Merge branch 'master' into export-slabh
[kernel.git] / arch / sparc / mm / init_64.c
1 /*
2  *  arch/sparc64/mm/init.c
3  *
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  */
7  
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/lmb.h>
27 #include <linux/mmzone.h>
28 #include <linux/gfp.h>
29
30 #include <asm/head.h>
31 #include <asm/system.h>
32 #include <asm/page.h>
33 #include <asm/pgalloc.h>
34 #include <asm/pgtable.h>
35 #include <asm/oplib.h>
36 #include <asm/iommu.h>
37 #include <asm/io.h>
38 #include <asm/uaccess.h>
39 #include <asm/mmu_context.h>
40 #include <asm/tlbflush.h>
41 #include <asm/dma.h>
42 #include <asm/starfire.h>
43 #include <asm/tlb.h>
44 #include <asm/spitfire.h>
45 #include <asm/sections.h>
46 #include <asm/tsb.h>
47 #include <asm/hypervisor.h>
48 #include <asm/prom.h>
49 #include <asm/mdesc.h>
50 #include <asm/cpudata.h>
51 #include <asm/irq.h>
52
53 #include "init_64.h"
54
55 unsigned long kern_linear_pte_xor[2] __read_mostly;
56
57 /* A bitmap, one bit for every 256MB of physical memory.  If the bit
58  * is clear, we should use a 4MB page (via kern_linear_pte_xor[0]) else
59  * if set we should use a 256MB page (via kern_linear_pte_xor[1]).
60  */
61 unsigned long kpte_linear_bitmap[KPTE_BITMAP_BYTES / sizeof(unsigned long)];
62
63 #ifndef CONFIG_DEBUG_PAGEALLOC
64 /* A special kernel TSB for 4MB and 256MB linear mappings.
65  * Space is allocated for this right after the trap table
66  * in arch/sparc64/kernel/head.S
67  */
68 extern struct tsb swapper_4m_tsb[KERNEL_TSB4M_NENTRIES];
69 #endif
70
71 #define MAX_BANKS       32
72
73 static struct linux_prom64_registers pavail[MAX_BANKS] __devinitdata;
74 static int pavail_ents __devinitdata;
75
76 static int cmp_p64(const void *a, const void *b)
77 {
78         const struct linux_prom64_registers *x = a, *y = b;
79
80         if (x->phys_addr > y->phys_addr)
81                 return 1;
82         if (x->phys_addr < y->phys_addr)
83                 return -1;
84         return 0;
85 }
86
87 static void __init read_obp_memory(const char *property,
88                                    struct linux_prom64_registers *regs,
89                                    int *num_ents)
90 {
91         int node = prom_finddevice("/memory");
92         int prop_size = prom_getproplen(node, property);
93         int ents, ret, i;
94
95         ents = prop_size / sizeof(struct linux_prom64_registers);
96         if (ents > MAX_BANKS) {
97                 prom_printf("The machine has more %s property entries than "
98                             "this kernel can support (%d).\n",
99                             property, MAX_BANKS);
100                 prom_halt();
101         }
102
103         ret = prom_getproperty(node, property, (char *) regs, prop_size);
104         if (ret == -1) {
105                 prom_printf("Couldn't get %s property from /memory.\n");
106                 prom_halt();
107         }
108
109         /* Sanitize what we got from the firmware, by page aligning
110          * everything.
111          */
112         for (i = 0; i < ents; i++) {
113                 unsigned long base, size;
114
115                 base = regs[i].phys_addr;
116                 size = regs[i].reg_size;
117
118                 size &= PAGE_MASK;
119                 if (base & ~PAGE_MASK) {
120                         unsigned long new_base = PAGE_ALIGN(base);
121
122                         size -= new_base - base;
123                         if ((long) size < 0L)
124                                 size = 0UL;
125                         base = new_base;
126                 }
127                 if (size == 0UL) {
128                         /* If it is empty, simply get rid of it.
129                          * This simplifies the logic of the other
130                          * functions that process these arrays.
131                          */
132                         memmove(&regs[i], &regs[i + 1],
133                                 (ents - i - 1) * sizeof(regs[0]));
134                         i--;
135                         ents--;
136                         continue;
137                 }
138                 regs[i].phys_addr = base;
139                 regs[i].reg_size = size;
140         }
141
142         *num_ents = ents;
143
144         sort(regs, ents, sizeof(struct linux_prom64_registers),
145              cmp_p64, NULL);
146 }
147
148 unsigned long sparc64_valid_addr_bitmap[VALID_ADDR_BITMAP_BYTES /
149                                         sizeof(unsigned long)];
150 EXPORT_SYMBOL(sparc64_valid_addr_bitmap);
151
152 /* Kernel physical address base and size in bytes.  */
153 unsigned long kern_base __read_mostly;
154 unsigned long kern_size __read_mostly;
155
156 /* Initial ramdisk setup */
157 extern unsigned long sparc_ramdisk_image64;
158 extern unsigned int sparc_ramdisk_image;
159 extern unsigned int sparc_ramdisk_size;
160
161 struct page *mem_map_zero __read_mostly;
162 EXPORT_SYMBOL(mem_map_zero);
163
164 unsigned int sparc64_highest_unlocked_tlb_ent __read_mostly;
165
166 unsigned long sparc64_kern_pri_context __read_mostly;
167 unsigned long sparc64_kern_pri_nuc_bits __read_mostly;
168 unsigned long sparc64_kern_sec_context __read_mostly;
169
170 int num_kernel_image_mappings;
171
172 #ifdef CONFIG_DEBUG_DCFLUSH
173 atomic_t dcpage_flushes = ATOMIC_INIT(0);
174 #ifdef CONFIG_SMP
175 atomic_t dcpage_flushes_xcall = ATOMIC_INIT(0);
176 #endif
177 #endif
178
179 inline void flush_dcache_page_impl(struct page *page)
180 {
181         BUG_ON(tlb_type == hypervisor);
182 #ifdef CONFIG_DEBUG_DCFLUSH
183         atomic_inc(&dcpage_flushes);
184 #endif
185
186 #ifdef DCACHE_ALIASING_POSSIBLE
187         __flush_dcache_page(page_address(page),
188                             ((tlb_type == spitfire) &&
189                              page_mapping(page) != NULL));
190 #else
191         if (page_mapping(page) != NULL &&
192             tlb_type == spitfire)
193                 __flush_icache_page(__pa(page_address(page)));
194 #endif
195 }
196
197 #define PG_dcache_dirty         PG_arch_1
198 #define PG_dcache_cpu_shift     32UL
199 #define PG_dcache_cpu_mask      \
200         ((1UL<<ilog2(roundup_pow_of_two(NR_CPUS)))-1UL)
201
202 #define dcache_dirty_cpu(page) \
203         (((page)->flags >> PG_dcache_cpu_shift) & PG_dcache_cpu_mask)
204
205 static inline void set_dcache_dirty(struct page *page, int this_cpu)
206 {
207         unsigned long mask = this_cpu;
208         unsigned long non_cpu_bits;
209
210         non_cpu_bits = ~(PG_dcache_cpu_mask << PG_dcache_cpu_shift);
211         mask = (mask << PG_dcache_cpu_shift) | (1UL << PG_dcache_dirty);
212
213         __asm__ __volatile__("1:\n\t"
214                              "ldx       [%2], %%g7\n\t"
215                              "and       %%g7, %1, %%g1\n\t"
216                              "or        %%g1, %0, %%g1\n\t"
217                              "casx      [%2], %%g7, %%g1\n\t"
218                              "cmp       %%g7, %%g1\n\t"
219                              "bne,pn    %%xcc, 1b\n\t"
220                              " nop"
221                              : /* no outputs */
222                              : "r" (mask), "r" (non_cpu_bits), "r" (&page->flags)
223                              : "g1", "g7");
224 }
225
226 static inline void clear_dcache_dirty_cpu(struct page *page, unsigned long cpu)
227 {
228         unsigned long mask = (1UL << PG_dcache_dirty);
229
230         __asm__ __volatile__("! test_and_clear_dcache_dirty\n"
231                              "1:\n\t"
232                              "ldx       [%2], %%g7\n\t"
233                              "srlx      %%g7, %4, %%g1\n\t"
234                              "and       %%g1, %3, %%g1\n\t"
235                              "cmp       %%g1, %0\n\t"
236                              "bne,pn    %%icc, 2f\n\t"
237                              " andn     %%g7, %1, %%g1\n\t"
238                              "casx      [%2], %%g7, %%g1\n\t"
239                              "cmp       %%g7, %%g1\n\t"
240                              "bne,pn    %%xcc, 1b\n\t"
241                              " nop\n"
242                              "2:"
243                              : /* no outputs */
244                              : "r" (cpu), "r" (mask), "r" (&page->flags),
245                                "i" (PG_dcache_cpu_mask),
246                                "i" (PG_dcache_cpu_shift)
247                              : "g1", "g7");
248 }
249
250 static inline void tsb_insert(struct tsb *ent, unsigned long tag, unsigned long pte)
251 {
252         unsigned long tsb_addr = (unsigned long) ent;
253
254         if (tlb_type == cheetah_plus || tlb_type == hypervisor)
255                 tsb_addr = __pa(tsb_addr);
256
257         __tsb_insert(tsb_addr, tag, pte);
258 }
259
260 unsigned long _PAGE_ALL_SZ_BITS __read_mostly;
261 unsigned long _PAGE_SZBITS __read_mostly;
262
263 static void flush_dcache(unsigned long pfn)
264 {
265         struct page *page;
266
267         page = pfn_to_page(pfn);
268         if (page) {
269                 unsigned long pg_flags;
270
271                 pg_flags = page->flags;
272                 if (pg_flags & (1UL << PG_dcache_dirty)) {
273                         int cpu = ((pg_flags >> PG_dcache_cpu_shift) &
274                                    PG_dcache_cpu_mask);
275                         int this_cpu = get_cpu();
276
277                         /* This is just to optimize away some function calls
278                          * in the SMP case.
279                          */
280                         if (cpu == this_cpu)
281                                 flush_dcache_page_impl(page);
282                         else
283                                 smp_flush_dcache_page_impl(page, cpu);
284
285                         clear_dcache_dirty_cpu(page, cpu);
286
287                         put_cpu();
288                 }
289         }
290 }
291
292 void update_mmu_cache(struct vm_area_struct *vma, unsigned long address, pte_t *ptep)
293 {
294         struct mm_struct *mm;
295         struct tsb *tsb;
296         unsigned long tag, flags;
297         unsigned long tsb_index, tsb_hash_shift;
298         pte_t pte = *ptep;
299
300         if (tlb_type != hypervisor) {
301                 unsigned long pfn = pte_pfn(pte);
302
303                 if (pfn_valid(pfn))
304                         flush_dcache(pfn);
305         }
306
307         mm = vma->vm_mm;
308
309         tsb_index = MM_TSB_BASE;
310         tsb_hash_shift = PAGE_SHIFT;
311
312         spin_lock_irqsave(&mm->context.lock, flags);
313
314 #ifdef CONFIG_HUGETLB_PAGE
315         if (mm->context.tsb_block[MM_TSB_HUGE].tsb != NULL) {
316                 if ((tlb_type == hypervisor &&
317                      (pte_val(pte) & _PAGE_SZALL_4V) == _PAGE_SZHUGE_4V) ||
318                     (tlb_type != hypervisor &&
319                      (pte_val(pte) & _PAGE_SZALL_4U) == _PAGE_SZHUGE_4U)) {
320                         tsb_index = MM_TSB_HUGE;
321                         tsb_hash_shift = HPAGE_SHIFT;
322                 }
323         }
324 #endif
325
326         tsb = mm->context.tsb_block[tsb_index].tsb;
327         tsb += ((address >> tsb_hash_shift) &
328                 (mm->context.tsb_block[tsb_index].tsb_nentries - 1UL));
329         tag = (address >> 22UL);
330         tsb_insert(tsb, tag, pte_val(pte));
331
332         spin_unlock_irqrestore(&mm->context.lock, flags);
333 }
334
335 void flush_dcache_page(struct page *page)
336 {
337         struct address_space *mapping;
338         int this_cpu;
339
340         if (tlb_type == hypervisor)
341                 return;
342
343         /* Do not bother with the expensive D-cache flush if it
344          * is merely the zero page.  The 'bigcore' testcase in GDB
345          * causes this case to run millions of times.
346          */
347         if (page == ZERO_PAGE(0))
348                 return;
349
350         this_cpu = get_cpu();
351
352         mapping = page_mapping(page);
353         if (mapping && !mapping_mapped(mapping)) {
354                 int dirty = test_bit(PG_dcache_dirty, &page->flags);
355                 if (dirty) {
356                         int dirty_cpu = dcache_dirty_cpu(page);
357
358                         if (dirty_cpu == this_cpu)
359                                 goto out;
360                         smp_flush_dcache_page_impl(page, dirty_cpu);
361                 }
362                 set_dcache_dirty(page, this_cpu);
363         } else {
364                 /* We could delay the flush for the !page_mapping
365                  * case too.  But that case is for exec env/arg
366                  * pages and those are %99 certainly going to get
367                  * faulted into the tlb (and thus flushed) anyways.
368                  */
369                 flush_dcache_page_impl(page);
370         }
371
372 out:
373         put_cpu();
374 }
375 EXPORT_SYMBOL(flush_dcache_page);
376
377 void __kprobes flush_icache_range(unsigned long start, unsigned long end)
378 {
379         /* Cheetah and Hypervisor platform cpus have coherent I-cache. */
380         if (tlb_type == spitfire) {
381                 unsigned long kaddr;
382
383                 /* This code only runs on Spitfire cpus so this is
384                  * why we can assume _PAGE_PADDR_4U.
385                  */
386                 for (kaddr = start; kaddr < end; kaddr += PAGE_SIZE) {
387                         unsigned long paddr, mask = _PAGE_PADDR_4U;
388
389                         if (kaddr >= PAGE_OFFSET)
390                                 paddr = kaddr & mask;
391                         else {
392                                 pgd_t *pgdp = pgd_offset_k(kaddr);
393                                 pud_t *pudp = pud_offset(pgdp, kaddr);
394                                 pmd_t *pmdp = pmd_offset(pudp, kaddr);
395                                 pte_t *ptep = pte_offset_kernel(pmdp, kaddr);
396
397                                 paddr = pte_val(*ptep) & mask;
398                         }
399                         __flush_icache_page(paddr);
400                 }
401         }
402 }
403 EXPORT_SYMBOL(flush_icache_range);
404
405 void mmu_info(struct seq_file *m)
406 {
407         if (tlb_type == cheetah)
408                 seq_printf(m, "MMU Type\t: Cheetah\n");
409         else if (tlb_type == cheetah_plus)
410                 seq_printf(m, "MMU Type\t: Cheetah+\n");
411         else if (tlb_type == spitfire)
412                 seq_printf(m, "MMU Type\t: Spitfire\n");
413         else if (tlb_type == hypervisor)
414                 seq_printf(m, "MMU Type\t: Hypervisor (sun4v)\n");
415         else
416                 seq_printf(m, "MMU Type\t: ???\n");
417
418 #ifdef CONFIG_DEBUG_DCFLUSH
419         seq_printf(m, "DCPageFlushes\t: %d\n",
420                    atomic_read(&dcpage_flushes));
421 #ifdef CONFIG_SMP
422         seq_printf(m, "DCPageFlushesXC\t: %d\n",
423                    atomic_read(&dcpage_flushes_xcall));
424 #endif /* CONFIG_SMP */
425 #endif /* CONFIG_DEBUG_DCFLUSH */
426 }
427
428 struct linux_prom_translation prom_trans[512] __read_mostly;
429 unsigned int prom_trans_ents __read_mostly;
430
431 unsigned long kern_locked_tte_data;
432
433 /* The obp translations are saved based on 8k pagesize, since obp can
434  * use a mixture of pagesizes. Misses to the LOW_OBP_ADDRESS ->
435  * HI_OBP_ADDRESS range are handled in ktlb.S.
436  */
437 static inline int in_obp_range(unsigned long vaddr)
438 {
439         return (vaddr >= LOW_OBP_ADDRESS &&
440                 vaddr < HI_OBP_ADDRESS);
441 }
442
443 static int cmp_ptrans(const void *a, const void *b)
444 {
445         const struct linux_prom_translation *x = a, *y = b;
446
447         if (x->virt > y->virt)
448                 return 1;
449         if (x->virt < y->virt)
450                 return -1;
451         return 0;
452 }
453
454 /* Read OBP translations property into 'prom_trans[]'.  */
455 static void __init read_obp_translations(void)
456 {
457         int n, node, ents, first, last, i;
458
459         node = prom_finddevice("/virtual-memory");
460         n = prom_getproplen(node, "translations");
461         if (unlikely(n == 0 || n == -1)) {
462                 prom_printf("prom_mappings: Couldn't get size.\n");
463                 prom_halt();
464         }
465         if (unlikely(n > sizeof(prom_trans))) {
466                 prom_printf("prom_mappings: Size %Zd is too big.\n", n);
467                 prom_halt();
468         }
469
470         if ((n = prom_getproperty(node, "translations",
471                                   (char *)&prom_trans[0],
472                                   sizeof(prom_trans))) == -1) {
473                 prom_printf("prom_mappings: Couldn't get property.\n");
474                 prom_halt();
475         }
476
477         n = n / sizeof(struct linux_prom_translation);
478
479         ents = n;
480
481         sort(prom_trans, ents, sizeof(struct linux_prom_translation),
482              cmp_ptrans, NULL);
483
484         /* Now kick out all the non-OBP entries.  */
485         for (i = 0; i < ents; i++) {
486                 if (in_obp_range(prom_trans[i].virt))
487                         break;
488         }
489         first = i;
490         for (; i < ents; i++) {
491                 if (!in_obp_range(prom_trans[i].virt))
492                         break;
493         }
494         last = i;
495
496         for (i = 0; i < (last - first); i++) {
497                 struct linux_prom_translation *src = &prom_trans[i + first];
498                 struct linux_prom_translation *dest = &prom_trans[i];
499
500                 *dest = *src;
501         }
502         for (; i < ents; i++) {
503                 struct linux_prom_translation *dest = &prom_trans[i];
504                 dest->virt = dest->size = dest->data = 0x0UL;
505         }
506
507         prom_trans_ents = last - first;
508
509         if (tlb_type == spitfire) {
510                 /* Clear diag TTE bits. */
511                 for (i = 0; i < prom_trans_ents; i++)
512                         prom_trans[i].data &= ~0x0003fe0000000000UL;
513         }
514 }
515
516 static void __init hypervisor_tlb_lock(unsigned long vaddr,
517                                        unsigned long pte,
518                                        unsigned long mmu)
519 {
520         unsigned long ret = sun4v_mmu_map_perm_addr(vaddr, 0, pte, mmu);
521
522         if (ret != 0) {
523                 prom_printf("hypervisor_tlb_lock[%lx:%lx:%lx:%lx]: "
524                             "errors with %lx\n", vaddr, 0, pte, mmu, ret);
525                 prom_halt();
526         }
527 }
528
529 static unsigned long kern_large_tte(unsigned long paddr);
530
531 static void __init remap_kernel(void)
532 {
533         unsigned long phys_page, tte_vaddr, tte_data;
534         int i, tlb_ent = sparc64_highest_locked_tlbent();
535
536         tte_vaddr = (unsigned long) KERNBASE;
537         phys_page = (prom_boot_mapping_phys_low >> 22UL) << 22UL;
538         tte_data = kern_large_tte(phys_page);
539
540         kern_locked_tte_data = tte_data;
541
542         /* Now lock us into the TLBs via Hypervisor or OBP. */
543         if (tlb_type == hypervisor) {
544                 for (i = 0; i < num_kernel_image_mappings; i++) {
545                         hypervisor_tlb_lock(tte_vaddr, tte_data, HV_MMU_DMMU);
546                         hypervisor_tlb_lock(tte_vaddr, tte_data, HV_MMU_IMMU);
547                         tte_vaddr += 0x400000;
548                         tte_data += 0x400000;
549                 }
550         } else {
551                 for (i = 0; i < num_kernel_image_mappings; i++) {
552                         prom_dtlb_load(tlb_ent - i, tte_data, tte_vaddr);
553                         prom_itlb_load(tlb_ent - i, tte_data, tte_vaddr);
554                         tte_vaddr += 0x400000;
555                         tte_data += 0x400000;
556                 }
557                 sparc64_highest_unlocked_tlb_ent = tlb_ent - i;
558         }
559         if (tlb_type == cheetah_plus) {
560                 sparc64_kern_pri_context = (CTX_CHEETAH_PLUS_CTX0 |
561                                             CTX_CHEETAH_PLUS_NUC);
562                 sparc64_kern_pri_nuc_bits = CTX_CHEETAH_PLUS_NUC;
563                 sparc64_kern_sec_context = CTX_CHEETAH_PLUS_CTX0;
564         }
565 }
566
567
568 static void __init inherit_prom_mappings(void)
569 {
570         /* Now fixup OBP's idea about where we really are mapped. */
571         printk("Remapping the kernel... ");
572         remap_kernel();
573         printk("done.\n");
574 }
575
576 void prom_world(int enter)
577 {
578         if (!enter)
579                 set_fs((mm_segment_t) { get_thread_current_ds() });
580
581         __asm__ __volatile__("flushw");
582 }
583
584 void __flush_dcache_range(unsigned long start, unsigned long end)
585 {
586         unsigned long va;
587
588         if (tlb_type == spitfire) {
589                 int n = 0;
590
591                 for (va = start; va < end; va += 32) {
592                         spitfire_put_dcache_tag(va & 0x3fe0, 0x0);
593                         if (++n >= 512)
594                                 break;
595                 }
596         } else if (tlb_type == cheetah || tlb_type == cheetah_plus) {
597                 start = __pa(start);
598                 end = __pa(end);
599                 for (va = start; va < end; va += 32)
600                         __asm__ __volatile__("stxa %%g0, [%0] %1\n\t"
601                                              "membar #Sync"
602                                              : /* no outputs */
603                                              : "r" (va),
604                                                "i" (ASI_DCACHE_INVALIDATE));
605         }
606 }
607 EXPORT_SYMBOL(__flush_dcache_range);
608
609 /* get_new_mmu_context() uses "cache + 1".  */
610 DEFINE_SPINLOCK(ctx_alloc_lock);
611 unsigned long tlb_context_cache = CTX_FIRST_VERSION - 1;
612 #define MAX_CTX_NR      (1UL << CTX_NR_BITS)
613 #define CTX_BMAP_SLOTS  BITS_TO_LONGS(MAX_CTX_NR)
614 DECLARE_BITMAP(mmu_context_bmap, MAX_CTX_NR);
615
616 /* Caller does TLB context flushing on local CPU if necessary.
617  * The caller also ensures that CTX_VALID(mm->context) is false.
618  *
619  * We must be careful about boundary cases so that we never
620  * let the user have CTX 0 (nucleus) or we ever use a CTX
621  * version of zero (and thus NO_CONTEXT would not be caught
622  * by version mis-match tests in mmu_context.h).
623  *
624  * Always invoked with interrupts disabled.
625  */
626 void get_new_mmu_context(struct mm_struct *mm)
627 {
628         unsigned long ctx, new_ctx;
629         unsigned long orig_pgsz_bits;
630         unsigned long flags;
631         int new_version;
632
633         spin_lock_irqsave(&ctx_alloc_lock, flags);
634         orig_pgsz_bits = (mm->context.sparc64_ctx_val & CTX_PGSZ_MASK);
635         ctx = (tlb_context_cache + 1) & CTX_NR_MASK;
636         new_ctx = find_next_zero_bit(mmu_context_bmap, 1 << CTX_NR_BITS, ctx);
637         new_version = 0;
638         if (new_ctx >= (1 << CTX_NR_BITS)) {
639                 new_ctx = find_next_zero_bit(mmu_context_bmap, ctx, 1);
640                 if (new_ctx >= ctx) {
641                         int i;
642                         new_ctx = (tlb_context_cache & CTX_VERSION_MASK) +
643                                 CTX_FIRST_VERSION;
644                         if (new_ctx == 1)
645                                 new_ctx = CTX_FIRST_VERSION;
646
647                         /* Don't call memset, for 16 entries that's just
648                          * plain silly...
649                          */
650                         mmu_context_bmap[0] = 3;
651                         mmu_context_bmap[1] = 0;
652                         mmu_context_bmap[2] = 0;
653                         mmu_context_bmap[3] = 0;
654                         for (i = 4; i < CTX_BMAP_SLOTS; i += 4) {
655                                 mmu_context_bmap[i + 0] = 0;
656                                 mmu_context_bmap[i + 1] = 0;
657                                 mmu_context_bmap[i + 2] = 0;
658                                 mmu_context_bmap[i + 3] = 0;
659                         }
660                         new_version = 1;
661                         goto out;
662                 }
663         }
664         mmu_context_bmap[new_ctx>>6] |= (1UL << (new_ctx & 63));
665         new_ctx |= (tlb_context_cache & CTX_VERSION_MASK);
666 out:
667         tlb_context_cache = new_ctx;
668         mm->context.sparc64_ctx_val = new_ctx | orig_pgsz_bits;
669         spin_unlock_irqrestore(&ctx_alloc_lock, flags);
670
671         if (unlikely(new_version))
672                 smp_new_mmu_context_version();
673 }
674
675 static int numa_enabled = 1;
676 static int numa_debug;
677
678 static int __init early_numa(char *p)
679 {
680         if (!p)
681                 return 0;
682
683         if (strstr(p, "off"))
684                 numa_enabled = 0;
685
686         if (strstr(p, "debug"))
687                 numa_debug = 1;
688
689         return 0;
690 }
691 early_param("numa", early_numa);
692
693 #define numadbg(f, a...) \
694 do {    if (numa_debug) \
695                 printk(KERN_INFO f, ## a); \
696 } while (0)
697
698 static void __init find_ramdisk(unsigned long phys_base)
699 {
700 #ifdef CONFIG_BLK_DEV_INITRD
701         if (sparc_ramdisk_image || sparc_ramdisk_image64) {
702                 unsigned long ramdisk_image;
703
704                 /* Older versions of the bootloader only supported a
705                  * 32-bit physical address for the ramdisk image
706                  * location, stored at sparc_ramdisk_image.  Newer
707                  * SILO versions set sparc_ramdisk_image to zero and
708                  * provide a full 64-bit physical address at
709                  * sparc_ramdisk_image64.
710                  */
711                 ramdisk_image = sparc_ramdisk_image;
712                 if (!ramdisk_image)
713                         ramdisk_image = sparc_ramdisk_image64;
714
715                 /* Another bootloader quirk.  The bootloader normalizes
716                  * the physical address to KERNBASE, so we have to
717                  * factor that back out and add in the lowest valid
718                  * physical page address to get the true physical address.
719                  */
720                 ramdisk_image -= KERNBASE;
721                 ramdisk_image += phys_base;
722
723                 numadbg("Found ramdisk at physical address 0x%lx, size %u\n",
724                         ramdisk_image, sparc_ramdisk_size);
725
726                 initrd_start = ramdisk_image;
727                 initrd_end = ramdisk_image + sparc_ramdisk_size;
728
729                 lmb_reserve(initrd_start, sparc_ramdisk_size);
730
731                 initrd_start += PAGE_OFFSET;
732                 initrd_end += PAGE_OFFSET;
733         }
734 #endif
735 }
736
737 struct node_mem_mask {
738         unsigned long mask;
739         unsigned long val;
740         unsigned long bootmem_paddr;
741 };
742 static struct node_mem_mask node_masks[MAX_NUMNODES];
743 static int num_node_masks;
744
745 int numa_cpu_lookup_table[NR_CPUS];
746 cpumask_t numa_cpumask_lookup_table[MAX_NUMNODES];
747
748 #ifdef CONFIG_NEED_MULTIPLE_NODES
749
750 struct mdesc_mblock {
751         u64     base;
752         u64     size;
753         u64     offset; /* RA-to-PA */
754 };
755 static struct mdesc_mblock *mblocks;
756 static int num_mblocks;
757
758 static unsigned long ra_to_pa(unsigned long addr)
759 {
760         int i;
761
762         for (i = 0; i < num_mblocks; i++) {
763                 struct mdesc_mblock *m = &mblocks[i];
764
765                 if (addr >= m->base &&
766                     addr < (m->base + m->size)) {
767                         addr += m->offset;
768                         break;
769                 }
770         }
771         return addr;
772 }
773
774 static int find_node(unsigned long addr)
775 {
776         int i;
777
778         addr = ra_to_pa(addr);
779         for (i = 0; i < num_node_masks; i++) {
780                 struct node_mem_mask *p = &node_masks[i];
781
782                 if ((addr & p->mask) == p->val)
783                         return i;
784         }
785         return -1;
786 }
787
788 static unsigned long long nid_range(unsigned long long start,
789                                     unsigned long long end, int *nid)
790 {
791         *nid = find_node(start);
792         start += PAGE_SIZE;
793         while (start < end) {
794                 int n = find_node(start);
795
796                 if (n != *nid)
797                         break;
798                 start += PAGE_SIZE;
799         }
800
801         if (start > end)
802                 start = end;
803
804         return start;
805 }
806 #else
807 static unsigned long long nid_range(unsigned long long start,
808                                     unsigned long long end, int *nid)
809 {
810         *nid = 0;
811         return end;
812 }
813 #endif
814
815 /* This must be invoked after performing all of the necessary
816  * add_active_range() calls for 'nid'.  We need to be able to get
817  * correct data from get_pfn_range_for_nid().
818  */
819 static void __init allocate_node_data(int nid)
820 {
821         unsigned long paddr, num_pages, start_pfn, end_pfn;
822         struct pglist_data *p;
823
824 #ifdef CONFIG_NEED_MULTIPLE_NODES
825         paddr = lmb_alloc_nid(sizeof(struct pglist_data),
826                               SMP_CACHE_BYTES, nid, nid_range);
827         if (!paddr) {
828                 prom_printf("Cannot allocate pglist_data for nid[%d]\n", nid);
829                 prom_halt();
830         }
831         NODE_DATA(nid) = __va(paddr);
832         memset(NODE_DATA(nid), 0, sizeof(struct pglist_data));
833
834         NODE_DATA(nid)->bdata = &bootmem_node_data[nid];
835 #endif
836
837         p = NODE_DATA(nid);
838
839         get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
840         p->node_start_pfn = start_pfn;
841         p->node_spanned_pages = end_pfn - start_pfn;
842
843         if (p->node_spanned_pages) {
844                 num_pages = bootmem_bootmap_pages(p->node_spanned_pages);
845
846                 paddr = lmb_alloc_nid(num_pages << PAGE_SHIFT, PAGE_SIZE, nid,
847                                       nid_range);
848                 if (!paddr) {
849                         prom_printf("Cannot allocate bootmap for nid[%d]\n",
850                                   nid);
851                         prom_halt();
852                 }
853                 node_masks[nid].bootmem_paddr = paddr;
854         }
855 }
856
857 static void init_node_masks_nonnuma(void)
858 {
859         int i;
860
861         numadbg("Initializing tables for non-numa.\n");
862
863         node_masks[0].mask = node_masks[0].val = 0;
864         num_node_masks = 1;
865
866         for (i = 0; i < NR_CPUS; i++)
867                 numa_cpu_lookup_table[i] = 0;
868
869         numa_cpumask_lookup_table[0] = CPU_MASK_ALL;
870 }
871
872 #ifdef CONFIG_NEED_MULTIPLE_NODES
873 struct pglist_data *node_data[MAX_NUMNODES];
874
875 EXPORT_SYMBOL(numa_cpu_lookup_table);
876 EXPORT_SYMBOL(numa_cpumask_lookup_table);
877 EXPORT_SYMBOL(node_data);
878
879 struct mdesc_mlgroup {
880         u64     node;
881         u64     latency;
882         u64     match;
883         u64     mask;
884 };
885 static struct mdesc_mlgroup *mlgroups;
886 static int num_mlgroups;
887
888 static int scan_pio_for_cfg_handle(struct mdesc_handle *md, u64 pio,
889                                    u32 cfg_handle)
890 {
891         u64 arc;
892
893         mdesc_for_each_arc(arc, md, pio, MDESC_ARC_TYPE_FWD) {
894                 u64 target = mdesc_arc_target(md, arc);
895                 const u64 *val;
896
897                 val = mdesc_get_property(md, target,
898                                          "cfg-handle", NULL);
899                 if (val && *val == cfg_handle)
900                         return 0;
901         }
902         return -ENODEV;
903 }
904
905 static int scan_arcs_for_cfg_handle(struct mdesc_handle *md, u64 grp,
906                                     u32 cfg_handle)
907 {
908         u64 arc, candidate, best_latency = ~(u64)0;
909
910         candidate = MDESC_NODE_NULL;
911         mdesc_for_each_arc(arc, md, grp, MDESC_ARC_TYPE_FWD) {
912                 u64 target = mdesc_arc_target(md, arc);
913                 const char *name = mdesc_node_name(md, target);
914                 const u64 *val;
915
916                 if (strcmp(name, "pio-latency-group"))
917                         continue;
918
919                 val = mdesc_get_property(md, target, "latency", NULL);
920                 if (!val)
921                         continue;
922
923                 if (*val < best_latency) {
924                         candidate = target;
925                         best_latency = *val;
926                 }
927         }
928
929         if (candidate == MDESC_NODE_NULL)
930                 return -ENODEV;
931
932         return scan_pio_for_cfg_handle(md, candidate, cfg_handle);
933 }
934
935 int of_node_to_nid(struct device_node *dp)
936 {
937         const struct linux_prom64_registers *regs;
938         struct mdesc_handle *md;
939         u32 cfg_handle;
940         int count, nid;
941         u64 grp;
942
943         /* This is the right thing to do on currently supported
944          * SUN4U NUMA platforms as well, as the PCI controller does
945          * not sit behind any particular memory controller.
946          */
947         if (!mlgroups)
948                 return -1;
949
950         regs = of_get_property(dp, "reg", NULL);
951         if (!regs)
952                 return -1;
953
954         cfg_handle = (regs->phys_addr >> 32UL) & 0x0fffffff;
955
956         md = mdesc_grab();
957
958         count = 0;
959         nid = -1;
960         mdesc_for_each_node_by_name(md, grp, "group") {
961                 if (!scan_arcs_for_cfg_handle(md, grp, cfg_handle)) {
962                         nid = count;
963                         break;
964                 }
965                 count++;
966         }
967
968         mdesc_release(md);
969
970         return nid;
971 }
972
973 static void __init add_node_ranges(void)
974 {
975         int i;
976
977         for (i = 0; i < lmb.memory.cnt; i++) {
978                 unsigned long size = lmb_size_bytes(&lmb.memory, i);
979                 unsigned long start, end;
980
981                 start = lmb.memory.region[i].base;
982                 end = start + size;
983                 while (start < end) {
984                         unsigned long this_end;
985                         int nid;
986
987                         this_end = nid_range(start, end, &nid);
988
989                         numadbg("Adding active range nid[%d] "
990                                 "start[%lx] end[%lx]\n",
991                                 nid, start, this_end);
992
993                         add_active_range(nid,
994                                          start >> PAGE_SHIFT,
995                                          this_end >> PAGE_SHIFT);
996
997                         start = this_end;
998                 }
999         }
1000 }
1001
1002 static int __init grab_mlgroups(struct mdesc_handle *md)
1003 {
1004         unsigned long paddr;
1005         int count = 0;
1006         u64 node;
1007
1008         mdesc_for_each_node_by_name(md, node, "memory-latency-group")
1009                 count++;
1010         if (!count)
1011                 return -ENOENT;
1012
1013         paddr = lmb_alloc(count * sizeof(struct mdesc_mlgroup),
1014                           SMP_CACHE_BYTES);
1015         if (!paddr)
1016                 return -ENOMEM;
1017
1018         mlgroups = __va(paddr);
1019         num_mlgroups = count;
1020
1021         count = 0;
1022         mdesc_for_each_node_by_name(md, node, "memory-latency-group") {
1023                 struct mdesc_mlgroup *m = &mlgroups[count++];
1024                 const u64 *val;
1025
1026                 m->node = node;
1027
1028                 val = mdesc_get_property(md, node, "latency", NULL);
1029                 m->latency = *val;
1030                 val = mdesc_get_property(md, node, "address-match", NULL);
1031                 m->match = *val;
1032                 val = mdesc_get_property(md, node, "address-mask", NULL);
1033                 m->mask = *val;
1034
1035                 numadbg("MLGROUP[%d]: node[%llx] latency[%llx] "
1036                         "match[%llx] mask[%llx]\n",
1037                         count - 1, m->node, m->latency, m->match, m->mask);
1038         }
1039
1040         return 0;
1041 }
1042
1043 static int __init grab_mblocks(struct mdesc_handle *md)
1044 {
1045         unsigned long paddr;
1046         int count = 0;
1047         u64 node;
1048
1049         mdesc_for_each_node_by_name(md, node, "mblock")
1050                 count++;
1051         if (!count)
1052                 return -ENOENT;
1053
1054         paddr = lmb_alloc(count * sizeof(struct mdesc_mblock),
1055                           SMP_CACHE_BYTES);
1056         if (!paddr)
1057                 return -ENOMEM;
1058
1059         mblocks = __va(paddr);
1060         num_mblocks = count;
1061
1062         count = 0;
1063         mdesc_for_each_node_by_name(md, node, "mblock") {
1064                 struct mdesc_mblock *m = &mblocks[count++];
1065                 const u64 *val;
1066
1067                 val = mdesc_get_property(md, node, "base", NULL);
1068                 m->base = *val;
1069                 val = mdesc_get_property(md, node, "size", NULL);
1070                 m->size = *val;
1071                 val = mdesc_get_property(md, node,
1072                                          "address-congruence-offset", NULL);
1073                 m->offset = *val;
1074
1075                 numadbg("MBLOCK[%d]: base[%llx] size[%llx] offset[%llx]\n",
1076                         count - 1, m->base, m->size, m->offset);
1077         }
1078
1079         return 0;
1080 }
1081
1082 static void __init numa_parse_mdesc_group_cpus(struct mdesc_handle *md,
1083                                                u64 grp, cpumask_t *mask)
1084 {
1085         u64 arc;
1086
1087         cpus_clear(*mask);
1088
1089         mdesc_for_each_arc(arc, md, grp, MDESC_ARC_TYPE_BACK) {
1090                 u64 target = mdesc_arc_target(md, arc);
1091                 const char *name = mdesc_node_name(md, target);
1092                 const u64 *id;
1093
1094                 if (strcmp(name, "cpu"))
1095                         continue;
1096                 id = mdesc_get_property(md, target, "id", NULL);
1097                 if (*id < nr_cpu_ids)
1098                         cpu_set(*id, *mask);
1099         }
1100 }
1101
1102 static struct mdesc_mlgroup * __init find_mlgroup(u64 node)
1103 {
1104         int i;
1105
1106         for (i = 0; i < num_mlgroups; i++) {
1107                 struct mdesc_mlgroup *m = &mlgroups[i];
1108                 if (m->node == node)
1109                         return m;
1110         }
1111         return NULL;
1112 }
1113
1114 static int __init numa_attach_mlgroup(struct mdesc_handle *md, u64 grp,
1115                                       int index)
1116 {
1117         struct mdesc_mlgroup *candidate = NULL;
1118         u64 arc, best_latency = ~(u64)0;
1119         struct node_mem_mask *n;
1120
1121         mdesc_for_each_arc(arc, md, grp, MDESC_ARC_TYPE_FWD) {
1122                 u64 target = mdesc_arc_target(md, arc);
1123                 struct mdesc_mlgroup *m = find_mlgroup(target);
1124                 if (!m)
1125                         continue;
1126                 if (m->latency < best_latency) {
1127                         candidate = m;
1128                         best_latency = m->latency;
1129                 }
1130         }
1131         if (!candidate)
1132                 return -ENOENT;
1133
1134         if (num_node_masks != index) {
1135                 printk(KERN_ERR "Inconsistent NUMA state, "
1136                        "index[%d] != num_node_masks[%d]\n",
1137                        index, num_node_masks);
1138                 return -EINVAL;
1139         }
1140
1141         n = &node_masks[num_node_masks++];
1142
1143         n->mask = candidate->mask;
1144         n->val = candidate->match;
1145
1146         numadbg("NUMA NODE[%d]: mask[%lx] val[%lx] (latency[%llx])\n",
1147                 index, n->mask, n->val, candidate->latency);
1148
1149         return 0;
1150 }
1151
1152 static int __init numa_parse_mdesc_group(struct mdesc_handle *md, u64 grp,
1153                                          int index)
1154 {
1155         cpumask_t mask;
1156         int cpu;
1157
1158         numa_parse_mdesc_group_cpus(md, grp, &mask);
1159
1160         for_each_cpu_mask(cpu, mask)
1161                 numa_cpu_lookup_table[cpu] = index;
1162         numa_cpumask_lookup_table[index] = mask;
1163
1164         if (numa_debug) {
1165                 printk(KERN_INFO "NUMA GROUP[%d]: cpus [ ", index);
1166                 for_each_cpu_mask(cpu, mask)
1167                         printk("%d ", cpu);
1168                 printk("]\n");
1169         }
1170
1171         return numa_attach_mlgroup(md, grp, index);
1172 }
1173
1174 static int __init numa_parse_mdesc(void)
1175 {
1176         struct mdesc_handle *md = mdesc_grab();
1177         int i, err, count;
1178         u64 node;
1179
1180         node = mdesc_node_by_name(md, MDESC_NODE_NULL, "latency-groups");
1181         if (node == MDESC_NODE_NULL) {
1182                 mdesc_release(md);
1183                 return -ENOENT;
1184         }
1185
1186         err = grab_mblocks(md);
1187         if (err < 0)
1188                 goto out;
1189
1190         err = grab_mlgroups(md);
1191         if (err < 0)
1192                 goto out;
1193
1194         count = 0;
1195         mdesc_for_each_node_by_name(md, node, "group") {
1196                 err = numa_parse_mdesc_group(md, node, count);
1197                 if (err < 0)
1198                         break;
1199                 count++;
1200         }
1201
1202         add_node_ranges();
1203
1204         for (i = 0; i < num_node_masks; i++) {
1205                 allocate_node_data(i);
1206                 node_set_online(i);
1207         }
1208
1209         err = 0;
1210 out:
1211         mdesc_release(md);
1212         return err;
1213 }
1214
1215 static int __init numa_parse_jbus(void)
1216 {
1217         unsigned long cpu, index;
1218
1219         /* NUMA node id is encoded in bits 36 and higher, and there is
1220          * a 1-to-1 mapping from CPU ID to NUMA node ID.
1221          */
1222         index = 0;
1223         for_each_present_cpu(cpu) {
1224                 numa_cpu_lookup_table[cpu] = index;
1225                 numa_cpumask_lookup_table[index] = cpumask_of_cpu(cpu);
1226                 node_masks[index].mask = ~((1UL << 36UL) - 1UL);
1227                 node_masks[index].val = cpu << 36UL;
1228
1229                 index++;
1230         }
1231         num_node_masks = index;
1232
1233         add_node_ranges();
1234
1235         for (index = 0; index < num_node_masks; index++) {
1236                 allocate_node_data(index);
1237                 node_set_online(index);
1238         }
1239
1240         return 0;
1241 }
1242
1243 static int __init numa_parse_sun4u(void)
1244 {
1245         if (tlb_type == cheetah || tlb_type == cheetah_plus) {
1246                 unsigned long ver;
1247
1248                 __asm__ ("rdpr %%ver, %0" : "=r" (ver));
1249                 if ((ver >> 32UL) == __JALAPENO_ID ||
1250                     (ver >> 32UL) == __SERRANO_ID)
1251                         return numa_parse_jbus();
1252         }
1253         return -1;
1254 }
1255
1256 static int __init bootmem_init_numa(void)
1257 {
1258         int err = -1;
1259
1260         numadbg("bootmem_init_numa()\n");
1261
1262         if (numa_enabled) {
1263                 if (tlb_type == hypervisor)
1264                         err = numa_parse_mdesc();
1265                 else
1266                         err = numa_parse_sun4u();
1267         }
1268         return err;
1269 }
1270
1271 #else
1272
1273 static int bootmem_init_numa(void)
1274 {
1275         return -1;
1276 }
1277
1278 #endif
1279
1280 static void __init bootmem_init_nonnuma(void)
1281 {
1282         unsigned long top_of_ram = lmb_end_of_DRAM();
1283         unsigned long total_ram = lmb_phys_mem_size();
1284         unsigned int i;
1285
1286         numadbg("bootmem_init_nonnuma()\n");
1287
1288         printk(KERN_INFO "Top of RAM: 0x%lx, Total RAM: 0x%lx\n",
1289                top_of_ram, total_ram);
1290         printk(KERN_INFO "Memory hole size: %ldMB\n",
1291                (top_of_ram - total_ram) >> 20);
1292
1293         init_node_masks_nonnuma();
1294
1295         for (i = 0; i < lmb.memory.cnt; i++) {
1296                 unsigned long size = lmb_size_bytes(&lmb.memory, i);
1297                 unsigned long start_pfn, end_pfn;
1298
1299                 if (!size)
1300                         continue;
1301
1302                 start_pfn = lmb.memory.region[i].base >> PAGE_SHIFT;
1303                 end_pfn = start_pfn + lmb_size_pages(&lmb.memory, i);
1304                 add_active_range(0, start_pfn, end_pfn);
1305         }
1306
1307         allocate_node_data(0);
1308
1309         node_set_online(0);
1310 }
1311
1312 static void __init reserve_range_in_node(int nid, unsigned long start,
1313                                          unsigned long end)
1314 {
1315         numadbg("    reserve_range_in_node(nid[%d],start[%lx],end[%lx]\n",
1316                 nid, start, end);
1317         while (start < end) {
1318                 unsigned long this_end;
1319                 int n;
1320
1321                 this_end = nid_range(start, end, &n);
1322                 if (n == nid) {
1323                         numadbg("      MATCH reserving range [%lx:%lx]\n",
1324                                 start, this_end);
1325                         reserve_bootmem_node(NODE_DATA(nid), start,
1326                                              (this_end - start), BOOTMEM_DEFAULT);
1327                 } else
1328                         numadbg("      NO MATCH, advancing start to %lx\n",
1329                                 this_end);
1330
1331                 start = this_end;
1332         }
1333 }
1334
1335 static void __init trim_reserved_in_node(int nid)
1336 {
1337         int i;
1338
1339         numadbg("  trim_reserved_in_node(%d)\n", nid);
1340
1341         for (i = 0; i < lmb.reserved.cnt; i++) {
1342                 unsigned long start = lmb.reserved.region[i].base;
1343                 unsigned long size = lmb_size_bytes(&lmb.reserved, i);
1344                 unsigned long end = start + size;
1345
1346                 reserve_range_in_node(nid, start, end);
1347         }
1348 }
1349
1350 static void __init bootmem_init_one_node(int nid)
1351 {
1352         struct pglist_data *p;
1353
1354         numadbg("bootmem_init_one_node(%d)\n", nid);
1355
1356         p = NODE_DATA(nid);
1357
1358         if (p->node_spanned_pages) {
1359                 unsigned long paddr = node_masks[nid].bootmem_paddr;
1360                 unsigned long end_pfn;
1361
1362                 end_pfn = p->node_start_pfn + p->node_spanned_pages;
1363
1364                 numadbg("  init_bootmem_node(%d, %lx, %lx, %lx)\n",
1365                         nid, paddr >> PAGE_SHIFT, p->node_start_pfn, end_pfn);
1366
1367                 init_bootmem_node(p, paddr >> PAGE_SHIFT,
1368                                   p->node_start_pfn, end_pfn);
1369
1370                 numadbg("  free_bootmem_with_active_regions(%d, %lx)\n",
1371                         nid, end_pfn);
1372                 free_bootmem_with_active_regions(nid, end_pfn);
1373
1374                 trim_reserved_in_node(nid);
1375
1376                 numadbg("  sparse_memory_present_with_active_regions(%d)\n",
1377                         nid);
1378                 sparse_memory_present_with_active_regions(nid);
1379         }
1380 }
1381
1382 static unsigned long __init bootmem_init(unsigned long phys_base)
1383 {
1384         unsigned long end_pfn;
1385         int nid;
1386
1387         end_pfn = lmb_end_of_DRAM() >> PAGE_SHIFT;
1388         max_pfn = max_low_pfn = end_pfn;
1389         min_low_pfn = (phys_base >> PAGE_SHIFT);
1390
1391         if (bootmem_init_numa() < 0)
1392                 bootmem_init_nonnuma();
1393
1394         /* XXX cpu notifier XXX */
1395
1396         for_each_online_node(nid)
1397                 bootmem_init_one_node(nid);
1398
1399         sparse_init();
1400
1401         return end_pfn;
1402 }
1403
1404 static struct linux_prom64_registers pall[MAX_BANKS] __initdata;
1405 static int pall_ents __initdata;
1406
1407 #ifdef CONFIG_DEBUG_PAGEALLOC
1408 static unsigned long __ref kernel_map_range(unsigned long pstart,
1409                                             unsigned long pend, pgprot_t prot)
1410 {
1411         unsigned long vstart = PAGE_OFFSET + pstart;
1412         unsigned long vend = PAGE_OFFSET + pend;
1413         unsigned long alloc_bytes = 0UL;
1414
1415         if ((vstart & ~PAGE_MASK) || (vend & ~PAGE_MASK)) {
1416                 prom_printf("kernel_map: Unaligned physmem[%lx:%lx]\n",
1417                             vstart, vend);
1418                 prom_halt();
1419         }
1420
1421         while (vstart < vend) {
1422                 unsigned long this_end, paddr = __pa(vstart);
1423                 pgd_t *pgd = pgd_offset_k(vstart);
1424                 pud_t *pud;
1425                 pmd_t *pmd;
1426                 pte_t *pte;
1427
1428                 pud = pud_offset(pgd, vstart);
1429                 if (pud_none(*pud)) {
1430                         pmd_t *new;
1431
1432                         new = __alloc_bootmem(PAGE_SIZE, PAGE_SIZE, PAGE_SIZE);
1433                         alloc_bytes += PAGE_SIZE;
1434                         pud_populate(&init_mm, pud, new);
1435                 }
1436
1437                 pmd = pmd_offset(pud, vstart);
1438                 if (!pmd_present(*pmd)) {
1439                         pte_t *new;
1440
1441                         new = __alloc_bootmem(PAGE_SIZE, PAGE_SIZE, PAGE_SIZE);
1442                         alloc_bytes += PAGE_SIZE;
1443                         pmd_populate_kernel(&init_mm, pmd, new);
1444                 }
1445
1446                 pte = pte_offset_kernel(pmd, vstart);
1447                 this_end = (vstart + PMD_SIZE) & PMD_MASK;
1448                 if (this_end > vend)
1449                         this_end = vend;
1450
1451                 while (vstart < this_end) {
1452                         pte_val(*pte) = (paddr | pgprot_val(prot));
1453
1454                         vstart += PAGE_SIZE;
1455                         paddr += PAGE_SIZE;
1456                         pte++;
1457                 }
1458         }
1459
1460         return alloc_bytes;
1461 }
1462
1463 extern unsigned int kvmap_linear_patch[1];
1464 #endif /* CONFIG_DEBUG_PAGEALLOC */
1465
1466 static void __init mark_kpte_bitmap(unsigned long start, unsigned long end)
1467 {
1468         const unsigned long shift_256MB = 28;
1469         const unsigned long mask_256MB = ((1UL << shift_256MB) - 1UL);
1470         const unsigned long size_256MB = (1UL << shift_256MB);
1471
1472         while (start < end) {
1473                 long remains;
1474
1475                 remains = end - start;
1476                 if (remains < size_256MB)
1477                         break;
1478
1479                 if (start & mask_256MB) {
1480                         start = (start + size_256MB) & ~mask_256MB;
1481                         continue;
1482                 }
1483
1484                 while (remains >= size_256MB) {
1485                         unsigned long index = start >> shift_256MB;
1486
1487                         __set_bit(index, kpte_linear_bitmap);
1488
1489                         start += size_256MB;
1490                         remains -= size_256MB;
1491                 }
1492         }
1493 }
1494
1495 static void __init init_kpte_bitmap(void)
1496 {
1497         unsigned long i;
1498
1499         for (i = 0; i < pall_ents; i++) {
1500                 unsigned long phys_start, phys_end;
1501
1502                 phys_start = pall[i].phys_addr;
1503                 phys_end = phys_start + pall[i].reg_size;
1504
1505                 mark_kpte_bitmap(phys_start, phys_end);
1506         }
1507 }
1508
1509 static void __init kernel_physical_mapping_init(void)
1510 {
1511 #ifdef CONFIG_DEBUG_PAGEALLOC
1512         unsigned long i, mem_alloced = 0UL;
1513
1514         for (i = 0; i < pall_ents; i++) {
1515                 unsigned long phys_start, phys_end;
1516
1517                 phys_start = pall[i].phys_addr;
1518                 phys_end = phys_start + pall[i].reg_size;
1519
1520                 mem_alloced += kernel_map_range(phys_start, phys_end,
1521                                                 PAGE_KERNEL);
1522         }
1523
1524         printk("Allocated %ld bytes for kernel page tables.\n",
1525                mem_alloced);
1526
1527         kvmap_linear_patch[0] = 0x01000000; /* nop */
1528         flushi(&kvmap_linear_patch[0]);
1529
1530         __flush_tlb_all();
1531 #endif
1532 }
1533
1534 #ifdef CONFIG_DEBUG_PAGEALLOC
1535 void kernel_map_pages(struct page *page, int numpages, int enable)
1536 {
1537         unsigned long phys_start = page_to_pfn(page) << PAGE_SHIFT;
1538         unsigned long phys_end = phys_start + (numpages * PAGE_SIZE);
1539
1540         kernel_map_range(phys_start, phys_end,
1541                          (enable ? PAGE_KERNEL : __pgprot(0)));
1542
1543         flush_tsb_kernel_range(PAGE_OFFSET + phys_start,
1544                                PAGE_OFFSET + phys_end);
1545
1546         /* we should perform an IPI and flush all tlbs,
1547          * but that can deadlock->flush only current cpu.
1548          */
1549         __flush_tlb_kernel_range(PAGE_OFFSET + phys_start,
1550                                  PAGE_OFFSET + phys_end);
1551 }
1552 #endif
1553
1554 unsigned long __init find_ecache_flush_span(unsigned long size)
1555 {
1556         int i;
1557
1558         for (i = 0; i < pavail_ents; i++) {
1559                 if (pavail[i].reg_size >= size)
1560                         return pavail[i].phys_addr;
1561         }
1562
1563         return ~0UL;
1564 }
1565
1566 static void __init tsb_phys_patch(void)
1567 {
1568         struct tsb_ldquad_phys_patch_entry *pquad;
1569         struct tsb_phys_patch_entry *p;
1570
1571         pquad = &__tsb_ldquad_phys_patch;
1572         while (pquad < &__tsb_ldquad_phys_patch_end) {
1573                 unsigned long addr = pquad->addr;
1574
1575                 if (tlb_type == hypervisor)
1576                         *(unsigned int *) addr = pquad->sun4v_insn;
1577                 else
1578                         *(unsigned int *) addr = pquad->sun4u_insn;
1579                 wmb();
1580                 __asm__ __volatile__("flush     %0"
1581                                      : /* no outputs */
1582                                      : "r" (addr));
1583
1584                 pquad++;
1585         }
1586
1587         p = &__tsb_phys_patch;
1588         while (p < &__tsb_phys_patch_end) {
1589                 unsigned long addr = p->addr;
1590
1591                 *(unsigned int *) addr = p->insn;
1592                 wmb();
1593                 __asm__ __volatile__("flush     %0"
1594                                      : /* no outputs */
1595                                      : "r" (addr));
1596
1597                 p++;
1598         }
1599 }
1600
1601 /* Don't mark as init, we give this to the Hypervisor.  */
1602 #ifndef CONFIG_DEBUG_PAGEALLOC
1603 #define NUM_KTSB_DESCR  2
1604 #else
1605 #define NUM_KTSB_DESCR  1
1606 #endif
1607 static struct hv_tsb_descr ktsb_descr[NUM_KTSB_DESCR];
1608 extern struct tsb swapper_tsb[KERNEL_TSB_NENTRIES];
1609
1610 static void __init sun4v_ktsb_init(void)
1611 {
1612         unsigned long ktsb_pa;
1613
1614         /* First KTSB for PAGE_SIZE mappings.  */
1615         ktsb_pa = kern_base + ((unsigned long)&swapper_tsb[0] - KERNBASE);
1616
1617         switch (PAGE_SIZE) {
1618         case 8 * 1024:
1619         default:
1620                 ktsb_descr[0].pgsz_idx = HV_PGSZ_IDX_8K;
1621                 ktsb_descr[0].pgsz_mask = HV_PGSZ_MASK_8K;
1622                 break;
1623
1624         case 64 * 1024:
1625                 ktsb_descr[0].pgsz_idx = HV_PGSZ_IDX_64K;
1626                 ktsb_descr[0].pgsz_mask = HV_PGSZ_MASK_64K;
1627                 break;
1628
1629         case 512 * 1024:
1630                 ktsb_descr[0].pgsz_idx = HV_PGSZ_IDX_512K;
1631                 ktsb_descr[0].pgsz_mask = HV_PGSZ_MASK_512K;
1632                 break;
1633
1634         case 4 * 1024 * 1024:
1635                 ktsb_descr[0].pgsz_idx = HV_PGSZ_IDX_4MB;
1636                 ktsb_descr[0].pgsz_mask = HV_PGSZ_MASK_4MB;
1637                 break;
1638         };
1639
1640         ktsb_descr[0].assoc = 1;
1641         ktsb_descr[0].num_ttes = KERNEL_TSB_NENTRIES;
1642         ktsb_descr[0].ctx_idx = 0;
1643         ktsb_descr[0].tsb_base = ktsb_pa;
1644         ktsb_descr[0].resv = 0;
1645
1646 #ifndef CONFIG_DEBUG_PAGEALLOC
1647         /* Second KTSB for 4MB/256MB mappings.  */
1648         ktsb_pa = (kern_base +
1649                    ((unsigned long)&swapper_4m_tsb[0] - KERNBASE));
1650
1651         ktsb_descr[1].pgsz_idx = HV_PGSZ_IDX_4MB;
1652         ktsb_descr[1].pgsz_mask = (HV_PGSZ_MASK_4MB |
1653                                    HV_PGSZ_MASK_256MB);
1654         ktsb_descr[1].assoc = 1;
1655         ktsb_descr[1].num_ttes = KERNEL_TSB4M_NENTRIES;
1656         ktsb_descr[1].ctx_idx = 0;
1657         ktsb_descr[1].tsb_base = ktsb_pa;
1658         ktsb_descr[1].resv = 0;
1659 #endif
1660 }
1661
1662 void __cpuinit sun4v_ktsb_register(void)
1663 {
1664         unsigned long pa, ret;
1665
1666         pa = kern_base + ((unsigned long)&ktsb_descr[0] - KERNBASE);
1667
1668         ret = sun4v_mmu_tsb_ctx0(NUM_KTSB_DESCR, pa);
1669         if (ret != 0) {
1670                 prom_printf("hypervisor_mmu_tsb_ctx0[%lx]: "
1671                             "errors with %lx\n", pa, ret);
1672                 prom_halt();
1673         }
1674 }
1675
1676 /* paging_init() sets up the page tables */
1677
1678 static unsigned long last_valid_pfn;
1679 pgd_t swapper_pg_dir[2048];
1680
1681 static void sun4u_pgprot_init(void);
1682 static void sun4v_pgprot_init(void);
1683
1684 void __init paging_init(void)
1685 {
1686         unsigned long end_pfn, shift, phys_base;
1687         unsigned long real_end, i;
1688
1689         /* These build time checkes make sure that the dcache_dirty_cpu()
1690          * page->flags usage will work.
1691          *
1692          * When a page gets marked as dcache-dirty, we store the
1693          * cpu number starting at bit 32 in the page->flags.  Also,
1694          * functions like clear_dcache_dirty_cpu use the cpu mask
1695          * in 13-bit signed-immediate instruction fields.
1696          */
1697
1698         /*
1699          * Page flags must not reach into upper 32 bits that are used
1700          * for the cpu number
1701          */
1702         BUILD_BUG_ON(NR_PAGEFLAGS > 32);
1703
1704         /*
1705          * The bit fields placed in the high range must not reach below
1706          * the 32 bit boundary. Otherwise we cannot place the cpu field
1707          * at the 32 bit boundary.
1708          */
1709         BUILD_BUG_ON(SECTIONS_WIDTH + NODES_WIDTH + ZONES_WIDTH +
1710                 ilog2(roundup_pow_of_two(NR_CPUS)) > 32);
1711
1712         BUILD_BUG_ON(NR_CPUS > 4096);
1713
1714         kern_base = (prom_boot_mapping_phys_low >> 22UL) << 22UL;
1715         kern_size = (unsigned long)&_end - (unsigned long)KERNBASE;
1716
1717         /* Invalidate both kernel TSBs.  */
1718         memset(swapper_tsb, 0x40, sizeof(swapper_tsb));
1719 #ifndef CONFIG_DEBUG_PAGEALLOC
1720         memset(swapper_4m_tsb, 0x40, sizeof(swapper_4m_tsb));
1721 #endif
1722
1723         if (tlb_type == hypervisor)
1724                 sun4v_pgprot_init();
1725         else
1726                 sun4u_pgprot_init();
1727
1728         if (tlb_type == cheetah_plus ||
1729             tlb_type == hypervisor)
1730                 tsb_phys_patch();
1731
1732         if (tlb_type == hypervisor) {
1733                 sun4v_patch_tlb_handlers();
1734                 sun4v_ktsb_init();
1735         }
1736
1737         lmb_init();
1738
1739         /* Find available physical memory...
1740          *
1741          * Read it twice in order to work around a bug in openfirmware.
1742          * The call to grab this table itself can cause openfirmware to
1743          * allocate memory, which in turn can take away some space from
1744          * the list of available memory.  Reading it twice makes sure
1745          * we really do get the final value.
1746          */
1747         read_obp_translations();
1748         read_obp_memory("reg", &pall[0], &pall_ents);
1749         read_obp_memory("available", &pavail[0], &pavail_ents);
1750         read_obp_memory("available", &pavail[0], &pavail_ents);
1751
1752         phys_base = 0xffffffffffffffffUL;
1753         for (i = 0; i < pavail_ents; i++) {
1754                 phys_base = min(phys_base, pavail[i].phys_addr);
1755                 lmb_add(pavail[i].phys_addr, pavail[i].reg_size);
1756         }
1757
1758         lmb_reserve(kern_base, kern_size);
1759
1760         find_ramdisk(phys_base);
1761
1762         lmb_enforce_memory_limit(cmdline_memory_size);
1763
1764         lmb_analyze();
1765         lmb_dump_all();
1766
1767         set_bit(0, mmu_context_bmap);
1768
1769         shift = kern_base + PAGE_OFFSET - ((unsigned long)KERNBASE);
1770
1771         real_end = (unsigned long)_end;
1772         num_kernel_image_mappings = DIV_ROUND_UP(real_end - KERNBASE, 1 << 22);
1773         printk("Kernel: Using %d locked TLB entries for main kernel image.\n",
1774                num_kernel_image_mappings);
1775
1776         /* Set kernel pgd to upper alias so physical page computations
1777          * work.
1778          */
1779         init_mm.pgd += ((shift) / (sizeof(pgd_t)));
1780         
1781         memset(swapper_low_pmd_dir, 0, sizeof(swapper_low_pmd_dir));
1782
1783         /* Now can init the kernel/bad page tables. */
1784         pud_set(pud_offset(&swapper_pg_dir[0], 0),
1785                 swapper_low_pmd_dir + (shift / sizeof(pgd_t)));
1786         
1787         inherit_prom_mappings();
1788         
1789         init_kpte_bitmap();
1790
1791         /* Ok, we can use our TLB miss and window trap handlers safely.  */
1792         setup_tba();
1793
1794         __flush_tlb_all();
1795
1796         if (tlb_type == hypervisor)
1797                 sun4v_ktsb_register();
1798
1799         prom_build_devicetree();
1800         of_populate_present_mask();
1801 #ifndef CONFIG_SMP
1802         of_fill_in_cpu_data();
1803 #endif
1804
1805         if (tlb_type == hypervisor) {
1806                 sun4v_mdesc_init();
1807                 mdesc_populate_present_mask(cpu_all_mask);
1808 #ifndef CONFIG_SMP
1809                 mdesc_fill_in_cpu_data(cpu_all_mask);
1810 #endif
1811         }
1812
1813         /* Once the OF device tree and MDESC have been setup, we know
1814          * the list of possible cpus.  Therefore we can allocate the
1815          * IRQ stacks.
1816          */
1817         for_each_possible_cpu(i) {
1818                 /* XXX Use node local allocations... XXX */
1819                 softirq_stack[i] = __va(lmb_alloc(THREAD_SIZE, THREAD_SIZE));
1820                 hardirq_stack[i] = __va(lmb_alloc(THREAD_SIZE, THREAD_SIZE));
1821         }
1822
1823         /* Setup bootmem... */
1824         last_valid_pfn = end_pfn = bootmem_init(phys_base);
1825
1826 #ifndef CONFIG_NEED_MULTIPLE_NODES
1827         max_mapnr = last_valid_pfn;
1828 #endif
1829         kernel_physical_mapping_init();
1830
1831         {
1832                 unsigned long max_zone_pfns[MAX_NR_ZONES];
1833
1834                 memset(max_zone_pfns, 0, sizeof(max_zone_pfns));
1835
1836                 max_zone_pfns[ZONE_NORMAL] = end_pfn;
1837
1838                 free_area_init_nodes(max_zone_pfns);
1839         }
1840
1841         printk("Booting Linux...\n");
1842 }
1843
1844 int __devinit page_in_phys_avail(unsigned long paddr)
1845 {
1846         int i;
1847
1848         paddr &= PAGE_MASK;
1849
1850         for (i = 0; i < pavail_ents; i++) {
1851                 unsigned long start, end;
1852
1853                 start = pavail[i].phys_addr;
1854                 end = start + pavail[i].reg_size;
1855
1856                 if (paddr >= start && paddr < end)
1857                         return 1;
1858         }
1859         if (paddr >= kern_base && paddr < (kern_base + kern_size))
1860                 return 1;
1861 #ifdef CONFIG_BLK_DEV_INITRD
1862         if (paddr >= __pa(initrd_start) &&
1863             paddr < __pa(PAGE_ALIGN(initrd_end)))
1864                 return 1;
1865 #endif
1866
1867         return 0;
1868 }
1869
1870 static struct linux_prom64_registers pavail_rescan[MAX_BANKS] __initdata;
1871 static int pavail_rescan_ents __initdata;
1872
1873 /* Certain OBP calls, such as fetching "available" properties, can
1874  * claim physical memory.  So, along with initializing the valid
1875  * address bitmap, what we do here is refetch the physical available
1876  * memory list again, and make sure it provides at least as much
1877  * memory as 'pavail' does.
1878  */
1879 static void __init setup_valid_addr_bitmap_from_pavail(unsigned long *bitmap)
1880 {
1881         int i;
1882
1883         read_obp_memory("available", &pavail_rescan[0], &pavail_rescan_ents);
1884
1885         for (i = 0; i < pavail_ents; i++) {
1886                 unsigned long old_start, old_end;
1887
1888                 old_start = pavail[i].phys_addr;
1889                 old_end = old_start + pavail[i].reg_size;
1890                 while (old_start < old_end) {
1891                         int n;
1892
1893                         for (n = 0; n < pavail_rescan_ents; n++) {
1894                                 unsigned long new_start, new_end;
1895
1896                                 new_start = pavail_rescan[n].phys_addr;
1897                                 new_end = new_start +
1898                                         pavail_rescan[n].reg_size;
1899
1900                                 if (new_start <= old_start &&
1901                                     new_end >= (old_start + PAGE_SIZE)) {
1902                                         set_bit(old_start >> 22, bitmap);
1903                                         goto do_next_page;
1904                                 }
1905                         }
1906
1907                         prom_printf("mem_init: Lost memory in pavail\n");
1908                         prom_printf("mem_init: OLD start[%lx] size[%lx]\n",
1909                                     pavail[i].phys_addr,
1910                                     pavail[i].reg_size);
1911                         prom_printf("mem_init: NEW start[%lx] size[%lx]\n",
1912                                     pavail_rescan[i].phys_addr,
1913                                     pavail_rescan[i].reg_size);
1914                         prom_printf("mem_init: Cannot continue, aborting.\n");
1915                         prom_halt();
1916
1917                 do_next_page:
1918                         old_start += PAGE_SIZE;
1919                 }
1920         }
1921 }
1922
1923 static void __init patch_tlb_miss_handler_bitmap(void)
1924 {
1925         extern unsigned int valid_addr_bitmap_insn[];
1926         extern unsigned int valid_addr_bitmap_patch[];
1927
1928         valid_addr_bitmap_insn[1] = valid_addr_bitmap_patch[1];
1929         mb();
1930         valid_addr_bitmap_insn[0] = valid_addr_bitmap_patch[0];
1931         flushi(&valid_addr_bitmap_insn[0]);
1932 }
1933
1934 void __init mem_init(void)
1935 {
1936         unsigned long codepages, datapages, initpages;
1937         unsigned long addr, last;
1938
1939         addr = PAGE_OFFSET + kern_base;
1940         last = PAGE_ALIGN(kern_size) + addr;
1941         while (addr < last) {
1942                 set_bit(__pa(addr) >> 22, sparc64_valid_addr_bitmap);
1943                 addr += PAGE_SIZE;
1944         }
1945
1946         setup_valid_addr_bitmap_from_pavail(sparc64_valid_addr_bitmap);
1947         patch_tlb_miss_handler_bitmap();
1948
1949         high_memory = __va(last_valid_pfn << PAGE_SHIFT);
1950
1951 #ifdef CONFIG_NEED_MULTIPLE_NODES
1952         {
1953                 int i;
1954                 for_each_online_node(i) {
1955                         if (NODE_DATA(i)->node_spanned_pages != 0) {
1956                                 totalram_pages +=
1957                                         free_all_bootmem_node(NODE_DATA(i));
1958                         }
1959                 }
1960         }
1961 #else
1962         totalram_pages = free_all_bootmem();
1963 #endif
1964
1965         /* We subtract one to account for the mem_map_zero page
1966          * allocated below.
1967          */
1968         totalram_pages -= 1;
1969         num_physpages = totalram_pages;
1970
1971         /*
1972          * Set up the zero page, mark it reserved, so that page count
1973          * is not manipulated when freeing the page from user ptes.
1974          */
1975         mem_map_zero = alloc_pages(GFP_KERNEL|__GFP_ZERO, 0);
1976         if (mem_map_zero == NULL) {
1977                 prom_printf("paging_init: Cannot alloc zero page.\n");
1978                 prom_halt();
1979         }
1980         SetPageReserved(mem_map_zero);
1981
1982         codepages = (((unsigned long) _etext) - ((unsigned long) _start));
1983         codepages = PAGE_ALIGN(codepages) >> PAGE_SHIFT;
1984         datapages = (((unsigned long) _edata) - ((unsigned long) _etext));
1985         datapages = PAGE_ALIGN(datapages) >> PAGE_SHIFT;
1986         initpages = (((unsigned long) __init_end) - ((unsigned long) __init_begin));
1987         initpages = PAGE_ALIGN(initpages) >> PAGE_SHIFT;
1988
1989         printk("Memory: %luk available (%ldk kernel code, %ldk data, %ldk init) [%016lx,%016lx]\n",
1990                nr_free_pages() << (PAGE_SHIFT-10),
1991                codepages << (PAGE_SHIFT-10),
1992                datapages << (PAGE_SHIFT-10), 
1993                initpages << (PAGE_SHIFT-10), 
1994                PAGE_OFFSET, (last_valid_pfn << PAGE_SHIFT));
1995
1996         if (tlb_type == cheetah || tlb_type == cheetah_plus)
1997                 cheetah_ecache_flush_init();
1998 }
1999
2000 void free_initmem(void)
2001 {
2002         unsigned long addr, initend;
2003         int do_free = 1;
2004
2005         /* If the physical memory maps were trimmed by kernel command
2006          * line options, don't even try freeing this initmem stuff up.
2007          * The kernel image could have been in the trimmed out region
2008          * and if so the freeing below will free invalid page structs.
2009          */
2010         if (cmdline_memory_size)
2011                 do_free = 0;
2012
2013         /*
2014          * The init section is aligned to 8k in vmlinux.lds. Page align for >8k pagesizes.
2015          */
2016         addr = PAGE_ALIGN((unsigned long)(__init_begin));
2017         initend = (unsigned long)(__init_end) & PAGE_MASK;
2018         for (; addr < initend; addr += PAGE_SIZE) {
2019                 unsigned long page;
2020                 struct page *p;
2021
2022                 page = (addr +
2023                         ((unsigned long) __va(kern_base)) -
2024                         ((unsigned long) KERNBASE));
2025                 memset((void *)addr, POISON_FREE_INITMEM, PAGE_SIZE);
2026
2027                 if (do_free) {
2028                         p = virt_to_page(page);
2029
2030                         ClearPageReserved(p);
2031                         init_page_count(p);
2032                         __free_page(p);
2033                         num_physpages++;
2034                         totalram_pages++;
2035                 }
2036         }
2037 }
2038
2039 #ifdef CONFIG_BLK_DEV_INITRD
2040 void free_initrd_mem(unsigned long start, unsigned long end)
2041 {
2042         if (start < end)
2043                 printk ("Freeing initrd memory: %ldk freed\n", (end - start) >> 10);
2044         for (; start < end; start += PAGE_SIZE) {
2045                 struct page *p = virt_to_page(start);
2046
2047                 ClearPageReserved(p);
2048                 init_page_count(p);
2049                 __free_page(p);
2050                 num_physpages++;
2051                 totalram_pages++;
2052         }
2053 }
2054 #endif
2055
2056 #define _PAGE_CACHE_4U  (_PAGE_CP_4U | _PAGE_CV_4U)
2057 #define _PAGE_CACHE_4V  (_PAGE_CP_4V | _PAGE_CV_4V)
2058 #define __DIRTY_BITS_4U  (_PAGE_MODIFIED_4U | _PAGE_WRITE_4U | _PAGE_W_4U)
2059 #define __DIRTY_BITS_4V  (_PAGE_MODIFIED_4V | _PAGE_WRITE_4V | _PAGE_W_4V)
2060 #define __ACCESS_BITS_4U (_PAGE_ACCESSED_4U | _PAGE_READ_4U | _PAGE_R)
2061 #define __ACCESS_BITS_4V (_PAGE_ACCESSED_4V | _PAGE_READ_4V | _PAGE_R)
2062
2063 pgprot_t PAGE_KERNEL __read_mostly;
2064 EXPORT_SYMBOL(PAGE_KERNEL);
2065
2066 pgprot_t PAGE_KERNEL_LOCKED __read_mostly;
2067 pgprot_t PAGE_COPY __read_mostly;
2068
2069 pgprot_t PAGE_SHARED __read_mostly;
2070 EXPORT_SYMBOL(PAGE_SHARED);
2071
2072 unsigned long pg_iobits __read_mostly;
2073
2074 unsigned long _PAGE_IE __read_mostly;
2075 EXPORT_SYMBOL(_PAGE_IE);
2076
2077 unsigned long _PAGE_E __read_mostly;
2078 EXPORT_SYMBOL(_PAGE_E);
2079
2080 unsigned long _PAGE_CACHE __read_mostly;
2081 EXPORT_SYMBOL(_PAGE_CACHE);
2082
2083 #ifdef CONFIG_SPARSEMEM_VMEMMAP
2084 unsigned long vmemmap_table[VMEMMAP_SIZE];
2085
2086 int __meminit vmemmap_populate(struct page *start, unsigned long nr, int node)
2087 {
2088         unsigned long vstart = (unsigned long) start;
2089         unsigned long vend = (unsigned long) (start + nr);
2090         unsigned long phys_start = (vstart - VMEMMAP_BASE);
2091         unsigned long phys_end = (vend - VMEMMAP_BASE);
2092         unsigned long addr = phys_start & VMEMMAP_CHUNK_MASK;
2093         unsigned long end = VMEMMAP_ALIGN(phys_end);
2094         unsigned long pte_base;
2095
2096         pte_base = (_PAGE_VALID | _PAGE_SZ4MB_4U |
2097                     _PAGE_CP_4U | _PAGE_CV_4U |
2098                     _PAGE_P_4U | _PAGE_W_4U);
2099         if (tlb_type == hypervisor)
2100                 pte_base = (_PAGE_VALID | _PAGE_SZ4MB_4V |
2101                             _PAGE_CP_4V | _PAGE_CV_4V |
2102                             _PAGE_P_4V | _PAGE_W_4V);
2103
2104         for (; addr < end; addr += VMEMMAP_CHUNK) {
2105                 unsigned long *vmem_pp =
2106                         vmemmap_table + (addr >> VMEMMAP_CHUNK_SHIFT);
2107                 void *block;
2108
2109                 if (!(*vmem_pp & _PAGE_VALID)) {
2110                         block = vmemmap_alloc_block(1UL << 22, node);
2111                         if (!block)
2112                                 return -ENOMEM;
2113
2114                         *vmem_pp = pte_base | __pa(block);
2115
2116                         printk(KERN_INFO "[%p-%p] page_structs=%lu "
2117                                "node=%d entry=%lu/%lu\n", start, block, nr,
2118                                node,
2119                                addr >> VMEMMAP_CHUNK_SHIFT,
2120                                VMEMMAP_SIZE);
2121                 }
2122         }
2123         return 0;
2124 }
2125 #endif /* CONFIG_SPARSEMEM_VMEMMAP */
2126
2127 static void prot_init_common(unsigned long page_none,
2128                              unsigned long page_shared,
2129                              unsigned long page_copy,
2130                              unsigned long page_readonly,
2131                              unsigned long page_exec_bit)
2132 {
2133         PAGE_COPY = __pgprot(page_copy);
2134         PAGE_SHARED = __pgprot(page_shared);
2135
2136         protection_map[0x0] = __pgprot(page_none);
2137         protection_map[0x1] = __pgprot(page_readonly & ~page_exec_bit);
2138         protection_map[0x2] = __pgprot(page_copy & ~page_exec_bit);
2139         protection_map[0x3] = __pgprot(page_copy & ~page_exec_bit);
2140         protection_map[0x4] = __pgprot(page_readonly);
2141         protection_map[0x5] = __pgprot(page_readonly);
2142         protection_map[0x6] = __pgprot(page_copy);
2143         protection_map[0x7] = __pgprot(page_copy);
2144         protection_map[0x8] = __pgprot(page_none);
2145         protection_map[0x9] = __pgprot(page_readonly & ~page_exec_bit);
2146         protection_map[0xa] = __pgprot(page_shared & ~page_exec_bit);
2147         protection_map[0xb] = __pgprot(page_shared & ~page_exec_bit);
2148         protection_map[0xc] = __pgprot(page_readonly);
2149         protection_map[0xd] = __pgprot(page_readonly);
2150         protection_map[0xe] = __pgprot(page_shared);
2151         protection_map[0xf] = __pgprot(page_shared);
2152 }
2153
2154 static void __init sun4u_pgprot_init(void)
2155 {
2156         unsigned long page_none, page_shared, page_copy, page_readonly;
2157         unsigned long page_exec_bit;
2158
2159         PAGE_KERNEL = __pgprot (_PAGE_PRESENT_4U | _PAGE_VALID |
2160                                 _PAGE_CACHE_4U | _PAGE_P_4U |
2161                                 __ACCESS_BITS_4U | __DIRTY_BITS_4U |
2162                                 _PAGE_EXEC_4U);
2163         PAGE_KERNEL_LOCKED = __pgprot (_PAGE_PRESENT_4U | _PAGE_VALID |
2164                                        _PAGE_CACHE_4U | _PAGE_P_4U |
2165                                        __ACCESS_BITS_4U | __DIRTY_BITS_4U |
2166                                        _PAGE_EXEC_4U | _PAGE_L_4U);
2167
2168         _PAGE_IE = _PAGE_IE_4U;
2169         _PAGE_E = _PAGE_E_4U;
2170         _PAGE_CACHE = _PAGE_CACHE_4U;
2171
2172         pg_iobits = (_PAGE_VALID | _PAGE_PRESENT_4U | __DIRTY_BITS_4U |
2173                      __ACCESS_BITS_4U | _PAGE_E_4U);
2174
2175 #ifdef CONFIG_DEBUG_PAGEALLOC
2176         kern_linear_pte_xor[0] = (_PAGE_VALID | _PAGE_SZBITS_4U) ^
2177                 0xfffff80000000000UL;
2178 #else
2179         kern_linear_pte_xor[0] = (_PAGE_VALID | _PAGE_SZ4MB_4U) ^
2180                 0xfffff80000000000UL;
2181 #endif
2182         kern_linear_pte_xor[0] |= (_PAGE_CP_4U | _PAGE_CV_4U |
2183                                    _PAGE_P_4U | _PAGE_W_4U);
2184
2185         /* XXX Should use 256MB on Panther. XXX */
2186         kern_linear_pte_xor[1] = kern_linear_pte_xor[0];
2187
2188         _PAGE_SZBITS = _PAGE_SZBITS_4U;
2189         _PAGE_ALL_SZ_BITS =  (_PAGE_SZ4MB_4U | _PAGE_SZ512K_4U |
2190                               _PAGE_SZ64K_4U | _PAGE_SZ8K_4U |
2191                               _PAGE_SZ32MB_4U | _PAGE_SZ256MB_4U);
2192
2193
2194         page_none = _PAGE_PRESENT_4U | _PAGE_ACCESSED_4U | _PAGE_CACHE_4U;
2195         page_shared = (_PAGE_VALID | _PAGE_PRESENT_4U | _PAGE_CACHE_4U |
2196                        __ACCESS_BITS_4U | _PAGE_WRITE_4U | _PAGE_EXEC_4U);
2197         page_copy   = (_PAGE_VALID | _PAGE_PRESENT_4U | _PAGE_CACHE_4U |
2198                        __ACCESS_BITS_4U | _PAGE_EXEC_4U);
2199         page_readonly   = (_PAGE_VALID | _PAGE_PRESENT_4U | _PAGE_CACHE_4U |
2200                            __ACCESS_BITS_4U | _PAGE_EXEC_4U);
2201
2202         page_exec_bit = _PAGE_EXEC_4U;
2203
2204         prot_init_common(page_none, page_shared, page_copy, page_readonly,
2205                          page_exec_bit);
2206 }
2207
2208 static void __init sun4v_pgprot_init(void)
2209 {
2210         unsigned long page_none, page_shared, page_copy, page_readonly;
2211         unsigned long page_exec_bit;
2212
2213         PAGE_KERNEL = __pgprot (_PAGE_PRESENT_4V | _PAGE_VALID |
2214                                 _PAGE_CACHE_4V | _PAGE_P_4V |
2215                                 __ACCESS_BITS_4V | __DIRTY_BITS_4V |
2216                                 _PAGE_EXEC_4V);
2217         PAGE_KERNEL_LOCKED = PAGE_KERNEL;
2218
2219         _PAGE_IE = _PAGE_IE_4V;
2220         _PAGE_E = _PAGE_E_4V;
2221         _PAGE_CACHE = _PAGE_CACHE_4V;
2222
2223 #ifdef CONFIG_DEBUG_PAGEALLOC
2224         kern_linear_pte_xor[0] = (_PAGE_VALID | _PAGE_SZBITS_4V) ^
2225                 0xfffff80000000000UL;
2226 #else
2227         kern_linear_pte_xor[0] = (_PAGE_VALID | _PAGE_SZ4MB_4V) ^
2228                 0xfffff80000000000UL;
2229 #endif
2230         kern_linear_pte_xor[0] |= (_PAGE_CP_4V | _PAGE_CV_4V |
2231                                    _PAGE_P_4V | _PAGE_W_4V);
2232
2233 #ifdef CONFIG_DEBUG_PAGEALLOC
2234         kern_linear_pte_xor[1] = (_PAGE_VALID | _PAGE_SZBITS_4V) ^
2235                 0xfffff80000000000UL;
2236 #else
2237         kern_linear_pte_xor[1] = (_PAGE_VALID | _PAGE_SZ256MB_4V) ^
2238                 0xfffff80000000000UL;
2239 #endif
2240         kern_linear_pte_xor[1] |= (_PAGE_CP_4V | _PAGE_CV_4V |
2241                                    _PAGE_P_4V | _PAGE_W_4V);
2242
2243         pg_iobits = (_PAGE_VALID | _PAGE_PRESENT_4V | __DIRTY_BITS_4V |
2244                      __ACCESS_BITS_4V | _PAGE_E_4V);
2245
2246         _PAGE_SZBITS = _PAGE_SZBITS_4V;
2247         _PAGE_ALL_SZ_BITS = (_PAGE_SZ16GB_4V | _PAGE_SZ2GB_4V |
2248                              _PAGE_SZ256MB_4V | _PAGE_SZ32MB_4V |
2249                              _PAGE_SZ4MB_4V | _PAGE_SZ512K_4V |
2250                              _PAGE_SZ64K_4V | _PAGE_SZ8K_4V);
2251
2252         page_none = _PAGE_PRESENT_4V | _PAGE_ACCESSED_4V | _PAGE_CACHE_4V;
2253         page_shared = (_PAGE_VALID | _PAGE_PRESENT_4V | _PAGE_CACHE_4V |
2254                        __ACCESS_BITS_4V | _PAGE_WRITE_4V | _PAGE_EXEC_4V);
2255         page_copy   = (_PAGE_VALID | _PAGE_PRESENT_4V | _PAGE_CACHE_4V |
2256                        __ACCESS_BITS_4V | _PAGE_EXEC_4V);
2257         page_readonly = (_PAGE_VALID | _PAGE_PRESENT_4V | _PAGE_CACHE_4V |
2258                          __ACCESS_BITS_4V | _PAGE_EXEC_4V);
2259
2260         page_exec_bit = _PAGE_EXEC_4V;
2261
2262         prot_init_common(page_none, page_shared, page_copy, page_readonly,
2263                          page_exec_bit);
2264 }
2265
2266 unsigned long pte_sz_bits(unsigned long sz)
2267 {
2268         if (tlb_type == hypervisor) {
2269                 switch (sz) {
2270                 case 8 * 1024:
2271                 default:
2272                         return _PAGE_SZ8K_4V;
2273                 case 64 * 1024:
2274                         return _PAGE_SZ64K_4V;
2275                 case 512 * 1024:
2276                         return _PAGE_SZ512K_4V;
2277                 case 4 * 1024 * 1024:
2278                         return _PAGE_SZ4MB_4V;
2279                 };
2280         } else {
2281                 switch (sz) {
2282                 case 8 * 1024:
2283                 default:
2284                         return _PAGE_SZ8K_4U;
2285                 case 64 * 1024:
2286                         return _PAGE_SZ64K_4U;
2287                 case 512 * 1024:
2288                         return _PAGE_SZ512K_4U;
2289                 case 4 * 1024 * 1024:
2290                         return _PAGE_SZ4MB_4U;
2291                 };
2292         }
2293 }
2294
2295 pte_t mk_pte_io(unsigned long page, pgprot_t prot, int space, unsigned long page_size)
2296 {
2297         pte_t pte;
2298
2299         pte_val(pte)  = page | pgprot_val(pgprot_noncached(prot));
2300         pte_val(pte) |= (((unsigned long)space) << 32);
2301         pte_val(pte) |= pte_sz_bits(page_size);
2302
2303         return pte;
2304 }
2305
2306 static unsigned long kern_large_tte(unsigned long paddr)
2307 {
2308         unsigned long val;
2309
2310         val = (_PAGE_VALID | _PAGE_SZ4MB_4U |
2311                _PAGE_CP_4U | _PAGE_CV_4U | _PAGE_P_4U |
2312                _PAGE_EXEC_4U | _PAGE_L_4U | _PAGE_W_4U);
2313         if (tlb_type == hypervisor)
2314                 val = (_PAGE_VALID | _PAGE_SZ4MB_4V |
2315                        _PAGE_CP_4V | _PAGE_CV_4V | _PAGE_P_4V |
2316                        _PAGE_EXEC_4V | _PAGE_W_4V);
2317
2318         return val | paddr;
2319 }
2320
2321 /* If not locked, zap it. */
2322 void __flush_tlb_all(void)
2323 {
2324         unsigned long pstate;
2325         int i;
2326
2327         __asm__ __volatile__("flushw\n\t"
2328                              "rdpr      %%pstate, %0\n\t"
2329                              "wrpr      %0, %1, %%pstate"
2330                              : "=r" (pstate)
2331                              : "i" (PSTATE_IE));
2332         if (tlb_type == hypervisor) {
2333                 sun4v_mmu_demap_all();
2334         } else if (tlb_type == spitfire) {
2335                 for (i = 0; i < 64; i++) {
2336                         /* Spitfire Errata #32 workaround */
2337                         /* NOTE: Always runs on spitfire, so no
2338                          *       cheetah+ page size encodings.
2339                          */
2340                         __asm__ __volatile__("stxa      %0, [%1] %2\n\t"
2341                                              "flush     %%g6"
2342                                              : /* No outputs */
2343                                              : "r" (0),
2344                                              "r" (PRIMARY_CONTEXT), "i" (ASI_DMMU));
2345
2346                         if (!(spitfire_get_dtlb_data(i) & _PAGE_L_4U)) {
2347                                 __asm__ __volatile__("stxa %%g0, [%0] %1\n\t"
2348                                                      "membar #Sync"
2349                                                      : /* no outputs */
2350                                                      : "r" (TLB_TAG_ACCESS), "i" (ASI_DMMU));
2351                                 spitfire_put_dtlb_data(i, 0x0UL);
2352                         }
2353
2354                         /* Spitfire Errata #32 workaround */
2355                         /* NOTE: Always runs on spitfire, so no
2356                          *       cheetah+ page size encodings.
2357                          */
2358                         __asm__ __volatile__("stxa      %0, [%1] %2\n\t"
2359                                              "flush     %%g6"
2360                                              : /* No outputs */
2361                                              : "r" (0),
2362                                              "r" (PRIMARY_CONTEXT), "i" (ASI_DMMU));
2363
2364                         if (!(spitfire_get_itlb_data(i) & _PAGE_L_4U)) {
2365                                 __asm__ __volatile__("stxa %%g0, [%0] %1\n\t"
2366                                                      "membar #Sync"
2367                                                      : /* no outputs */
2368                                                      : "r" (TLB_TAG_ACCESS), "i" (ASI_IMMU));
2369                                 spitfire_put_itlb_data(i, 0x0UL);
2370                         }
2371                 }
2372         } else if (tlb_type == cheetah || tlb_type == cheetah_plus) {
2373                 cheetah_flush_dtlb_all();
2374                 cheetah_flush_itlb_all();
2375         }
2376         __asm__ __volatile__("wrpr      %0, 0, %%pstate"
2377                              : : "r" (pstate));
2378 }