6"
2531 : /* No outputs */
2532 : "r" (0),
2533 "r" (PRIMARY_CONTEXT), "i" (ASI_DMMU));
2535 if (!(spitfire_get_itlb_data(i) & _PAGE_L_4U)) {
2536 __asm__ __volatile__("stxa %%g0, [%0] %1\n\t"
2537 "membar #Sync"
2538 : /* no outputs */
2539 : "r" (TLB_TAG_ACCESS), "i" (ASI_IMMU));
2540 spitfire_put_itlb_data(i, 0x0UL);
2543 } else if (tlb_type == cheetah || tlb_type == cheetah_plus) {
2544 cheetah_flush_dtlb_all();
2545 cheetah_flush_itlb_all();
2547 __asm__ __volatile__("wrpr %0, 0, %%pstate"
2548 : : "r" (pstate));
2551 pte_t *pte_alloc_one_kernel(struct mm_struct *mm,
2552 unsigned long address)
2554 struct page *page = alloc_page(GFP_KERNEL | __GFP_NOTRACK |
2555 __GFP_REPEAT | __GFP_ZERO);
2556 pte_t *pte = NULL;
2558 if (page)
2559 pte = (pte_t *) page_address(page);
2561 return pte;
2564 pgtable_t pte_alloc_one(struct mm_struct *mm,
2565 unsigned long address)
2567 struct page *page = alloc_page(GFP_KERNEL | __GFP_NOTRACK |
2568 __GFP_REPEAT | __GFP_ZERO);
2569 if (!page)
2570 return NULL;
2571 if (!pgtable_page_ctor(page)) {
2572 free_hot_cold_page(page, 0);
2573 return NULL;
2575 return (pte_t *) page_address(page);
2578 void pte_free_kernel(struct mm_struct *mm, pte_t *pte)
2580 free_page((unsigned long)pte);
2583 static void __pte_free(pgtable_t pte)
2585 struct page *page = virt_to_page(pte);
2587 pgtable_page_dtor(page);
2588 __free_page(page);
2591 void pte_free(struct mm_struct *mm, pgtable_t pte)
2593 __pte_free(pte);
2596 void pgtable_free(void *table, bool is_page)
2598 if (is_page)
2599 __pte_free(table);
2600 else
2601 kmem_cache_free(pgtable_cache, table);
2604 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
2605 void update_mmu_cache_pmd(struct vm_area_struct *vma, unsigned long addr,
2606 pmd_t *pmd)
2608 unsigned long pte, flags;
2609 struct mm_struct *mm;
2610 pmd_t entry = *pmd;
2612 if (!pmd_large(entry) || !pmd_young(entry))
2613 return;
2615 pte = pmd_val(entry);
2617 /* We are fabricating 8MB pages using 4MB real hw pages. */
2618 pte |= (addr & (1UL << REAL_HPAGE_SHIFT));
2620 mm = vma->vm_mm;
2622 spin_lock_irqsave(&mm->context.lock, flags);
2624 if (mm->context.tsb_block[MM_TSB_HUGE].tsb != NULL)
2625 __update_mmu_tsb_insert(mm, MM_TSB_HUGE, REAL_HPAGE_SHIFT,
2626 addr, pte);
2628 spin_unlock_irqrestore(&mm->context.lock, flags);
2630 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
2632 #if defined(CONFIG_HUGETLB_PAGE) || defined(CONFIG_TRANSPARENT_HUGEPAGE)
2633 static void context_reload(void *__data)
2635 struct mm_struct *mm = __data;
2637 if (mm == current->mm)
2638 load_secondary_context(mm);
2641 void hugetlb_setup(struct pt_regs *regs)
2643 struct mm_struct *mm = current->mm;
2644 struct tsb_config *tp;
2646 if (in_atomic() || !mm) {
2647 const struct exception_table_entry *entry;
2649 entry = search_exception_tables(regs->tpc);
2650 if (entry) {
2651 regs->tpc = entry->fixup;
2652 regs->tnpc = regs->tpc + 4;
2653 return;
2655 pr_alert("Unexpected HugeTLB setup in atomic context.\n");
2656 die_if_kernel("HugeTSB in atomic", regs);
2659 tp = &mm->context.tsb_block[MM_TSB_HUGE];
2660 if (likely(tp->tsb == NULL))
2661 tsb_grow(mm, MM_TSB_HUGE, 0);
2663 tsb_context_switch(mm);
2664 smp_tsb_sync(mm);
2666 /* On UltraSPARC-III+ and later, configure the second half of
2667 * the Data-TLB for huge pages.
2669 if (tlb_type == cheetah_plus) {
2670 unsigned long ctx;
2672 spin_lock(&ctx_alloc_lock);
2673 ctx = mm->context.sparc64_ctx_val;
2674 ctx &= ~CTX_PGSZ_MASK;
2675 ctx |= CTX_PGSZ_BASE << CTX_PGSZ0_SHIFT;
2676 ctx |= CTX_PGSZ_HUGE << CTX_PGSZ1_SHIFT;
2678 if (ctx != mm->context.sparc64_ctx_val) {
2679 /* When changing the page size fields, we
2680 * must perform a context flush so that no
2681 * stale entries match. This flush must
2682 * occur with the original context register
2683 * settings.
2685 do_flush_tlb_mm(mm);
2687 /* Reload the context register of all processors
2688 * also executing in this address space.
2690 mm->context.sparc64_ctx_val = ctx;
2691 on_each_cpu(context_reload, mm, 0);
2693 spin_unlock(&ctx_alloc_lock);
2696 #endif