Merge branch 'merge' of git://git.kernel.org/pub/scm/linux/kernel/git/paulus/powerpc
[linux-2.6] / arch / sparc64 / mm / init.c
1 /*  $Id: init.c,v 1.209 2002/02/09 19:49:31 davem Exp $
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/slab.h>
17 #include <linux/initrd.h>
18 #include <linux/swap.h>
19 #include <linux/pagemap.h>
20 #include <linux/poison.h>
21 #include <linux/fs.h>
22 #include <linux/seq_file.h>
23 #include <linux/kprobes.h>
24 #include <linux/cache.h>
25 #include <linux/sort.h>
26 #include <linux/percpu.h>
27 #include <linux/lmb.h>
28 #include <linux/mmzone.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/sstate.h>
50 #include <asm/mdesc.h>
51 #include <asm/cpudata.h>
52
53 #define MAX_PHYS_ADDRESS        (1UL << 42UL)
54 #define KPTE_BITMAP_CHUNK_SZ    (256UL * 1024UL * 1024UL)
55 #define KPTE_BITMAP_BYTES       \
56         ((MAX_PHYS_ADDRESS / KPTE_BITMAP_CHUNK_SZ) / 8)
57
58 unsigned long kern_linear_pte_xor[2] __read_mostly;
59
60 /* A bitmap, one bit for every 256MB of physical memory.  If the bit
61  * is clear, we should use a 4MB page (via kern_linear_pte_xor[0]) else
62  * if set we should use a 256MB page (via kern_linear_pte_xor[1]).
63  */
64 unsigned long kpte_linear_bitmap[KPTE_BITMAP_BYTES / sizeof(unsigned long)];
65
66 #ifndef CONFIG_DEBUG_PAGEALLOC
67 /* A special kernel TSB for 4MB and 256MB linear mappings.
68  * Space is allocated for this right after the trap table
69  * in arch/sparc64/kernel/head.S
70  */
71 extern struct tsb swapper_4m_tsb[KERNEL_TSB4M_NENTRIES];
72 #endif
73
74 #define MAX_BANKS       32
75
76 static struct linux_prom64_registers pavail[MAX_BANKS] __initdata;
77 static int pavail_ents __initdata;
78
79 static int cmp_p64(const void *a, const void *b)
80 {
81         const struct linux_prom64_registers *x = a, *y = b;
82
83         if (x->phys_addr > y->phys_addr)
84                 return 1;
85         if (x->phys_addr < y->phys_addr)
86                 return -1;
87         return 0;
88 }
89
90 static void __init read_obp_memory(const char *property,
91                                    struct linux_prom64_registers *regs,
92                                    int *num_ents)
93 {
94         int node = prom_finddevice("/memory");
95         int prop_size = prom_getproplen(node, property);
96         int ents, ret, i;
97
98         ents = prop_size / sizeof(struct linux_prom64_registers);
99         if (ents > MAX_BANKS) {
100                 prom_printf("The machine has more %s property entries than "
101                             "this kernel can support (%d).\n",
102                             property, MAX_BANKS);
103                 prom_halt();
104         }
105
106         ret = prom_getproperty(node, property, (char *) regs, prop_size);
107         if (ret == -1) {
108                 prom_printf("Couldn't get %s property from /memory.\n");
109                 prom_halt();
110         }
111
112         /* Sanitize what we got from the firmware, by page aligning
113          * everything.
114          */
115         for (i = 0; i < ents; i++) {
116                 unsigned long base, size;
117
118                 base = regs[i].phys_addr;
119                 size = regs[i].reg_size;
120
121                 size &= PAGE_MASK;
122                 if (base & ~PAGE_MASK) {
123                         unsigned long new_base = PAGE_ALIGN(base);
124
125                         size -= new_base - base;
126                         if ((long) size < 0L)
127                                 size = 0UL;
128                         base = new_base;
129                 }
130                 if (size == 0UL) {
131                         /* If it is empty, simply get rid of it.
132                          * This simplifies the logic of the other
133                          * functions that process these arrays.
134                          */
135                         memmove(&regs[i], &regs[i + 1],
136                                 (ents - i - 1) * sizeof(regs[0]));
137                         i--;
138                         ents--;
139                         continue;
140                 }
141                 regs[i].phys_addr = base;
142                 regs[i].reg_size = size;
143         }
144
145         *num_ents = ents;
146
147         sort(regs, ents, sizeof(struct linux_prom64_registers),
148              cmp_p64, NULL);
149 }
150
151 unsigned long *sparc64_valid_addr_bitmap __read_mostly;
152
153 /* Kernel physical address base and size in bytes.  */
154 unsigned long kern_base __read_mostly;
155 unsigned long kern_size __read_mostly;
156
157 /* Initial ramdisk setup */
158 extern unsigned long sparc_ramdisk_image64;
159 extern unsigned int sparc_ramdisk_image;
160 extern unsigned int sparc_ramdisk_size;
161
162 struct page *mem_map_zero __read_mostly;
163 EXPORT_SYMBOL(mem_map_zero);
164
165 unsigned int sparc64_highest_unlocked_tlb_ent __read_mostly;
166
167 unsigned long sparc64_kern_pri_context __read_mostly;
168 unsigned long sparc64_kern_pri_nuc_bits __read_mostly;
169 unsigned long sparc64_kern_sec_context __read_mostly;
170
171 int num_kernel_image_mappings;
172
173 #ifdef CONFIG_DEBUG_DCFLUSH
174 atomic_t dcpage_flushes = ATOMIC_INIT(0);
175 #ifdef CONFIG_SMP
176 atomic_t dcpage_flushes_xcall = ATOMIC_INIT(0);
177 #endif
178 #endif
179
180 inline void flush_dcache_page_impl(struct page *page)
181 {
182         BUG_ON(tlb_type == hypervisor);
183 #ifdef CONFIG_DEBUG_DCFLUSH
184         atomic_inc(&dcpage_flushes);
185 #endif
186
187 #ifdef DCACHE_ALIASING_POSSIBLE
188         __flush_dcache_page(page_address(page),
189                             ((tlb_type == spitfire) &&
190                              page_mapping(page) != NULL));
191 #else
192         if (page_mapping(page) != NULL &&
193             tlb_type == spitfire)
194                 __flush_icache_page(__pa(page_address(page)));
195 #endif
196 }
197
198 #define PG_dcache_dirty         PG_arch_1
199 #define PG_dcache_cpu_shift     32UL
200 #define PG_dcache_cpu_mask      \
201         ((1UL<<ilog2(roundup_pow_of_two(NR_CPUS)))-1UL)
202
203 #define dcache_dirty_cpu(page) \
204         (((page)->flags >> PG_dcache_cpu_shift) & PG_dcache_cpu_mask)
205
206 static inline void set_dcache_dirty(struct page *page, int this_cpu)
207 {
208         unsigned long mask = this_cpu;
209         unsigned long non_cpu_bits;
210
211         non_cpu_bits = ~(PG_dcache_cpu_mask << PG_dcache_cpu_shift);
212         mask = (mask << PG_dcache_cpu_shift) | (1UL << PG_dcache_dirty);
213
214         __asm__ __volatile__("1:\n\t"
215                              "ldx       [%2], %%g7\n\t"
216                              "and       %%g7, %1, %%g1\n\t"
217                              "or        %%g1, %0, %%g1\n\t"
218                              "casx      [%2], %%g7, %%g1\n\t"
219                              "cmp       %%g7, %%g1\n\t"
220                              "membar    #StoreLoad | #StoreStore\n\t"
221                              "bne,pn    %%xcc, 1b\n\t"
222                              " nop"
223                              : /* no outputs */
224                              : "r" (mask), "r" (non_cpu_bits), "r" (&page->flags)
225                              : "g1", "g7");
226 }
227
228 static inline void clear_dcache_dirty_cpu(struct page *page, unsigned long cpu)
229 {
230         unsigned long mask = (1UL << PG_dcache_dirty);
231
232         __asm__ __volatile__("! test_and_clear_dcache_dirty\n"
233                              "1:\n\t"
234                              "ldx       [%2], %%g7\n\t"
235                              "srlx      %%g7, %4, %%g1\n\t"
236                              "and       %%g1, %3, %%g1\n\t"
237                              "cmp       %%g1, %0\n\t"
238                              "bne,pn    %%icc, 2f\n\t"
239                              " andn     %%g7, %1, %%g1\n\t"
240                              "casx      [%2], %%g7, %%g1\n\t"
241                              "cmp       %%g7, %%g1\n\t"
242                              "membar    #StoreLoad | #StoreStore\n\t"
243                              "bne,pn    %%xcc, 1b\n\t"
244                              " nop\n"
245                              "2:"
246                              : /* no outputs */
247                              : "r" (cpu), "r" (mask), "r" (&page->flags),
248                                "i" (PG_dcache_cpu_mask),
249                                "i" (PG_dcache_cpu_shift)
250                              : "g1", "g7");
251 }
252
253 static inline void tsb_insert(struct tsb *ent, unsigned long tag, unsigned long pte)
254 {
255         unsigned long tsb_addr = (unsigned long) ent;
256
257         if (tlb_type == cheetah_plus || tlb_type == hypervisor)
258                 tsb_addr = __pa(tsb_addr);
259
260         __tsb_insert(tsb_addr, tag, pte);
261 }
262
263 unsigned long _PAGE_ALL_SZ_BITS __read_mostly;
264 unsigned long _PAGE_SZBITS __read_mostly;
265
266 void update_mmu_cache(struct vm_area_struct *vma, unsigned long address, pte_t pte)
267 {
268         struct mm_struct *mm;
269         struct tsb *tsb;
270         unsigned long tag, flags;
271         unsigned long tsb_index, tsb_hash_shift;
272
273         if (tlb_type != hypervisor) {
274                 unsigned long pfn = pte_pfn(pte);
275                 unsigned long pg_flags;
276                 struct page *page;
277
278                 if (pfn_valid(pfn) &&
279                     (page = pfn_to_page(pfn), page_mapping(page)) &&
280                     ((pg_flags = page->flags) & (1UL << PG_dcache_dirty))) {
281                         int cpu = ((pg_flags >> PG_dcache_cpu_shift) &
282                                    PG_dcache_cpu_mask);
283                         int this_cpu = get_cpu();
284
285                         /* This is just to optimize away some function calls
286                          * in the SMP case.
287                          */
288                         if (cpu == this_cpu)
289                                 flush_dcache_page_impl(page);
290                         else
291                                 smp_flush_dcache_page_impl(page, cpu);
292
293                         clear_dcache_dirty_cpu(page, cpu);
294
295                         put_cpu();
296                 }
297         }
298
299         mm = vma->vm_mm;
300
301         tsb_index = MM_TSB_BASE;
302         tsb_hash_shift = PAGE_SHIFT;
303
304         spin_lock_irqsave(&mm->context.lock, flags);
305
306 #ifdef CONFIG_HUGETLB_PAGE
307         if (mm->context.tsb_block[MM_TSB_HUGE].tsb != NULL) {
308                 if ((tlb_type == hypervisor &&
309                      (pte_val(pte) & _PAGE_SZALL_4V) == _PAGE_SZHUGE_4V) ||
310                     (tlb_type != hypervisor &&
311                      (pte_val(pte) & _PAGE_SZALL_4U) == _PAGE_SZHUGE_4U)) {
312                         tsb_index = MM_TSB_HUGE;
313                         tsb_hash_shift = HPAGE_SHIFT;
314                 }
315         }
316 #endif
317
318         tsb = mm->context.tsb_block[tsb_index].tsb;
319         tsb += ((address >> tsb_hash_shift) &
320                 (mm->context.tsb_block[tsb_index].tsb_nentries - 1UL));
321         tag = (address >> 22UL);
322         tsb_insert(tsb, tag, pte_val(pte));
323
324         spin_unlock_irqrestore(&mm->context.lock, flags);
325 }
326
327 void flush_dcache_page(struct page *page)
328 {
329         struct address_space *mapping;
330         int this_cpu;
331
332         if (tlb_type == hypervisor)
333                 return;
334
335         /* Do not bother with the expensive D-cache flush if it
336          * is merely the zero page.  The 'bigcore' testcase in GDB
337          * causes this case to run millions of times.
338          */
339         if (page == ZERO_PAGE(0))
340                 return;
341
342         this_cpu = get_cpu();
343
344         mapping = page_mapping(page);
345         if (mapping && !mapping_mapped(mapping)) {
346                 int dirty = test_bit(PG_dcache_dirty, &page->flags);
347                 if (dirty) {
348                         int dirty_cpu = dcache_dirty_cpu(page);
349
350                         if (dirty_cpu == this_cpu)
351                                 goto out;
352                         smp_flush_dcache_page_impl(page, dirty_cpu);
353                 }
354                 set_dcache_dirty(page, this_cpu);
355         } else {
356                 /* We could delay the flush for the !page_mapping
357                  * case too.  But that case is for exec env/arg
358                  * pages and those are %99 certainly going to get
359                  * faulted into the tlb (and thus flushed) anyways.
360                  */
361                 flush_dcache_page_impl(page);
362         }
363
364 out:
365         put_cpu();
366 }
367
368 void __kprobes flush_icache_range(unsigned long start, unsigned long end)
369 {
370         /* Cheetah and Hypervisor platform cpus have coherent I-cache. */
371         if (tlb_type == spitfire) {
372                 unsigned long kaddr;
373
374                 /* This code only runs on Spitfire cpus so this is
375                  * why we can assume _PAGE_PADDR_4U.
376                  */
377                 for (kaddr = start; kaddr < end; kaddr += PAGE_SIZE) {
378                         unsigned long paddr, mask = _PAGE_PADDR_4U;
379
380                         if (kaddr >= PAGE_OFFSET)
381                                 paddr = kaddr & mask;
382                         else {
383                                 pgd_t *pgdp = pgd_offset_k(kaddr);
384                                 pud_t *pudp = pud_offset(pgdp, kaddr);
385                                 pmd_t *pmdp = pmd_offset(pudp, kaddr);
386                                 pte_t *ptep = pte_offset_kernel(pmdp, kaddr);
387
388                                 paddr = pte_val(*ptep) & mask;
389                         }
390                         __flush_icache_page(paddr);
391                 }
392         }
393 }
394
395 void show_mem(void)
396 {
397         unsigned long total = 0, reserved = 0;
398         unsigned long shared = 0, cached = 0;
399         pg_data_t *pgdat;
400
401         printk(KERN_INFO "Mem-info:\n");
402         show_free_areas();
403         printk(KERN_INFO "Free swap:       %6ldkB\n",
404                nr_swap_pages << (PAGE_SHIFT-10));
405         for_each_online_pgdat(pgdat) {
406                 unsigned long i, flags;
407
408                 pgdat_resize_lock(pgdat, &flags);
409                 for (i = 0; i < pgdat->node_spanned_pages; i++) {
410                         struct page *page = pgdat_page_nr(pgdat, i);
411                         total++;
412                         if (PageReserved(page))
413                                 reserved++;
414                         else if (PageSwapCache(page))
415                                 cached++;
416                         else if (page_count(page))
417                                 shared += page_count(page) - 1;
418                 }
419                 pgdat_resize_unlock(pgdat, &flags);
420         }
421
422         printk(KERN_INFO "%lu pages of RAM\n", total);
423         printk(KERN_INFO "%lu reserved pages\n", reserved);
424         printk(KERN_INFO "%lu pages shared\n", shared);
425         printk(KERN_INFO "%lu pages swap cached\n", cached);
426
427         printk(KERN_INFO "%lu pages dirty\n",
428                global_page_state(NR_FILE_DIRTY));
429         printk(KERN_INFO "%lu pages writeback\n",
430                global_page_state(NR_WRITEBACK));
431         printk(KERN_INFO "%lu pages mapped\n",
432                global_page_state(NR_FILE_MAPPED));
433         printk(KERN_INFO "%lu pages slab\n",
434                 global_page_state(NR_SLAB_RECLAIMABLE) +
435                 global_page_state(NR_SLAB_UNRECLAIMABLE));
436         printk(KERN_INFO "%lu pages pagetables\n",
437                global_page_state(NR_PAGETABLE));
438 }
439
440 void mmu_info(struct seq_file *m)
441 {
442         if (tlb_type == cheetah)
443                 seq_printf(m, "MMU Type\t: Cheetah\n");
444         else if (tlb_type == cheetah_plus)
445                 seq_printf(m, "MMU Type\t: Cheetah+\n");
446         else if (tlb_type == spitfire)
447                 seq_printf(m, "MMU Type\t: Spitfire\n");
448         else if (tlb_type == hypervisor)
449                 seq_printf(m, "MMU Type\t: Hypervisor (sun4v)\n");
450         else
451                 seq_printf(m, "MMU Type\t: ???\n");
452
453 #ifdef CONFIG_DEBUG_DCFLUSH
454         seq_printf(m, "DCPageFlushes\t: %d\n",
455                    atomic_read(&dcpage_flushes));
456 #ifdef CONFIG_SMP
457         seq_printf(m, "DCPageFlushesXC\t: %d\n",
458                    atomic_read(&dcpage_flushes_xcall));
459 #endif /* CONFIG_SMP */
460 #endif /* CONFIG_DEBUG_DCFLUSH */
461 }
462
463 struct linux_prom_translation {
464         unsigned long virt;
465         unsigned long size;
466         unsigned long data;
467 };
468
469 /* Exported for kernel TLB miss handling in ktlb.S */
470 struct linux_prom_translation prom_trans[512] __read_mostly;
471 unsigned int prom_trans_ents __read_mostly;
472
473 /* Exported for SMP bootup purposes. */
474 unsigned long kern_locked_tte_data;
475
476 /* The obp translations are saved based on 8k pagesize, since obp can
477  * use a mixture of pagesizes. Misses to the LOW_OBP_ADDRESS ->
478  * HI_OBP_ADDRESS range are handled in ktlb.S.
479  */
480 static inline int in_obp_range(unsigned long vaddr)
481 {
482         return (vaddr >= LOW_OBP_ADDRESS &&
483                 vaddr < HI_OBP_ADDRESS);
484 }
485
486 static int cmp_ptrans(const void *a, const void *b)
487 {
488         const struct linux_prom_translation *x = a, *y = b;
489
490         if (x->virt > y->virt)
491                 return 1;
492         if (x->virt < y->virt)
493                 return -1;
494         return 0;
495 }
496
497 /* Read OBP translations property into 'prom_trans[]'.  */
498 static void __init read_obp_translations(void)
499 {
500         int n, node, ents, first, last, i;
501
502         node = prom_finddevice("/virtual-memory");
503         n = prom_getproplen(node, "translations");
504         if (unlikely(n == 0 || n == -1)) {
505                 prom_printf("prom_mappings: Couldn't get size.\n");
506                 prom_halt();
507         }
508         if (unlikely(n > sizeof(prom_trans))) {
509                 prom_printf("prom_mappings: Size %Zd is too big.\n", n);
510                 prom_halt();
511         }
512
513         if ((n = prom_getproperty(node, "translations",
514                                   (char *)&prom_trans[0],
515                                   sizeof(prom_trans))) == -1) {
516                 prom_printf("prom_mappings: Couldn't get property.\n");
517                 prom_halt();
518         }
519
520         n = n / sizeof(struct linux_prom_translation);
521
522         ents = n;
523
524         sort(prom_trans, ents, sizeof(struct linux_prom_translation),
525              cmp_ptrans, NULL);
526
527         /* Now kick out all the non-OBP entries.  */
528         for (i = 0; i < ents; i++) {
529                 if (in_obp_range(prom_trans[i].virt))
530                         break;
531         }
532         first = i;
533         for (; i < ents; i++) {
534                 if (!in_obp_range(prom_trans[i].virt))
535                         break;
536         }
537         last = i;
538
539         for (i = 0; i < (last - first); i++) {
540                 struct linux_prom_translation *src = &prom_trans[i + first];
541                 struct linux_prom_translation *dest = &prom_trans[i];
542
543                 *dest = *src;
544         }
545         for (; i < ents; i++) {
546                 struct linux_prom_translation *dest = &prom_trans[i];
547                 dest->virt = dest->size = dest->data = 0x0UL;
548         }
549
550         prom_trans_ents = last - first;
551
552         if (tlb_type == spitfire) {
553                 /* Clear diag TTE bits. */
554                 for (i = 0; i < prom_trans_ents; i++)
555                         prom_trans[i].data &= ~0x0003fe0000000000UL;
556         }
557 }
558
559 static void __init hypervisor_tlb_lock(unsigned long vaddr,
560                                        unsigned long pte,
561                                        unsigned long mmu)
562 {
563         unsigned long ret = sun4v_mmu_map_perm_addr(vaddr, 0, pte, mmu);
564
565         if (ret != 0) {
566                 prom_printf("hypervisor_tlb_lock[%lx:%lx:%lx:%lx]: "
567                             "errors with %lx\n", vaddr, 0, pte, mmu, ret);
568                 prom_halt();
569         }
570 }
571
572 static unsigned long kern_large_tte(unsigned long paddr);
573
574 static void __init remap_kernel(void)
575 {
576         unsigned long phys_page, tte_vaddr, tte_data;
577         int i, tlb_ent = sparc64_highest_locked_tlbent();
578
579         tte_vaddr = (unsigned long) KERNBASE;
580         phys_page = (prom_boot_mapping_phys_low >> 22UL) << 22UL;
581         tte_data = kern_large_tte(phys_page);
582
583         kern_locked_tte_data = tte_data;
584
585         /* Now lock us into the TLBs via Hypervisor or OBP. */
586         if (tlb_type == hypervisor) {
587                 for (i = 0; i < num_kernel_image_mappings; i++) {
588                         hypervisor_tlb_lock(tte_vaddr, tte_data, HV_MMU_DMMU);
589                         hypervisor_tlb_lock(tte_vaddr, tte_data, HV_MMU_IMMU);
590                         tte_vaddr += 0x400000;
591                         tte_data += 0x400000;
592                 }
593         } else {
594                 for (i = 0; i < num_kernel_image_mappings; i++) {
595                         prom_dtlb_load(tlb_ent - i, tte_data, tte_vaddr);
596                         prom_itlb_load(tlb_ent - i, tte_data, tte_vaddr);
597                         tte_vaddr += 0x400000;
598                         tte_data += 0x400000;
599                 }
600                 sparc64_highest_unlocked_tlb_ent = tlb_ent - i;
601         }
602         if (tlb_type == cheetah_plus) {
603                 sparc64_kern_pri_context = (CTX_CHEETAH_PLUS_CTX0 |
604                                             CTX_CHEETAH_PLUS_NUC);
605                 sparc64_kern_pri_nuc_bits = CTX_CHEETAH_PLUS_NUC;
606                 sparc64_kern_sec_context = CTX_CHEETAH_PLUS_CTX0;
607         }
608 }
609
610
611 static void __init inherit_prom_mappings(void)
612 {
613         /* Now fixup OBP's idea about where we really are mapped. */
614         printk("Remapping the kernel... ");
615         remap_kernel();
616         printk("done.\n");
617 }
618
619 void prom_world(int enter)
620 {
621         if (!enter)
622                 set_fs((mm_segment_t) { get_thread_current_ds() });
623
624         __asm__ __volatile__("flushw");
625 }
626
627 void __flush_dcache_range(unsigned long start, unsigned long end)
628 {
629         unsigned long va;
630
631         if (tlb_type == spitfire) {
632                 int n = 0;
633
634                 for (va = start; va < end; va += 32) {
635                         spitfire_put_dcache_tag(va & 0x3fe0, 0x0);
636                         if (++n >= 512)
637                                 break;
638                 }
639         } else if (tlb_type == cheetah || tlb_type == cheetah_plus) {
640                 start = __pa(start);
641                 end = __pa(end);
642                 for (va = start; va < end; va += 32)
643                         __asm__ __volatile__("stxa %%g0, [%0] %1\n\t"
644                                              "membar #Sync"
645                                              : /* no outputs */
646                                              : "r" (va),
647                                                "i" (ASI_DCACHE_INVALIDATE));
648         }
649 }
650
651 /* get_new_mmu_context() uses "cache + 1".  */
652 DEFINE_SPINLOCK(ctx_alloc_lock);
653 unsigned long tlb_context_cache = CTX_FIRST_VERSION - 1;
654 #define MAX_CTX_NR      (1UL << CTX_NR_BITS)
655 #define CTX_BMAP_SLOTS  BITS_TO_LONGS(MAX_CTX_NR)
656 DECLARE_BITMAP(mmu_context_bmap, MAX_CTX_NR);
657
658 /* Caller does TLB context flushing on local CPU if necessary.
659  * The caller also ensures that CTX_VALID(mm->context) is false.
660  *
661  * We must be careful about boundary cases so that we never
662  * let the user have CTX 0 (nucleus) or we ever use a CTX
663  * version of zero (and thus NO_CONTEXT would not be caught
664  * by version mis-match tests in mmu_context.h).
665  *
666  * Always invoked with interrupts disabled.
667  */
668 void get_new_mmu_context(struct mm_struct *mm)
669 {
670         unsigned long ctx, new_ctx;
671         unsigned long orig_pgsz_bits;
672         unsigned long flags;
673         int new_version;
674
675         spin_lock_irqsave(&ctx_alloc_lock, flags);
676         orig_pgsz_bits = (mm->context.sparc64_ctx_val & CTX_PGSZ_MASK);
677         ctx = (tlb_context_cache + 1) & CTX_NR_MASK;
678         new_ctx = find_next_zero_bit(mmu_context_bmap, 1 << CTX_NR_BITS, ctx);
679         new_version = 0;
680         if (new_ctx >= (1 << CTX_NR_BITS)) {
681                 new_ctx = find_next_zero_bit(mmu_context_bmap, ctx, 1);
682                 if (new_ctx >= ctx) {
683                         int i;
684                         new_ctx = (tlb_context_cache & CTX_VERSION_MASK) +
685                                 CTX_FIRST_VERSION;
686                         if (new_ctx == 1)
687                                 new_ctx = CTX_FIRST_VERSION;
688
689                         /* Don't call memset, for 16 entries that's just
690                          * plain silly...
691                          */
692                         mmu_context_bmap[0] = 3;
693                         mmu_context_bmap[1] = 0;
694                         mmu_context_bmap[2] = 0;
695                         mmu_context_bmap[3] = 0;
696                         for (i = 4; i < CTX_BMAP_SLOTS; i += 4) {
697                                 mmu_context_bmap[i + 0] = 0;
698                                 mmu_context_bmap[i + 1] = 0;
699                                 mmu_context_bmap[i + 2] = 0;
700                                 mmu_context_bmap[i + 3] = 0;
701                         }
702                         new_version = 1;
703                         goto out;
704                 }
705         }
706         mmu_context_bmap[new_ctx>>6] |= (1UL << (new_ctx & 63));
707         new_ctx |= (tlb_context_cache & CTX_VERSION_MASK);
708 out:
709         tlb_context_cache = new_ctx;
710         mm->context.sparc64_ctx_val = new_ctx | orig_pgsz_bits;
711         spin_unlock_irqrestore(&ctx_alloc_lock, flags);
712
713         if (unlikely(new_version))
714                 smp_new_mmu_context_version();
715 }
716
717 static int numa_enabled = 1;
718 static int numa_debug;
719
720 static int __init early_numa(char *p)
721 {
722         if (!p)
723                 return 0;
724
725         if (strstr(p, "off"))
726                 numa_enabled = 0;
727
728         if (strstr(p, "debug"))
729                 numa_debug = 1;
730
731         return 0;
732 }
733 early_param("numa", early_numa);
734
735 #define numadbg(f, a...) \
736 do {    if (numa_debug) \
737                 printk(KERN_INFO f, ## a); \
738 } while (0)
739
740 static void __init find_ramdisk(unsigned long phys_base)
741 {
742 #ifdef CONFIG_BLK_DEV_INITRD
743         if (sparc_ramdisk_image || sparc_ramdisk_image64) {
744                 unsigned long ramdisk_image;
745
746                 /* Older versions of the bootloader only supported a
747                  * 32-bit physical address for the ramdisk image
748                  * location, stored at sparc_ramdisk_image.  Newer
749                  * SILO versions set sparc_ramdisk_image to zero and
750                  * provide a full 64-bit physical address at
751                  * sparc_ramdisk_image64.
752                  */
753                 ramdisk_image = sparc_ramdisk_image;
754                 if (!ramdisk_image)
755                         ramdisk_image = sparc_ramdisk_image64;
756
757                 /* Another bootloader quirk.  The bootloader normalizes
758                  * the physical address to KERNBASE, so we have to
759                  * factor that back out and add in the lowest valid
760                  * physical page address to get the true physical address.
761                  */
762                 ramdisk_image -= KERNBASE;
763                 ramdisk_image += phys_base;
764
765                 numadbg("Found ramdisk at physical address 0x%lx, size %u\n",
766                         ramdisk_image, sparc_ramdisk_size);
767
768                 initrd_start = ramdisk_image;
769                 initrd_end = ramdisk_image + sparc_ramdisk_size;
770
771                 lmb_reserve(initrd_start, initrd_end);
772
773                 initrd_start += PAGE_OFFSET;
774                 initrd_end += PAGE_OFFSET;
775         }
776 #endif
777 }
778
779 struct node_mem_mask {
780         unsigned long mask;
781         unsigned long val;
782         unsigned long bootmem_paddr;
783 };
784 static struct node_mem_mask node_masks[MAX_NUMNODES];
785 static int num_node_masks;
786
787 int numa_cpu_lookup_table[NR_CPUS];
788 cpumask_t numa_cpumask_lookup_table[MAX_NUMNODES];
789
790 #ifdef CONFIG_NEED_MULTIPLE_NODES
791 static bootmem_data_t plat_node_bdata[MAX_NUMNODES];
792
793 struct mdesc_mblock {
794         u64     base;
795         u64     size;
796         u64     offset; /* RA-to-PA */
797 };
798 static struct mdesc_mblock *mblocks;
799 static int num_mblocks;
800
801 static unsigned long ra_to_pa(unsigned long addr)
802 {
803         int i;
804
805         for (i = 0; i < num_mblocks; i++) {
806                 struct mdesc_mblock *m = &mblocks[i];
807
808                 if (addr >= m->base &&
809                     addr < (m->base + m->size)) {
810                         addr += m->offset;
811                         break;
812                 }
813         }
814         return addr;
815 }
816
817 static int find_node(unsigned long addr)
818 {
819         int i;
820
821         addr = ra_to_pa(addr);
822         for (i = 0; i < num_node_masks; i++) {
823                 struct node_mem_mask *p = &node_masks[i];
824
825                 if ((addr & p->mask) == p->val)
826                         return i;
827         }
828         return -1;
829 }
830
831 static unsigned long nid_range(unsigned long start, unsigned long end,
832                                int *nid)
833 {
834         *nid = find_node(start);
835         start += PAGE_SIZE;
836         while (start < end) {
837                 int n = find_node(start);
838
839                 if (n != *nid)
840                         break;
841                 start += PAGE_SIZE;
842         }
843
844         return start;
845 }
846 #else
847 static unsigned long nid_range(unsigned long start, unsigned long end,
848                                int *nid)
849 {
850         *nid = 0;
851         return end;
852 }
853 #endif
854
855 /* This must be invoked after performing all of the necessary
856  * add_active_range() calls for 'nid'.  We need to be able to get
857  * correct data from get_pfn_range_for_nid().
858  */
859 static void __init allocate_node_data(int nid)
860 {
861         unsigned long paddr, num_pages, start_pfn, end_pfn;
862         struct pglist_data *p;
863
864 #ifdef CONFIG_NEED_MULTIPLE_NODES
865         paddr = lmb_alloc_nid(sizeof(struct pglist_data),
866                               SMP_CACHE_BYTES, nid, nid_range);
867         if (!paddr) {
868                 prom_printf("Cannot allocate pglist_data for nid[%d]\n", nid);
869                 prom_halt();
870         }
871         NODE_DATA(nid) = __va(paddr);
872         memset(NODE_DATA(nid), 0, sizeof(struct pglist_data));
873
874         NODE_DATA(nid)->bdata = &plat_node_bdata[nid];
875 #endif
876
877         p = NODE_DATA(nid);
878
879         get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
880         p->node_start_pfn = start_pfn;
881         p->node_spanned_pages = end_pfn - start_pfn;
882
883         if (p->node_spanned_pages) {
884                 num_pages = bootmem_bootmap_pages(p->node_spanned_pages);
885
886                 paddr = lmb_alloc_nid(num_pages << PAGE_SHIFT, PAGE_SIZE, nid,
887                                       nid_range);
888                 if (!paddr) {
889                         prom_printf("Cannot allocate bootmap for nid[%d]\n",
890                                   nid);
891                         prom_halt();
892                 }
893                 node_masks[nid].bootmem_paddr = paddr;
894         }
895 }
896
897 static void init_node_masks_nonnuma(void)
898 {
899         int i;
900
901         numadbg("Initializing tables for non-numa.\n");
902
903         node_masks[0].mask = node_masks[0].val = 0;
904         num_node_masks = 1;
905
906         for (i = 0; i < NR_CPUS; i++)
907                 numa_cpu_lookup_table[i] = 0;
908
909         numa_cpumask_lookup_table[0] = CPU_MASK_ALL;
910 }
911
912 #ifdef CONFIG_NEED_MULTIPLE_NODES
913 struct pglist_data *node_data[MAX_NUMNODES];
914
915 EXPORT_SYMBOL(numa_cpu_lookup_table);
916 EXPORT_SYMBOL(numa_cpumask_lookup_table);
917 EXPORT_SYMBOL(node_data);
918
919 struct mdesc_mlgroup {
920         u64     node;
921         u64     latency;
922         u64     match;
923         u64     mask;
924 };
925 static struct mdesc_mlgroup *mlgroups;
926 static int num_mlgroups;
927
928 static int scan_pio_for_cfg_handle(struct mdesc_handle *md, u64 pio,
929                                    u32 cfg_handle)
930 {
931         u64 arc;
932
933         mdesc_for_each_arc(arc, md, pio, MDESC_ARC_TYPE_FWD) {
934                 u64 target = mdesc_arc_target(md, arc);
935                 const u64 *val;
936
937                 val = mdesc_get_property(md, target,
938                                          "cfg-handle", NULL);
939                 if (val && *val == cfg_handle)
940                         return 0;
941         }
942         return -ENODEV;
943 }
944
945 static int scan_arcs_for_cfg_handle(struct mdesc_handle *md, u64 grp,
946                                     u32 cfg_handle)
947 {
948         u64 arc, candidate, best_latency = ~(u64)0;
949
950         candidate = MDESC_NODE_NULL;
951         mdesc_for_each_arc(arc, md, grp, MDESC_ARC_TYPE_FWD) {
952                 u64 target = mdesc_arc_target(md, arc);
953                 const char *name = mdesc_node_name(md, target);
954                 const u64 *val;
955
956                 if (strcmp(name, "pio-latency-group"))
957                         continue;
958
959                 val = mdesc_get_property(md, target, "latency", NULL);
960                 if (!val)
961                         continue;
962
963                 if (*val < best_latency) {
964                         candidate = target;
965                         best_latency = *val;
966                 }
967         }
968
969         if (candidate == MDESC_NODE_NULL)
970                 return -ENODEV;
971
972         return scan_pio_for_cfg_handle(md, candidate, cfg_handle);
973 }
974
975 int of_node_to_nid(struct device_node *dp)
976 {
977         const struct linux_prom64_registers *regs;
978         struct mdesc_handle *md;
979         u32 cfg_handle;
980         int count, nid;
981         u64 grp;
982
983         if (!mlgroups)
984                 return -1;
985
986         regs = of_get_property(dp, "reg", NULL);
987         if (!regs)
988                 return -1;
989
990         cfg_handle = (regs->phys_addr >> 32UL) & 0x0fffffff;
991
992         md = mdesc_grab();
993
994         count = 0;
995         nid = -1;
996         mdesc_for_each_node_by_name(md, grp, "group") {
997                 if (!scan_arcs_for_cfg_handle(md, grp, cfg_handle)) {
998                         nid = count;
999                         break;
1000                 }
1001                 count++;
1002         }
1003
1004         mdesc_release(md);
1005
1006         return nid;
1007 }
1008
1009 static void add_node_ranges(void)
1010 {
1011         int i;
1012
1013         for (i = 0; i < lmb.memory.cnt; i++) {
1014                 unsigned long size = lmb_size_bytes(&lmb.memory, i);
1015                 unsigned long start, end;
1016
1017                 start = lmb.memory.region[i].base;
1018                 end = start + size;
1019                 while (start < end) {
1020                         unsigned long this_end;
1021                         int nid;
1022
1023                         this_end = nid_range(start, end, &nid);
1024
1025                         numadbg("Adding active range nid[%d] "
1026                                 "start[%lx] end[%lx]\n",
1027                                 nid, start, this_end);
1028
1029                         add_active_range(nid,
1030                                          start >> PAGE_SHIFT,
1031                                          this_end >> PAGE_SHIFT);
1032
1033                         start = this_end;
1034                 }
1035         }
1036 }
1037
1038 static int __init grab_mlgroups(struct mdesc_handle *md)
1039 {
1040         unsigned long paddr;
1041         int count = 0;
1042         u64 node;
1043
1044         mdesc_for_each_node_by_name(md, node, "memory-latency-group")
1045                 count++;
1046         if (!count)
1047                 return -ENOENT;
1048
1049         paddr = lmb_alloc(count * sizeof(struct mdesc_mlgroup),
1050                           SMP_CACHE_BYTES);
1051         if (!paddr)
1052                 return -ENOMEM;
1053
1054         mlgroups = __va(paddr);
1055         num_mlgroups = count;
1056
1057         count = 0;
1058         mdesc_for_each_node_by_name(md, node, "memory-latency-group") {
1059                 struct mdesc_mlgroup *m = &mlgroups[count++];
1060                 const u64 *val;
1061
1062                 m->node = node;
1063
1064                 val = mdesc_get_property(md, node, "latency", NULL);
1065                 m->latency = *val;
1066                 val = mdesc_get_property(md, node, "address-match", NULL);
1067                 m->match = *val;
1068                 val = mdesc_get_property(md, node, "address-mask", NULL);
1069                 m->mask = *val;
1070
1071                 numadbg("MLGROUP[%d]: node[%lx] latency[%lx] "
1072                         "match[%lx] mask[%lx]\n",
1073                         count - 1, m->node, m->latency, m->match, m->mask);
1074         }
1075
1076         return 0;
1077 }
1078
1079 static int __init grab_mblocks(struct mdesc_handle *md)
1080 {
1081         unsigned long paddr;
1082         int count = 0;
1083         u64 node;
1084
1085         mdesc_for_each_node_by_name(md, node, "mblock")
1086                 count++;
1087         if (!count)
1088                 return -ENOENT;
1089
1090         paddr = lmb_alloc(count * sizeof(struct mdesc_mblock),
1091                           SMP_CACHE_BYTES);
1092         if (!paddr)
1093                 return -ENOMEM;
1094
1095         mblocks = __va(paddr);
1096         num_mblocks = count;
1097
1098         count = 0;
1099         mdesc_for_each_node_by_name(md, node, "mblock") {
1100                 struct mdesc_mblock *m = &mblocks[count++];
1101                 const u64 *val;
1102
1103                 val = mdesc_get_property(md, node, "base", NULL);
1104                 m->base = *val;
1105                 val = mdesc_get_property(md, node, "size", NULL);
1106                 m->size = *val;
1107                 val = mdesc_get_property(md, node,
1108                                          "address-congruence-offset", NULL);
1109                 m->offset = *val;
1110
1111                 numadbg("MBLOCK[%d]: base[%lx] size[%lx] offset[%lx]\n",
1112                         count - 1, m->base, m->size, m->offset);
1113         }
1114
1115         return 0;
1116 }
1117
1118 static void __init numa_parse_mdesc_group_cpus(struct mdesc_handle *md,
1119                                                u64 grp, cpumask_t *mask)
1120 {
1121         u64 arc;
1122
1123         cpus_clear(*mask);
1124
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;
1129
1130                 if (strcmp(name, "cpu"))
1131                         continue;
1132                 id = mdesc_get_property(md, target, "id", NULL);
1133                 if (*id < NR_CPUS)
1134                         cpu_set(*id, *mask);
1135         }
1136 }
1137
1138 static struct mdesc_mlgroup * __init find_mlgroup(u64 node)
1139 {
1140         int i;
1141
1142         for (i = 0; i < num_mlgroups; i++) {
1143                 struct mdesc_mlgroup *m = &mlgroups[i];
1144                 if (m->node == node)
1145                         return m;
1146         }
1147         return NULL;
1148 }
1149
1150 static int __init numa_attach_mlgroup(struct mdesc_handle *md, u64 grp,
1151                                       int index)
1152 {
1153         struct mdesc_mlgroup *candidate = NULL;
1154         u64 arc, best_latency = ~(u64)0;
1155         struct node_mem_mask *n;
1156
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;
1165                 }
1166         }
1167         if (!candidate)
1168                 return -ENOENT;
1169
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;
1175         }
1176
1177         n = &node_masks[num_node_masks++];
1178
1179         n->mask = candidate->mask;
1180         n->val = candidate->match;
1181
1182         numadbg("NUMA NODE[%d]: mask[%lx] val[%lx] (latency[%lx])\n",
1183                 index, n->mask, n->val, candidate->latency);
1184
1185         return 0;
1186 }
1187
1188 static int __init numa_parse_mdesc_group(struct mdesc_handle *md, u64 grp,
1189                                          int index)
1190 {
1191         cpumask_t mask;
1192         int cpu;
1193
1194         numa_parse_mdesc_group_cpus(md, grp, &mask);
1195
1196         for_each_cpu_mask(cpu, mask)
1197                 numa_cpu_lookup_table[cpu] = index;
1198         numa_cpumask_lookup_table[index] = mask;
1199
1200         if (numa_debug) {
1201                 printk(KERN_INFO "NUMA GROUP[%d]: cpus [ ", index);
1202                 for_each_cpu_mask(cpu, mask)
1203                         printk("%d ", cpu);
1204                 printk("]\n");
1205         }
1206
1207         return numa_attach_mlgroup(md, grp, index);
1208 }
1209
1210 static int __init numa_parse_mdesc(void)
1211 {
1212         struct mdesc_handle *md = mdesc_grab();
1213         int i, err, count;
1214         u64 node;
1215
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;
1220         }
1221
1222         err = grab_mblocks(md);
1223         if (err < 0)
1224                 goto out;
1225
1226         err = grab_mlgroups(md);
1227         if (err < 0)
1228                 goto out;
1229
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++;
1236         }
1237
1238         add_node_ranges();
1239
1240         for (i = 0; i < num_node_masks; i++) {
1241                 allocate_node_data(i);
1242                 node_set_online(i);
1243         }
1244
1245         err = 0;
1246 out:
1247         mdesc_release(md);
1248         return err;
1249 }
1250
1251 static int __init numa_parse_sun4u(void)
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 extern void central_probe(void);
1679
1680 static unsigned long last_valid_pfn;
1681 pgd_t swapper_pg_dir[2048];
1682
1683 static void sun4u_pgprot_init(void);
1684 static void sun4v_pgprot_init(void);
1685
1686 /* Dummy function */
1687 void __init setup_per_cpu_areas(void)
1688 {
1689 }
1690
1691 void __init paging_init(void)
1692 {
1693         unsigned long end_pfn, shift, phys_base;
1694         unsigned long real_end, i;
1695
1696         /* These build time checkes make sure that the dcache_dirty_cpu()
1697          * page->flags usage will work.
1698          *
1699          * When a page gets marked as dcache-dirty, we store the
1700          * cpu number starting at bit 32 in the page->flags.  Also,
1701          * functions like clear_dcache_dirty_cpu use the cpu mask
1702          * in 13-bit signed-immediate instruction fields.
1703          */
1704
1705         /*
1706          * Page flags must not reach into upper 32 bits that are used
1707          * for the cpu number
1708          */
1709         BUILD_BUG_ON(NR_PAGEFLAGS > 32);
1710
1711         /*
1712          * The bit fields placed in the high range must not reach below
1713          * the 32 bit boundary. Otherwise we cannot place the cpu field
1714          * at the 32 bit boundary.
1715          */
1716         BUILD_BUG_ON(SECTIONS_WIDTH + NODES_WIDTH + ZONES_WIDTH +
1717                 ilog2(roundup_pow_of_two(NR_CPUS)) > 32);
1718
1719         BUILD_BUG_ON(NR_CPUS > 4096);
1720
1721         kern_base = (prom_boot_mapping_phys_low >> 22UL) << 22UL;
1722         kern_size = (unsigned long)&_end - (unsigned long)KERNBASE;
1723
1724         sstate_booting();
1725
1726         /* Invalidate both kernel TSBs.  */
1727         memset(swapper_tsb, 0x40, sizeof(swapper_tsb));
1728 #ifndef CONFIG_DEBUG_PAGEALLOC
1729         memset(swapper_4m_tsb, 0x40, sizeof(swapper_4m_tsb));
1730 #endif
1731
1732         if (tlb_type == hypervisor)
1733                 sun4v_pgprot_init();
1734         else
1735                 sun4u_pgprot_init();
1736
1737         if (tlb_type == cheetah_plus ||
1738             tlb_type == hypervisor)
1739                 tsb_phys_patch();
1740
1741         if (tlb_type == hypervisor) {
1742                 sun4v_patch_tlb_handlers();
1743                 sun4v_ktsb_init();
1744         }
1745
1746         lmb_init();
1747
1748         /* Find available physical memory...
1749          *
1750          * Read it twice in order to work around a bug in openfirmware.
1751          * The call to grab this table itself can cause openfirmware to
1752          * allocate memory, which in turn can take away some space from
1753          * the list of available memory.  Reading it twice makes sure
1754          * we really do get the final value.
1755          */
1756         read_obp_translations();
1757         read_obp_memory("reg", &pall[0], &pall_ents);
1758         read_obp_memory("available", &pavail[0], &pavail_ents);
1759         read_obp_memory("available", &pavail[0], &pavail_ents);
1760
1761         phys_base = 0xffffffffffffffffUL;
1762         for (i = 0; i < pavail_ents; i++) {
1763                 phys_base = min(phys_base, pavail[i].phys_addr);
1764                 lmb_add(pavail[i].phys_addr, pavail[i].reg_size);
1765         }
1766
1767         lmb_reserve(kern_base, kern_size);
1768
1769         find_ramdisk(phys_base);
1770
1771         if (cmdline_memory_size)
1772                 lmb_enforce_memory_limit(phys_base + cmdline_memory_size);
1773
1774         lmb_analyze();
1775         lmb_dump_all();
1776
1777         set_bit(0, mmu_context_bmap);
1778
1779         shift = kern_base + PAGE_OFFSET - ((unsigned long)KERNBASE);
1780
1781         real_end = (unsigned long)_end;
1782         num_kernel_image_mappings = DIV_ROUND_UP(real_end - KERNBASE, 1 << 22);
1783         printk("Kernel: Using %d locked TLB entries for main kernel image.\n",
1784                num_kernel_image_mappings);
1785
1786         /* Set kernel pgd to upper alias so physical page computations
1787          * work.
1788          */
1789         init_mm.pgd += ((shift) / (sizeof(pgd_t)));
1790         
1791         memset(swapper_low_pmd_dir, 0, sizeof(swapper_low_pmd_dir));
1792
1793         /* Now can init the kernel/bad page tables. */
1794         pud_set(pud_offset(&swapper_pg_dir[0], 0),
1795                 swapper_low_pmd_dir + (shift / sizeof(pgd_t)));
1796         
1797         inherit_prom_mappings();
1798         
1799         init_kpte_bitmap();
1800
1801         /* Ok, we can use our TLB miss and window trap handlers safely.  */
1802         setup_tba();
1803
1804         __flush_tlb_all();
1805
1806         if (tlb_type == hypervisor)
1807                 sun4v_ktsb_register();
1808
1809         /* We must setup the per-cpu areas before we pull in the
1810          * PROM and the MDESC.  The code there fills in cpu and
1811          * other information into per-cpu data structures.
1812          */
1813         real_setup_per_cpu_areas();
1814
1815         prom_build_devicetree();
1816
1817         if (tlb_type == hypervisor)
1818                 sun4v_mdesc_init();
1819
1820         /* Setup bootmem... */
1821         last_valid_pfn = end_pfn = bootmem_init(phys_base);
1822
1823 #ifndef CONFIG_NEED_MULTIPLE_NODES
1824         max_mapnr = last_valid_pfn;
1825 #endif
1826         kernel_physical_mapping_init();
1827
1828         {
1829                 unsigned long max_zone_pfns[MAX_NR_ZONES];
1830
1831                 memset(max_zone_pfns, 0, sizeof(max_zone_pfns));
1832
1833                 max_zone_pfns[ZONE_NORMAL] = end_pfn;
1834
1835                 free_area_init_nodes(max_zone_pfns);
1836         }
1837
1838         printk("Booting Linux...\n");
1839
1840         central_probe();
1841         cpu_probe();
1842 }
1843
1844 int __init 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 setup_valid_addr_bitmap_from_pavail(void)
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,
1903                                                 sparc64_valid_addr_bitmap);
1904                                         goto do_next_page;
1905                                 }
1906                         }
1907
1908                         prom_printf("mem_init: Lost memory in pavail\n");
1909                         prom_printf("mem_init: OLD start[%lx] size[%lx]\n",
1910                                     pavail[i].phys_addr,
1911                                     pavail[i].reg_size);
1912                         prom_printf("mem_init: NEW start[%lx] size[%lx]\n",
1913                                     pavail_rescan[i].phys_addr,
1914                                     pavail_rescan[i].reg_size);
1915                         prom_printf("mem_init: Cannot continue, aborting.\n");
1916                         prom_halt();
1917
1918                 do_next_page:
1919                         old_start += PAGE_SIZE;
1920                 }
1921         }
1922 }
1923
1924 void __init mem_init(void)
1925 {
1926         unsigned long codepages, datapages, initpages;
1927         unsigned long addr, last;
1928         int i;
1929
1930         i = last_valid_pfn >> ((22 - PAGE_SHIFT) + 6);
1931         i += 1;
1932         sparc64_valid_addr_bitmap = (unsigned long *) alloc_bootmem(i << 3);
1933         if (sparc64_valid_addr_bitmap == NULL) {
1934                 prom_printf("mem_init: Cannot alloc valid_addr_bitmap.\n");
1935                 prom_halt();
1936         }
1937         memset(sparc64_valid_addr_bitmap, 0, i << 3);
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();
1947
1948         high_memory = __va(last_valid_pfn << PAGE_SHIFT);
1949
1950 #ifdef CONFIG_NEED_MULTIPLE_NODES
1951         for_each_online_node(i) {
1952                 if (NODE_DATA(i)->node_spanned_pages != 0) {
1953                         totalram_pages +=
1954                                 free_all_bootmem_node(NODE_DATA(i));
1955                 }
1956         }
1957 #else
1958         totalram_pages = free_all_bootmem();
1959 #endif
1960
1961         /* We subtract one to account for the mem_map_zero page
1962          * allocated below.
1963          */
1964         totalram_pages -= 1;
1965         num_physpages = totalram_pages;
1966
1967         /*
1968          * Set up the zero page, mark it reserved, so that page count
1969          * is not manipulated when freeing the page from user ptes.
1970          */
1971         mem_map_zero = alloc_pages(GFP_KERNEL|__GFP_ZERO, 0);
1972         if (mem_map_zero == NULL) {
1973                 prom_printf("paging_init: Cannot alloc zero page.\n");
1974                 prom_halt();
1975         }
1976         SetPageReserved(mem_map_zero);
1977
1978         codepages = (((unsigned long) _etext) - ((unsigned long) _start));
1979         codepages = PAGE_ALIGN(codepages) >> PAGE_SHIFT;
1980         datapages = (((unsigned long) _edata) - ((unsigned long) _etext));
1981         datapages = PAGE_ALIGN(datapages) >> PAGE_SHIFT;
1982         initpages = (((unsigned long) __init_end) - ((unsigned long) __init_begin));
1983         initpages = PAGE_ALIGN(initpages) >> PAGE_SHIFT;
1984
1985         printk("Memory: %luk available (%ldk kernel code, %ldk data, %ldk init) [%016lx,%016lx]\n",
1986                nr_free_pages() << (PAGE_SHIFT-10),
1987                codepages << (PAGE_SHIFT-10),
1988                datapages << (PAGE_SHIFT-10), 
1989                initpages << (PAGE_SHIFT-10), 
1990                PAGE_OFFSET, (last_valid_pfn << PAGE_SHIFT));
1991
1992         if (tlb_type == cheetah || tlb_type == cheetah_plus)
1993                 cheetah_ecache_flush_init();
1994 }
1995
1996 void free_initmem(void)
1997 {
1998         unsigned long addr, initend;
1999
2000         /*
2001          * The init section is aligned to 8k in vmlinux.lds. Page align for >8k pagesizes.
2002          */
2003         addr = PAGE_ALIGN((unsigned long)(__init_begin));
2004         initend = (unsigned long)(__init_end) & PAGE_MASK;
2005         for (; addr < initend; addr += PAGE_SIZE) {
2006                 unsigned long page;
2007                 struct page *p;
2008
2009                 page = (addr +
2010                         ((unsigned long) __va(kern_base)) -
2011                         ((unsigned long) KERNBASE));
2012                 memset((void *)addr, POISON_FREE_INITMEM, PAGE_SIZE);
2013                 p = virt_to_page(page);
2014
2015                 ClearPageReserved(p);
2016                 init_page_count(p);
2017                 __free_page(p);
2018                 num_physpages++;
2019                 totalram_pages++;
2020         }
2021 }
2022
2023 #ifdef CONFIG_BLK_DEV_INITRD
2024 void free_initrd_mem(unsigned long start, unsigned long end)
2025 {
2026         if (start < end)
2027                 printk ("Freeing initrd memory: %ldk freed\n", (end - start) >> 10);
2028         for (; start < end; start += PAGE_SIZE) {
2029                 struct page *p = virt_to_page(start);
2030
2031                 ClearPageReserved(p);
2032                 init_page_count(p);
2033                 __free_page(p);
2034                 num_physpages++;
2035                 totalram_pages++;
2036         }
2037 }
2038 #endif
2039
2040 #define _PAGE_CACHE_4U  (_PAGE_CP_4U | _PAGE_CV_4U)
2041 #define _PAGE_CACHE_4V  (_PAGE_CP_4V | _PAGE_CV_4V)
2042 #define __DIRTY_BITS_4U  (_PAGE_MODIFIED_4U | _PAGE_WRITE_4U | _PAGE_W_4U)
2043 #define __DIRTY_BITS_4V  (_PAGE_MODIFIED_4V | _PAGE_WRITE_4V | _PAGE_W_4V)
2044 #define __ACCESS_BITS_4U (_PAGE_ACCESSED_4U | _PAGE_READ_4U | _PAGE_R)
2045 #define __ACCESS_BITS_4V (_PAGE_ACCESSED_4V | _PAGE_READ_4V | _PAGE_R)
2046
2047 pgprot_t PAGE_KERNEL __read_mostly;
2048 EXPORT_SYMBOL(PAGE_KERNEL);
2049
2050 pgprot_t PAGE_KERNEL_LOCKED __read_mostly;
2051 pgprot_t PAGE_COPY __read_mostly;
2052
2053 pgprot_t PAGE_SHARED __read_mostly;
2054 EXPORT_SYMBOL(PAGE_SHARED);
2055
2056 pgprot_t PAGE_EXEC __read_mostly;
2057 unsigned long pg_iobits __read_mostly;
2058
2059 unsigned long _PAGE_IE __read_mostly;
2060 EXPORT_SYMBOL(_PAGE_IE);
2061
2062 unsigned long _PAGE_E __read_mostly;
2063 EXPORT_SYMBOL(_PAGE_E);
2064
2065 unsigned long _PAGE_CACHE __read_mostly;
2066 EXPORT_SYMBOL(_PAGE_CACHE);
2067
2068 #ifdef CONFIG_SPARSEMEM_VMEMMAP
2069
2070 #define VMEMMAP_CHUNK_SHIFT     22
2071 #define VMEMMAP_CHUNK           (1UL << VMEMMAP_CHUNK_SHIFT)
2072 #define VMEMMAP_CHUNK_MASK      ~(VMEMMAP_CHUNK - 1UL)
2073 #define VMEMMAP_ALIGN(x)        (((x)+VMEMMAP_CHUNK-1UL)&VMEMMAP_CHUNK_MASK)
2074
2075 #define VMEMMAP_SIZE    ((((1UL << MAX_PHYSADDR_BITS) >> PAGE_SHIFT) * \
2076                           sizeof(struct page *)) >> VMEMMAP_CHUNK_SHIFT)
2077 unsigned long vmemmap_table[VMEMMAP_SIZE];
2078
2079 int __meminit vmemmap_populate(struct page *start, unsigned long nr, int node)
2080 {
2081         unsigned long vstart = (unsigned long) start;
2082         unsigned long vend = (unsigned long) (start + nr);
2083         unsigned long phys_start = (vstart - VMEMMAP_BASE);
2084         unsigned long phys_end = (vend - VMEMMAP_BASE);
2085         unsigned long addr = phys_start & VMEMMAP_CHUNK_MASK;
2086         unsigned long end = VMEMMAP_ALIGN(phys_end);
2087         unsigned long pte_base;
2088
2089         pte_base = (_PAGE_VALID | _PAGE_SZ4MB_4U |
2090                     _PAGE_CP_4U | _PAGE_CV_4U |
2091                     _PAGE_P_4U | _PAGE_W_4U);
2092         if (tlb_type == hypervisor)
2093                 pte_base = (_PAGE_VALID | _PAGE_SZ4MB_4V |
2094                             _PAGE_CP_4V | _PAGE_CV_4V |
2095                             _PAGE_P_4V | _PAGE_W_4V);
2096
2097         for (; addr < end; addr += VMEMMAP_CHUNK) {
2098                 unsigned long *vmem_pp =
2099                         vmemmap_table + (addr >> VMEMMAP_CHUNK_SHIFT);
2100                 void *block;
2101
2102                 if (!(*vmem_pp & _PAGE_VALID)) {
2103                         block = vmemmap_alloc_block(1UL << 22, node);
2104                         if (!block)
2105                                 return -ENOMEM;
2106
2107                         *vmem_pp = pte_base | __pa(block);
2108
2109                         printk(KERN_INFO "[%p-%p] page_structs=%lu "
2110                                "node=%d entry=%lu/%lu\n", start, block, nr,
2111                                node,
2112                                addr >> VMEMMAP_CHUNK_SHIFT,
2113                                VMEMMAP_SIZE >> VMEMMAP_CHUNK_SHIFT);
2114                 }
2115         }
2116         return 0;
2117 }
2118 #endif /* CONFIG_SPARSEMEM_VMEMMAP */
2119
2120 static void prot_init_common(unsigned long page_none,
2121                              unsigned long page_shared,
2122                              unsigned long page_copy,
2123                              unsigned long page_readonly,
2124                              unsigned long page_exec_bit)
2125 {
2126         PAGE_COPY = __pgprot(page_copy);
2127         PAGE_SHARED = __pgprot(page_shared);
2128
2129         protection_map[0x0] = __pgprot(page_none);
2130         protection_map[0x1] = __pgprot(page_readonly & ~page_exec_bit);
2131         protection_map[0x2] = __pgprot(page_copy & ~page_exec_bit);
2132         protection_map[0x3] = __pgprot(page_copy & ~page_exec_bit);
2133         protection_map[0x4] = __pgprot(page_readonly);
2134         protection_map[0x5] = __pgprot(page_readonly);
2135         protection_map[0x6] = __pgprot(page_copy);
2136         protection_map[0x7] = __pgprot(page_copy);
2137         protection_map[0x8] = __pgprot(page_none);
2138         protection_map[0x9] = __pgprot(page_readonly & ~page_exec_bit);
2139         protection_map[0xa] = __pgprot(page_shared & ~page_exec_bit);
2140         protection_map[0xb] = __pgprot(page_shared & ~page_exec_bit);
2141         protection_map[0xc] = __pgprot(page_readonly);
2142         protection_map[0xd] = __pgprot(page_readonly);
2143         protection_map[0xe] = __pgprot(page_shared);
2144         protection_map[0xf] = __pgprot(page_shared);
2145 }
2146
2147 static void __init sun4u_pgprot_init(void)
2148 {
2149         unsigned long page_none, page_shared, page_copy, page_readonly;
2150         unsigned long page_exec_bit;
2151
2152         PAGE_KERNEL = __pgprot (_PAGE_PRESENT_4U | _PAGE_VALID |
2153                                 _PAGE_CACHE_4U | _PAGE_P_4U |
2154                                 __ACCESS_BITS_4U | __DIRTY_BITS_4U |
2155                                 _PAGE_EXEC_4U);
2156         PAGE_KERNEL_LOCKED = __pgprot (_PAGE_PRESENT_4U | _PAGE_VALID |
2157                                        _PAGE_CACHE_4U | _PAGE_P_4U |
2158                                        __ACCESS_BITS_4U | __DIRTY_BITS_4U |
2159                                        _PAGE_EXEC_4U | _PAGE_L_4U);
2160         PAGE_EXEC = __pgprot(_PAGE_EXEC_4U);
2161
2162         _PAGE_IE = _PAGE_IE_4U;
2163         _PAGE_E = _PAGE_E_4U;
2164         _PAGE_CACHE = _PAGE_CACHE_4U;
2165
2166         pg_iobits = (_PAGE_VALID | _PAGE_PRESENT_4U | __DIRTY_BITS_4U |
2167                      __ACCESS_BITS_4U | _PAGE_E_4U);
2168
2169 #ifdef CONFIG_DEBUG_PAGEALLOC
2170         kern_linear_pte_xor[0] = (_PAGE_VALID | _PAGE_SZBITS_4U) ^
2171                 0xfffff80000000000;
2172 #else
2173         kern_linear_pte_xor[0] = (_PAGE_VALID | _PAGE_SZ4MB_4U) ^
2174                 0xfffff80000000000;
2175 #endif
2176         kern_linear_pte_xor[0] |= (_PAGE_CP_4U | _PAGE_CV_4U |
2177                                    _PAGE_P_4U | _PAGE_W_4U);
2178
2179         /* XXX Should use 256MB on Panther. XXX */
2180         kern_linear_pte_xor[1] = kern_linear_pte_xor[0];
2181
2182         _PAGE_SZBITS = _PAGE_SZBITS_4U;
2183         _PAGE_ALL_SZ_BITS =  (_PAGE_SZ4MB_4U | _PAGE_SZ512K_4U |
2184                               _PAGE_SZ64K_4U | _PAGE_SZ8K_4U |
2185                               _PAGE_SZ32MB_4U | _PAGE_SZ256MB_4U);
2186
2187
2188         page_none = _PAGE_PRESENT_4U | _PAGE_ACCESSED_4U | _PAGE_CACHE_4U;
2189         page_shared = (_PAGE_VALID | _PAGE_PRESENT_4U | _PAGE_CACHE_4U |
2190                        __ACCESS_BITS_4U | _PAGE_WRITE_4U | _PAGE_EXEC_4U);
2191         page_copy   = (_PAGE_VALID | _PAGE_PRESENT_4U | _PAGE_CACHE_4U |
2192                        __ACCESS_BITS_4U | _PAGE_EXEC_4U);
2193         page_readonly   = (_PAGE_VALID | _PAGE_PRESENT_4U | _PAGE_CACHE_4U |
2194                            __ACCESS_BITS_4U | _PAGE_EXEC_4U);
2195
2196         page_exec_bit = _PAGE_EXEC_4U;
2197
2198         prot_init_common(page_none, page_shared, page_copy, page_readonly,
2199                          page_exec_bit);
2200 }
2201
2202 static void __init sun4v_pgprot_init(void)
2203 {
2204         unsigned long page_none, page_shared, page_copy, page_readonly;
2205         unsigned long page_exec_bit;
2206
2207         PAGE_KERNEL = __pgprot (_PAGE_PRESENT_4V | _PAGE_VALID |
2208                                 _PAGE_CACHE_4V | _PAGE_P_4V |
2209                                 __ACCESS_BITS_4V | __DIRTY_BITS_4V |
2210                                 _PAGE_EXEC_4V);
2211         PAGE_KERNEL_LOCKED = PAGE_KERNEL;
2212         PAGE_EXEC = __pgprot(_PAGE_EXEC_4V);
2213
2214         _PAGE_IE = _PAGE_IE_4V;
2215         _PAGE_E = _PAGE_E_4V;
2216         _PAGE_CACHE = _PAGE_CACHE_4V;
2217
2218 #ifdef CONFIG_DEBUG_PAGEALLOC
2219         kern_linear_pte_xor[0] = (_PAGE_VALID | _PAGE_SZBITS_4V) ^
2220                 0xfffff80000000000;
2221 #else
2222         kern_linear_pte_xor[0] = (_PAGE_VALID | _PAGE_SZ4MB_4V) ^
2223                 0xfffff80000000000;
2224 #endif
2225         kern_linear_pte_xor[0] |= (_PAGE_CP_4V | _PAGE_CV_4V |
2226                                    _PAGE_P_4V | _PAGE_W_4V);
2227
2228 #ifdef CONFIG_DEBUG_PAGEALLOC
2229         kern_linear_pte_xor[1] = (_PAGE_VALID | _PAGE_SZBITS_4V) ^
2230                 0xfffff80000000000;
2231 #else
2232         kern_linear_pte_xor[1] = (_PAGE_VALID | _PAGE_SZ256MB_4V) ^
2233                 0xfffff80000000000;
2234 #endif
2235         kern_linear_pte_xor[1] |= (_PAGE_CP_4V | _PAGE_CV_4V |
2236                                    _PAGE_P_4V | _PAGE_W_4V);
2237
2238         pg_iobits = (_PAGE_VALID | _PAGE_PRESENT_4V | __DIRTY_BITS_4V |
2239                      __ACCESS_BITS_4V | _PAGE_E_4V);
2240
2241         _PAGE_SZBITS = _PAGE_SZBITS_4V;
2242         _PAGE_ALL_SZ_BITS = (_PAGE_SZ16GB_4V | _PAGE_SZ2GB_4V |
2243                              _PAGE_SZ256MB_4V | _PAGE_SZ32MB_4V |
2244                              _PAGE_SZ4MB_4V | _PAGE_SZ512K_4V |
2245                              _PAGE_SZ64K_4V | _PAGE_SZ8K_4V);
2246
2247         page_none = _PAGE_PRESENT_4V | _PAGE_ACCESSED_4V | _PAGE_CACHE_4V;
2248         page_shared = (_PAGE_VALID | _PAGE_PRESENT_4V | _PAGE_CACHE_4V |
2249                        __ACCESS_BITS_4V | _PAGE_WRITE_4V | _PAGE_EXEC_4V);
2250         page_copy   = (_PAGE_VALID | _PAGE_PRESENT_4V | _PAGE_CACHE_4V |
2251                        __ACCESS_BITS_4V | _PAGE_EXEC_4V);
2252         page_readonly = (_PAGE_VALID | _PAGE_PRESENT_4V | _PAGE_CACHE_4V |
2253                          __ACCESS_BITS_4V | _PAGE_EXEC_4V);
2254
2255         page_exec_bit = _PAGE_EXEC_4V;
2256
2257         prot_init_common(page_none, page_shared, page_copy, page_readonly,
2258                          page_exec_bit);
2259 }
2260
2261 unsigned long pte_sz_bits(unsigned long sz)
2262 {
2263         if (tlb_type == hypervisor) {
2264                 switch (sz) {
2265                 case 8 * 1024:
2266                 default:
2267                         return _PAGE_SZ8K_4V;
2268                 case 64 * 1024:
2269                         return _PAGE_SZ64K_4V;
2270                 case 512 * 1024:
2271                         return _PAGE_SZ512K_4V;
2272                 case 4 * 1024 * 1024:
2273                         return _PAGE_SZ4MB_4V;
2274                 };
2275         } else {
2276                 switch (sz) {
2277                 case 8 * 1024:
2278                 default:
2279                         return _PAGE_SZ8K_4U;
2280                 case 64 * 1024:
2281                         return _PAGE_SZ64K_4U;
2282                 case 512 * 1024:
2283                         return _PAGE_SZ512K_4U;
2284                 case 4 * 1024 * 1024:
2285                         return _PAGE_SZ4MB_4U;
2286                 };
2287         }
2288 }
2289
2290 pte_t mk_pte_io(unsigned long page, pgprot_t prot, int space, unsigned long page_size)
2291 {
2292         pte_t pte;
2293
2294         pte_val(pte)  = page | pgprot_val(pgprot_noncached(prot));
2295         pte_val(pte) |= (((unsigned long)space) << 32);
2296         pte_val(pte) |= pte_sz_bits(page_size);
2297
2298         return pte;
2299 }
2300
2301 static unsigned long kern_large_tte(unsigned long paddr)
2302 {
2303         unsigned long val;
2304
2305         val = (_PAGE_VALID | _PAGE_SZ4MB_4U |
2306                _PAGE_CP_4U | _PAGE_CV_4U | _PAGE_P_4U |
2307                _PAGE_EXEC_4U | _PAGE_L_4U | _PAGE_W_4U);
2308         if (tlb_type == hypervisor)
2309                 val = (_PAGE_VALID | _PAGE_SZ4MB_4V |
2310                        _PAGE_CP_4V | _PAGE_CV_4V | _PAGE_P_4V |
2311                        _PAGE_EXEC_4V | _PAGE_W_4V);
2312
2313         return val | paddr;
2314 }
2315
2316 /* If not locked, zap it. */
2317 void __flush_tlb_all(void)
2318 {
2319         unsigned long pstate;
2320         int i;
2321
2322         __asm__ __volatile__("flushw\n\t"
2323                              "rdpr      %%pstate, %0\n\t"
2324                              "wrpr      %0, %1, %%pstate"
2325                              : "=r" (pstate)
2326                              : "i" (PSTATE_IE));
2327         if (tlb_type == hypervisor) {
2328                 sun4v_mmu_demap_all();
2329         } else if (tlb_type == spitfire) {
2330                 for (i = 0; i < 64; i++) {
2331                         /* Spitfire Errata #32 workaround */
2332                         /* NOTE: Always runs on spitfire, so no
2333                          *       cheetah+ page size encodings.
2334                          */
2335                         __asm__ __volatile__("stxa      %0, [%1] %2\n\t"
2336                                              "flush     %%g6"
2337                                              : /* No outputs */
2338                                              : "r" (0),
2339                                              "r" (PRIMARY_CONTEXT), "i" (ASI_DMMU));
2340
2341                         if (!(spitfire_get_dtlb_data(i) & _PAGE_L_4U)) {
2342                                 __asm__ __volatile__("stxa %%g0, [%0] %1\n\t"
2343                                                      "membar #Sync"
2344                                                      : /* no outputs */
2345                                                      : "r" (TLB_TAG_ACCESS), "i" (ASI_DMMU));
2346                                 spitfire_put_dtlb_data(i, 0x0UL);
2347                         }
2348
2349                         /* Spitfire Errata #32 workaround */
2350                         /* NOTE: Always runs on spitfire, so no
2351                          *       cheetah+ page size encodings.
2352                          */
2353                         __asm__ __volatile__("stxa      %0, [%1] %2\n\t"
2354                                              "flush     %%g6"
2355                                              : /* No outputs */
2356                                              : "r" (0),
2357                                              "r" (PRIMARY_CONTEXT), "i" (ASI_DMMU));
2358
2359                         if (!(spitfire_get_itlb_data(i) & _PAGE_L_4U)) {
2360                                 __asm__ __volatile__("stxa %%g0, [%0] %1\n\t"
2361                                                      "membar #Sync"
2362                                                      : /* no outputs */
2363                                                      : "r" (TLB_TAG_ACCESS), "i" (ASI_IMMU));
2364                                 spitfire_put_itlb_data(i, 0x0UL);
2365                         }
2366                 }
2367         } else if (tlb_type == cheetah || tlb_type == cheetah_plus) {
2368                 cheetah_flush_dtlb_all();
2369                 cheetah_flush_itlb_all();
2370         }
2371         __asm__ __volatile__("wrpr      %0, 0, %%pstate"
2372                              : : "r" (pstate));
2373 }