Merge branch 'master'
[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/config.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/fs.h>
21 #include <linux/seq_file.h>
22 #include <linux/kprobes.h>
23 #include <linux/cache.h>
24 #include <linux/sort.h>
25
26 #include <asm/head.h>
27 #include <asm/system.h>
28 #include <asm/page.h>
29 #include <asm/pgalloc.h>
30 #include <asm/pgtable.h>
31 #include <asm/oplib.h>
32 #include <asm/iommu.h>
33 #include <asm/io.h>
34 #include <asm/uaccess.h>
35 #include <asm/mmu_context.h>
36 #include <asm/tlbflush.h>
37 #include <asm/dma.h>
38 #include <asm/starfire.h>
39 #include <asm/tlb.h>
40 #include <asm/spitfire.h>
41 #include <asm/sections.h>
42
43 extern void device_scan(void);
44
45 #define MAX_BANKS       32
46
47 static struct linux_prom64_registers pavail[MAX_BANKS] __initdata;
48 static struct linux_prom64_registers pavail_rescan[MAX_BANKS] __initdata;
49 static int pavail_ents __initdata;
50 static int pavail_rescan_ents __initdata;
51
52 static int cmp_p64(const void *a, const void *b)
53 {
54         const struct linux_prom64_registers *x = a, *y = b;
55
56         if (x->phys_addr > y->phys_addr)
57                 return 1;
58         if (x->phys_addr < y->phys_addr)
59                 return -1;
60         return 0;
61 }
62
63 static void __init read_obp_memory(const char *property,
64                                    struct linux_prom64_registers *regs,
65                                    int *num_ents)
66 {
67         int node = prom_finddevice("/memory");
68         int prop_size = prom_getproplen(node, property);
69         int ents, ret, i;
70
71         ents = prop_size / sizeof(struct linux_prom64_registers);
72         if (ents > MAX_BANKS) {
73                 prom_printf("The machine has more %s property entries than "
74                             "this kernel can support (%d).\n",
75                             property, MAX_BANKS);
76                 prom_halt();
77         }
78
79         ret = prom_getproperty(node, property, (char *) regs, prop_size);
80         if (ret == -1) {
81                 prom_printf("Couldn't get %s property from /memory.\n");
82                 prom_halt();
83         }
84
85         *num_ents = ents;
86
87         /* Sanitize what we got from the firmware, by page aligning
88          * everything.
89          */
90         for (i = 0; i < ents; i++) {
91                 unsigned long base, size;
92
93                 base = regs[i].phys_addr;
94                 size = regs[i].reg_size;
95
96                 size &= PAGE_MASK;
97                 if (base & ~PAGE_MASK) {
98                         unsigned long new_base = PAGE_ALIGN(base);
99
100                         size -= new_base - base;
101                         if ((long) size < 0L)
102                                 size = 0UL;
103                         base = new_base;
104                 }
105                 regs[i].phys_addr = base;
106                 regs[i].reg_size = size;
107         }
108         sort(regs, ents, sizeof(struct linux_prom64_registers),
109              cmp_p64, NULL);
110 }
111
112 unsigned long *sparc64_valid_addr_bitmap __read_mostly;
113
114 /* Ugly, but necessary... -DaveM */
115 unsigned long phys_base __read_mostly;
116 unsigned long kern_base __read_mostly;
117 unsigned long kern_size __read_mostly;
118 unsigned long pfn_base __read_mostly;
119
120 /* get_new_mmu_context() uses "cache + 1".  */
121 DEFINE_SPINLOCK(ctx_alloc_lock);
122 unsigned long tlb_context_cache = CTX_FIRST_VERSION - 1;
123 #define CTX_BMAP_SLOTS (1UL << (CTX_NR_BITS - 6))
124 unsigned long mmu_context_bmap[CTX_BMAP_SLOTS];
125
126 /* References to special section boundaries */
127 extern char  _start[], _end[];
128
129 /* Initial ramdisk setup */
130 extern unsigned long sparc_ramdisk_image64;
131 extern unsigned int sparc_ramdisk_image;
132 extern unsigned int sparc_ramdisk_size;
133
134 struct page *mem_map_zero __read_mostly;
135
136 unsigned int sparc64_highest_unlocked_tlb_ent __read_mostly;
137
138 unsigned long sparc64_kern_pri_context __read_mostly;
139 unsigned long sparc64_kern_pri_nuc_bits __read_mostly;
140 unsigned long sparc64_kern_sec_context __read_mostly;
141
142 int bigkernel = 0;
143
144 /* XXX Tune this... */
145 #define PGT_CACHE_LOW   25
146 #define PGT_CACHE_HIGH  50
147
148 void check_pgt_cache(void)
149 {
150         preempt_disable();
151         if (pgtable_cache_size > PGT_CACHE_HIGH) {
152                 do {
153                         if (pgd_quicklist)
154                                 free_pgd_slow(get_pgd_fast());
155                         if (pte_quicklist[0])
156                                 free_pte_slow(pte_alloc_one_fast(NULL, 0));
157                         if (pte_quicklist[1])
158                                 free_pte_slow(pte_alloc_one_fast(NULL, 1 << (PAGE_SHIFT + 10)));
159                 } while (pgtable_cache_size > PGT_CACHE_LOW);
160         }
161         preempt_enable();
162 }
163
164 #ifdef CONFIG_DEBUG_DCFLUSH
165 atomic_t dcpage_flushes = ATOMIC_INIT(0);
166 #ifdef CONFIG_SMP
167 atomic_t dcpage_flushes_xcall = ATOMIC_INIT(0);
168 #endif
169 #endif
170
171 __inline__ void flush_dcache_page_impl(struct page *page)
172 {
173 #ifdef CONFIG_DEBUG_DCFLUSH
174         atomic_inc(&dcpage_flushes);
175 #endif
176
177 #ifdef DCACHE_ALIASING_POSSIBLE
178         __flush_dcache_page(page_address(page),
179                             ((tlb_type == spitfire) &&
180                              page_mapping(page) != NULL));
181 #else
182         if (page_mapping(page) != NULL &&
183             tlb_type == spitfire)
184                 __flush_icache_page(__pa(page_address(page)));
185 #endif
186 }
187
188 #define PG_dcache_dirty         PG_arch_1
189 #define PG_dcache_cpu_shift     24
190 #define PG_dcache_cpu_mask      (256 - 1)
191
192 #if NR_CPUS > 256
193 #error D-cache dirty tracking and thread_info->cpu need fixing for > 256 cpus
194 #endif
195
196 #define dcache_dirty_cpu(page) \
197         (((page)->flags >> PG_dcache_cpu_shift) & PG_dcache_cpu_mask)
198
199 static __inline__ void set_dcache_dirty(struct page *page, int this_cpu)
200 {
201         unsigned long mask = this_cpu;
202         unsigned long non_cpu_bits;
203
204         non_cpu_bits = ~(PG_dcache_cpu_mask << PG_dcache_cpu_shift);
205         mask = (mask << PG_dcache_cpu_shift) | (1UL << PG_dcache_dirty);
206
207         __asm__ __volatile__("1:\n\t"
208                              "ldx       [%2], %%g7\n\t"
209                              "and       %%g7, %1, %%g1\n\t"
210                              "or        %%g1, %0, %%g1\n\t"
211                              "casx      [%2], %%g7, %%g1\n\t"
212                              "cmp       %%g7, %%g1\n\t"
213                              "membar    #StoreLoad | #StoreStore\n\t"
214                              "bne,pn    %%xcc, 1b\n\t"
215                              " nop"
216                              : /* no outputs */
217                              : "r" (mask), "r" (non_cpu_bits), "r" (&page->flags)
218                              : "g1", "g7");
219 }
220
221 static __inline__ void clear_dcache_dirty_cpu(struct page *page, unsigned long cpu)
222 {
223         unsigned long mask = (1UL << PG_dcache_dirty);
224
225         __asm__ __volatile__("! test_and_clear_dcache_dirty\n"
226                              "1:\n\t"
227                              "ldx       [%2], %%g7\n\t"
228                              "srlx      %%g7, %4, %%g1\n\t"
229                              "and       %%g1, %3, %%g1\n\t"
230                              "cmp       %%g1, %0\n\t"
231                              "bne,pn    %%icc, 2f\n\t"
232                              " andn     %%g7, %1, %%g1\n\t"
233                              "casx      [%2], %%g7, %%g1\n\t"
234                              "cmp       %%g7, %%g1\n\t"
235                              "membar    #StoreLoad | #StoreStore\n\t"
236                              "bne,pn    %%xcc, 1b\n\t"
237                              " nop\n"
238                              "2:"
239                              : /* no outputs */
240                              : "r" (cpu), "r" (mask), "r" (&page->flags),
241                                "i" (PG_dcache_cpu_mask),
242                                "i" (PG_dcache_cpu_shift)
243                              : "g1", "g7");
244 }
245
246 void update_mmu_cache(struct vm_area_struct *vma, unsigned long address, pte_t pte)
247 {
248         struct page *page;
249         unsigned long pfn;
250         unsigned long pg_flags;
251
252         pfn = pte_pfn(pte);
253         if (pfn_valid(pfn) &&
254             (page = pfn_to_page(pfn), page_mapping(page)) &&
255             ((pg_flags = page->flags) & (1UL << PG_dcache_dirty))) {
256                 int cpu = ((pg_flags >> PG_dcache_cpu_shift) &
257                            PG_dcache_cpu_mask);
258                 int this_cpu = get_cpu();
259
260                 /* This is just to optimize away some function calls
261                  * in the SMP case.
262                  */
263                 if (cpu == this_cpu)
264                         flush_dcache_page_impl(page);
265                 else
266                         smp_flush_dcache_page_impl(page, cpu);
267
268                 clear_dcache_dirty_cpu(page, cpu);
269
270                 put_cpu();
271         }
272 }
273
274 void flush_dcache_page(struct page *page)
275 {
276         struct address_space *mapping;
277         int this_cpu;
278
279         /* Do not bother with the expensive D-cache flush if it
280          * is merely the zero page.  The 'bigcore' testcase in GDB
281          * causes this case to run millions of times.
282          */
283         if (page == ZERO_PAGE(0))
284                 return;
285
286         this_cpu = get_cpu();
287
288         mapping = page_mapping(page);
289         if (mapping && !mapping_mapped(mapping)) {
290                 int dirty = test_bit(PG_dcache_dirty, &page->flags);
291                 if (dirty) {
292                         int dirty_cpu = dcache_dirty_cpu(page);
293
294                         if (dirty_cpu == this_cpu)
295                                 goto out;
296                         smp_flush_dcache_page_impl(page, dirty_cpu);
297                 }
298                 set_dcache_dirty(page, this_cpu);
299         } else {
300                 /* We could delay the flush for the !page_mapping
301                  * case too.  But that case is for exec env/arg
302                  * pages and those are %99 certainly going to get
303                  * faulted into the tlb (and thus flushed) anyways.
304                  */
305                 flush_dcache_page_impl(page);
306         }
307
308 out:
309         put_cpu();
310 }
311
312 void __kprobes flush_icache_range(unsigned long start, unsigned long end)
313 {
314         /* Cheetah has coherent I-cache. */
315         if (tlb_type == spitfire) {
316                 unsigned long kaddr;
317
318                 for (kaddr = start; kaddr < end; kaddr += PAGE_SIZE)
319                         __flush_icache_page(__get_phys(kaddr));
320         }
321 }
322
323 unsigned long page_to_pfn(struct page *page)
324 {
325         return (unsigned long) ((page - mem_map) + pfn_base);
326 }
327
328 struct page *pfn_to_page(unsigned long pfn)
329 {
330         return (mem_map + (pfn - pfn_base));
331 }
332
333 void show_mem(void)
334 {
335         printk("Mem-info:\n");
336         show_free_areas();
337         printk("Free swap:       %6ldkB\n",
338                nr_swap_pages << (PAGE_SHIFT-10));
339         printk("%ld pages of RAM\n", num_physpages);
340         printk("%d free pages\n", nr_free_pages());
341         printk("%d pages in page table cache\n",pgtable_cache_size);
342 }
343
344 void mmu_info(struct seq_file *m)
345 {
346         if (tlb_type == cheetah)
347                 seq_printf(m, "MMU Type\t: Cheetah\n");
348         else if (tlb_type == cheetah_plus)
349                 seq_printf(m, "MMU Type\t: Cheetah+\n");
350         else if (tlb_type == spitfire)
351                 seq_printf(m, "MMU Type\t: Spitfire\n");
352         else
353                 seq_printf(m, "MMU Type\t: ???\n");
354
355 #ifdef CONFIG_DEBUG_DCFLUSH
356         seq_printf(m, "DCPageFlushes\t: %d\n",
357                    atomic_read(&dcpage_flushes));
358 #ifdef CONFIG_SMP
359         seq_printf(m, "DCPageFlushesXC\t: %d\n",
360                    atomic_read(&dcpage_flushes_xcall));
361 #endif /* CONFIG_SMP */
362 #endif /* CONFIG_DEBUG_DCFLUSH */
363 }
364
365 struct linux_prom_translation {
366         unsigned long virt;
367         unsigned long size;
368         unsigned long data;
369 };
370
371 /* Exported for kernel TLB miss handling in ktlb.S */
372 struct linux_prom_translation prom_trans[512] __read_mostly;
373 unsigned int prom_trans_ents __read_mostly;
374 unsigned int swapper_pgd_zero __read_mostly;
375
376 extern unsigned long prom_boot_page;
377 extern void prom_remap(unsigned long physpage, unsigned long virtpage, int mmu_ihandle);
378 extern int prom_get_mmu_ihandle(void);
379 extern void register_prom_callbacks(void);
380
381 /* Exported for SMP bootup purposes. */
382 unsigned long kern_locked_tte_data;
383
384 /*
385  * Translate PROM's mapping we capture at boot time into physical address.
386  * The second parameter is only set from prom_callback() invocations.
387  */
388 unsigned long prom_virt_to_phys(unsigned long promva, int *error)
389 {
390         int i;
391
392         for (i = 0; i < prom_trans_ents; i++) {
393                 struct linux_prom_translation *p = &prom_trans[i];
394
395                 if (promva >= p->virt &&
396                     promva < (p->virt + p->size)) {
397                         unsigned long base = p->data & _PAGE_PADDR;
398
399                         if (error)
400                                 *error = 0;
401                         return base + (promva & (8192 - 1));
402                 }
403         }
404         if (error)
405                 *error = 1;
406         return 0UL;
407 }
408
409 /* The obp translations are saved based on 8k pagesize, since obp can
410  * use a mixture of pagesizes. Misses to the LOW_OBP_ADDRESS ->
411  * HI_OBP_ADDRESS range are handled in ktlb.S and do not use the vpte
412  * scheme (also, see rant in inherit_locked_prom_mappings()).
413  */
414 static inline int in_obp_range(unsigned long vaddr)
415 {
416         return (vaddr >= LOW_OBP_ADDRESS &&
417                 vaddr < HI_OBP_ADDRESS);
418 }
419
420 static int cmp_ptrans(const void *a, const void *b)
421 {
422         const struct linux_prom_translation *x = a, *y = b;
423
424         if (x->virt > y->virt)
425                 return 1;
426         if (x->virt < y->virt)
427                 return -1;
428         return 0;
429 }
430
431 /* Read OBP translations property into 'prom_trans[]'.  */
432 static void __init read_obp_translations(void)
433 {
434         int n, node, ents, first, last, i;
435
436         node = prom_finddevice("/virtual-memory");
437         n = prom_getproplen(node, "translations");
438         if (unlikely(n == 0 || n == -1)) {
439                 prom_printf("prom_mappings: Couldn't get size.\n");
440                 prom_halt();
441         }
442         if (unlikely(n > sizeof(prom_trans))) {
443                 prom_printf("prom_mappings: Size %Zd is too big.\n", n);
444                 prom_halt();
445         }
446
447         if ((n = prom_getproperty(node, "translations",
448                                   (char *)&prom_trans[0],
449                                   sizeof(prom_trans))) == -1) {
450                 prom_printf("prom_mappings: Couldn't get property.\n");
451                 prom_halt();
452         }
453
454         n = n / sizeof(struct linux_prom_translation);
455
456         ents = n;
457
458         sort(prom_trans, ents, sizeof(struct linux_prom_translation),
459              cmp_ptrans, NULL);
460
461         /* Now kick out all the non-OBP entries.  */
462         for (i = 0; i < ents; i++) {
463                 if (in_obp_range(prom_trans[i].virt))
464                         break;
465         }
466         first = i;
467         for (; i < ents; i++) {
468                 if (!in_obp_range(prom_trans[i].virt))
469                         break;
470         }
471         last = i;
472
473         for (i = 0; i < (last - first); i++) {
474                 struct linux_prom_translation *src = &prom_trans[i + first];
475                 struct linux_prom_translation *dest = &prom_trans[i];
476
477                 *dest = *src;
478         }
479         for (; i < ents; i++) {
480                 struct linux_prom_translation *dest = &prom_trans[i];
481                 dest->virt = dest->size = dest->data = 0x0UL;
482         }
483
484         prom_trans_ents = last - first;
485
486         if (tlb_type == spitfire) {
487                 /* Clear diag TTE bits. */
488                 for (i = 0; i < prom_trans_ents; i++)
489                         prom_trans[i].data &= ~0x0003fe0000000000UL;
490         }
491 }
492
493 static void __init remap_kernel(void)
494 {
495         unsigned long phys_page, tte_vaddr, tte_data;
496         int tlb_ent = sparc64_highest_locked_tlbent();
497
498         tte_vaddr = (unsigned long) KERNBASE;
499         phys_page = (prom_boot_mapping_phys_low >> 22UL) << 22UL;
500         tte_data = (phys_page | (_PAGE_VALID | _PAGE_SZ4MB |
501                                  _PAGE_CP | _PAGE_CV | _PAGE_P |
502                                  _PAGE_L | _PAGE_W));
503
504         kern_locked_tte_data = tte_data;
505
506         /* Now lock us into the TLBs via OBP. */
507         prom_dtlb_load(tlb_ent, tte_data, tte_vaddr);
508         prom_itlb_load(tlb_ent, tte_data, tte_vaddr);
509         if (bigkernel) {
510                 tlb_ent -= 1;
511                 prom_dtlb_load(tlb_ent,
512                                tte_data + 0x400000, 
513                                tte_vaddr + 0x400000);
514                 prom_itlb_load(tlb_ent,
515                                tte_data + 0x400000, 
516                                tte_vaddr + 0x400000);
517         }
518         sparc64_highest_unlocked_tlb_ent = tlb_ent - 1;
519         if (tlb_type == cheetah_plus) {
520                 sparc64_kern_pri_context = (CTX_CHEETAH_PLUS_CTX0 |
521                                             CTX_CHEETAH_PLUS_NUC);
522                 sparc64_kern_pri_nuc_bits = CTX_CHEETAH_PLUS_NUC;
523                 sparc64_kern_sec_context = CTX_CHEETAH_PLUS_CTX0;
524         }
525 }
526
527
528 static void __init inherit_prom_mappings(void)
529 {
530         read_obp_translations();
531
532         /* Now fixup OBP's idea about where we really are mapped. */
533         prom_printf("Remapping the kernel... ");
534         remap_kernel();
535         prom_printf("done.\n");
536
537         prom_printf("Registering callbacks... ");
538         register_prom_callbacks();
539         prom_printf("done.\n");
540 }
541
542 /* The OBP specifications for sun4u mark 0xfffffffc00000000 and
543  * upwards as reserved for use by the firmware (I wonder if this
544  * will be the same on Cheetah...).  We use this virtual address
545  * range for the VPTE table mappings of the nucleus so we need
546  * to zap them when we enter the PROM.  -DaveM
547  */
548 static void __flush_nucleus_vptes(void)
549 {
550         unsigned long prom_reserved_base = 0xfffffffc00000000UL;
551         int i;
552
553         /* Only DTLB must be checked for VPTE entries. */
554         if (tlb_type == spitfire) {
555                 for (i = 0; i < 63; i++) {
556                         unsigned long tag;
557
558                         /* Spitfire Errata #32 workaround */
559                         /* NOTE: Always runs on spitfire, so no cheetah+
560                          *       page size encodings.
561                          */
562                         __asm__ __volatile__("stxa      %0, [%1] %2\n\t"
563                                              "flush     %%g6"
564                                              : /* No outputs */
565                                              : "r" (0),
566                                              "r" (PRIMARY_CONTEXT), "i" (ASI_DMMU));
567
568                         tag = spitfire_get_dtlb_tag(i);
569                         if (((tag & ~(PAGE_MASK)) == 0) &&
570                             ((tag &  (PAGE_MASK)) >= prom_reserved_base)) {
571                                 __asm__ __volatile__("stxa %%g0, [%0] %1\n\t"
572                                                      "membar #Sync"
573                                                      : /* no outputs */
574                                                      : "r" (TLB_TAG_ACCESS), "i" (ASI_DMMU));
575                                 spitfire_put_dtlb_data(i, 0x0UL);
576                         }
577                 }
578         } else if (tlb_type == cheetah || tlb_type == cheetah_plus) {
579                 for (i = 0; i < 512; i++) {
580                         unsigned long tag = cheetah_get_dtlb_tag(i, 2);
581
582                         if ((tag & ~PAGE_MASK) == 0 &&
583                             (tag & PAGE_MASK) >= prom_reserved_base) {
584                                 __asm__ __volatile__("stxa %%g0, [%0] %1\n\t"
585                                                      "membar #Sync"
586                                                      : /* no outputs */
587                                                      : "r" (TLB_TAG_ACCESS), "i" (ASI_DMMU));
588                                 cheetah_put_dtlb_data(i, 0x0UL, 2);
589                         }
590
591                         if (tlb_type != cheetah_plus)
592                                 continue;
593
594                         tag = cheetah_get_dtlb_tag(i, 3);
595
596                         if ((tag & ~PAGE_MASK) == 0 &&
597                             (tag & PAGE_MASK) >= prom_reserved_base) {
598                                 __asm__ __volatile__("stxa %%g0, [%0] %1\n\t"
599                                                      "membar #Sync"
600                                                      : /* no outputs */
601                                                      : "r" (TLB_TAG_ACCESS), "i" (ASI_DMMU));
602                                 cheetah_put_dtlb_data(i, 0x0UL, 3);
603                         }
604                 }
605         } else {
606                 /* Implement me :-) */
607                 BUG();
608         }
609 }
610
611 static int prom_ditlb_set;
612 struct prom_tlb_entry {
613         int             tlb_ent;
614         unsigned long   tlb_tag;
615         unsigned long   tlb_data;
616 };
617 struct prom_tlb_entry prom_itlb[16], prom_dtlb[16];
618
619 void prom_world(int enter)
620 {
621         unsigned long pstate;
622         int i;
623
624         if (!enter)
625                 set_fs((mm_segment_t) { get_thread_current_ds() });
626
627         if (!prom_ditlb_set)
628                 return;
629
630         /* Make sure the following runs atomically. */
631         __asm__ __volatile__("flushw\n\t"
632                              "rdpr      %%pstate, %0\n\t"
633                              "wrpr      %0, %1, %%pstate"
634                              : "=r" (pstate)
635                              : "i" (PSTATE_IE));
636
637         if (enter) {
638                 /* Kick out nucleus VPTEs. */
639                 __flush_nucleus_vptes();
640
641                 /* Install PROM world. */
642                 for (i = 0; i < 16; i++) {
643                         if (prom_dtlb[i].tlb_ent != -1) {
644                                 __asm__ __volatile__("stxa %0, [%1] %2\n\t"
645                                                      "membar #Sync"
646                                         : : "r" (prom_dtlb[i].tlb_tag), "r" (TLB_TAG_ACCESS),
647                                         "i" (ASI_DMMU));
648                                 if (tlb_type == spitfire)
649                                         spitfire_put_dtlb_data(prom_dtlb[i].tlb_ent,
650                                                                prom_dtlb[i].tlb_data);
651                                 else if (tlb_type == cheetah || tlb_type == cheetah_plus)
652                                         cheetah_put_ldtlb_data(prom_dtlb[i].tlb_ent,
653                                                                prom_dtlb[i].tlb_data);
654                         }
655                         if (prom_itlb[i].tlb_ent != -1) {
656                                 __asm__ __volatile__("stxa %0, [%1] %2\n\t"
657                                                      "membar #Sync"
658                                                      : : "r" (prom_itlb[i].tlb_tag),
659                                                      "r" (TLB_TAG_ACCESS),
660                                                      "i" (ASI_IMMU));
661                                 if (tlb_type == spitfire)
662                                         spitfire_put_itlb_data(prom_itlb[i].tlb_ent,
663                                                                prom_itlb[i].tlb_data);
664                                 else if (tlb_type == cheetah || tlb_type == cheetah_plus)
665                                         cheetah_put_litlb_data(prom_itlb[i].tlb_ent,
666                                                                prom_itlb[i].tlb_data);
667                         }
668                 }
669         } else {
670                 for (i = 0; i < 16; i++) {
671                         if (prom_dtlb[i].tlb_ent != -1) {
672                                 __asm__ __volatile__("stxa %%g0, [%0] %1\n\t"
673                                                      "membar #Sync"
674                                         : : "r" (TLB_TAG_ACCESS), "i" (ASI_DMMU));
675                                 if (tlb_type == spitfire)
676                                         spitfire_put_dtlb_data(prom_dtlb[i].tlb_ent, 0x0UL);
677                                 else
678                                         cheetah_put_ldtlb_data(prom_dtlb[i].tlb_ent, 0x0UL);
679                         }
680                         if (prom_itlb[i].tlb_ent != -1) {
681                                 __asm__ __volatile__("stxa %%g0, [%0] %1\n\t"
682                                                      "membar #Sync"
683                                                      : : "r" (TLB_TAG_ACCESS),
684                                                      "i" (ASI_IMMU));
685                                 if (tlb_type == spitfire)
686                                         spitfire_put_itlb_data(prom_itlb[i].tlb_ent, 0x0UL);
687                                 else
688                                         cheetah_put_litlb_data(prom_itlb[i].tlb_ent, 0x0UL);
689                         }
690                 }
691         }
692         __asm__ __volatile__("wrpr      %0, 0, %%pstate"
693                              : : "r" (pstate));
694 }
695
696 void inherit_locked_prom_mappings(int save_p)
697 {
698         int i;
699         int dtlb_seen = 0;
700         int itlb_seen = 0;
701
702         /* Fucking losing PROM has more mappings in the TLB, but
703          * it (conveniently) fails to mention any of these in the
704          * translations property.  The only ones that matter are
705          * the locked PROM tlb entries, so we impose the following
706          * irrecovable rule on the PROM, it is allowed 8 locked
707          * entries in the ITLB and 8 in the DTLB.
708          *
709          * Supposedly the upper 16GB of the address space is
710          * reserved for OBP, BUT I WISH THIS WAS DOCUMENTED
711          * SOMEWHERE!!!!!!!!!!!!!!!!!  Furthermore the entire interface
712          * used between the client program and the firmware on sun5
713          * systems to coordinate mmu mappings is also COMPLETELY
714          * UNDOCUMENTED!!!!!! Thanks S(t)un!
715          */
716         if (save_p) {
717                 for (i = 0; i < 16; i++) {
718                         prom_itlb[i].tlb_ent = -1;
719                         prom_dtlb[i].tlb_ent = -1;
720                 }
721         }
722         if (tlb_type == spitfire) {
723                 int high = sparc64_highest_unlocked_tlb_ent;
724                 for (i = 0; i <= high; i++) {
725                         unsigned long data;
726
727                         /* Spitfire Errata #32 workaround */
728                         /* NOTE: Always runs on spitfire, so no cheetah+
729                          *       page size encodings.
730                          */
731                         __asm__ __volatile__("stxa      %0, [%1] %2\n\t"
732                                              "flush     %%g6"
733                                              : /* No outputs */
734                                              : "r" (0),
735                                              "r" (PRIMARY_CONTEXT), "i" (ASI_DMMU));
736
737                         data = spitfire_get_dtlb_data(i);
738                         if ((data & (_PAGE_L|_PAGE_VALID)) == (_PAGE_L|_PAGE_VALID)) {
739                                 unsigned long tag;
740
741                                 /* Spitfire Errata #32 workaround */
742                                 /* NOTE: Always runs on spitfire, so no
743                                  *       cheetah+ page size encodings.
744                                  */
745                                 __asm__ __volatile__("stxa      %0, [%1] %2\n\t"
746                                                      "flush     %%g6"
747                                                      : /* No outputs */
748                                                      : "r" (0),
749                                                      "r" (PRIMARY_CONTEXT), "i" (ASI_DMMU));
750
751                                 tag = spitfire_get_dtlb_tag(i);
752                                 if (save_p) {
753                                         prom_dtlb[dtlb_seen].tlb_ent = i;
754                                         prom_dtlb[dtlb_seen].tlb_tag = tag;
755                                         prom_dtlb[dtlb_seen].tlb_data = data;
756                                 }
757                                 __asm__ __volatile__("stxa %%g0, [%0] %1\n\t"
758                                                      "membar #Sync"
759                                                      : : "r" (TLB_TAG_ACCESS), "i" (ASI_DMMU));
760                                 spitfire_put_dtlb_data(i, 0x0UL);
761
762                                 dtlb_seen++;
763                                 if (dtlb_seen > 15)
764                                         break;
765                         }
766                 }
767
768                 for (i = 0; i < high; i++) {
769                         unsigned long data;
770
771                         /* Spitfire Errata #32 workaround */
772                         /* NOTE: Always runs on spitfire, so no
773                          *       cheetah+ page size encodings.
774                          */
775                         __asm__ __volatile__("stxa      %0, [%1] %2\n\t"
776                                              "flush     %%g6"
777                                              : /* No outputs */
778                                              : "r" (0),
779                                              "r" (PRIMARY_CONTEXT), "i" (ASI_DMMU));
780
781                         data = spitfire_get_itlb_data(i);
782                         if ((data & (_PAGE_L|_PAGE_VALID)) == (_PAGE_L|_PAGE_VALID)) {
783                                 unsigned long tag;
784
785                                 /* Spitfire Errata #32 workaround */
786                                 /* NOTE: Always runs on spitfire, so no
787                                  *       cheetah+ page size encodings.
788                                  */
789                                 __asm__ __volatile__("stxa      %0, [%1] %2\n\t"
790                                                      "flush     %%g6"
791                                                      : /* No outputs */
792                                                      : "r" (0),
793                                                      "r" (PRIMARY_CONTEXT), "i" (ASI_DMMU));
794
795                                 tag = spitfire_get_itlb_tag(i);
796                                 if (save_p) {
797                                         prom_itlb[itlb_seen].tlb_ent = i;
798                                         prom_itlb[itlb_seen].tlb_tag = tag;
799                                         prom_itlb[itlb_seen].tlb_data = data;
800                                 }
801                                 __asm__ __volatile__("stxa %%g0, [%0] %1\n\t"
802                                                      "membar #Sync"
803                                                      : : "r" (TLB_TAG_ACCESS), "i" (ASI_IMMU));
804                                 spitfire_put_itlb_data(i, 0x0UL);
805
806                                 itlb_seen++;
807                                 if (itlb_seen > 15)
808                                         break;
809                         }
810                 }
811         } else if (tlb_type == cheetah || tlb_type == cheetah_plus) {
812                 int high = sparc64_highest_unlocked_tlb_ent;
813
814                 for (i = 0; i <= high; i++) {
815                         unsigned long data;
816
817                         data = cheetah_get_ldtlb_data(i);
818                         if ((data & (_PAGE_L|_PAGE_VALID)) == (_PAGE_L|_PAGE_VALID)) {
819                                 unsigned long tag;
820
821                                 tag = cheetah_get_ldtlb_tag(i);
822                                 if (save_p) {
823                                         prom_dtlb[dtlb_seen].tlb_ent = i;
824                                         prom_dtlb[dtlb_seen].tlb_tag = tag;
825                                         prom_dtlb[dtlb_seen].tlb_data = data;
826                                 }
827                                 __asm__ __volatile__("stxa %%g0, [%0] %1\n\t"
828                                                      "membar #Sync"
829                                                      : : "r" (TLB_TAG_ACCESS), "i" (ASI_DMMU));
830                                 cheetah_put_ldtlb_data(i, 0x0UL);
831
832                                 dtlb_seen++;
833                                 if (dtlb_seen > 15)
834                                         break;
835                         }
836                 }
837
838                 for (i = 0; i < high; i++) {
839                         unsigned long data;
840
841                         data = cheetah_get_litlb_data(i);
842                         if ((data & (_PAGE_L|_PAGE_VALID)) == (_PAGE_L|_PAGE_VALID)) {
843                                 unsigned long tag;
844
845                                 tag = cheetah_get_litlb_tag(i);
846                                 if (save_p) {
847                                         prom_itlb[itlb_seen].tlb_ent = i;
848                                         prom_itlb[itlb_seen].tlb_tag = tag;
849                                         prom_itlb[itlb_seen].tlb_data = data;
850                                 }
851                                 __asm__ __volatile__("stxa %%g0, [%0] %1\n\t"
852                                                      "membar #Sync"
853                                                      : : "r" (TLB_TAG_ACCESS), "i" (ASI_IMMU));
854                                 cheetah_put_litlb_data(i, 0x0UL);
855
856                                 itlb_seen++;
857                                 if (itlb_seen > 15)
858                                         break;
859                         }
860                 }
861         } else {
862                 /* Implement me :-) */
863                 BUG();
864         }
865         if (save_p)
866                 prom_ditlb_set = 1;
867 }
868
869 /* Give PROM back his world, done during reboots... */
870 void prom_reload_locked(void)
871 {
872         int i;
873
874         for (i = 0; i < 16; i++) {
875                 if (prom_dtlb[i].tlb_ent != -1) {
876                         __asm__ __volatile__("stxa %0, [%1] %2\n\t"
877                                              "membar #Sync"
878                                 : : "r" (prom_dtlb[i].tlb_tag), "r" (TLB_TAG_ACCESS),
879                                 "i" (ASI_DMMU));
880                         if (tlb_type == spitfire)
881                                 spitfire_put_dtlb_data(prom_dtlb[i].tlb_ent,
882                                                        prom_dtlb[i].tlb_data);
883                         else if (tlb_type == cheetah || tlb_type == cheetah_plus)
884                                 cheetah_put_ldtlb_data(prom_dtlb[i].tlb_ent,
885                                                       prom_dtlb[i].tlb_data);
886                 }
887
888                 if (prom_itlb[i].tlb_ent != -1) {
889                         __asm__ __volatile__("stxa %0, [%1] %2\n\t"
890                                              "membar #Sync"
891                                              : : "r" (prom_itlb[i].tlb_tag),
892                                              "r" (TLB_TAG_ACCESS),
893                                              "i" (ASI_IMMU));
894                         if (tlb_type == spitfire)
895                                 spitfire_put_itlb_data(prom_itlb[i].tlb_ent,
896                                                        prom_itlb[i].tlb_data);
897                         else
898                                 cheetah_put_litlb_data(prom_itlb[i].tlb_ent,
899                                                        prom_itlb[i].tlb_data);
900                 }
901         }
902 }
903
904 #ifdef DCACHE_ALIASING_POSSIBLE
905 void __flush_dcache_range(unsigned long start, unsigned long end)
906 {
907         unsigned long va;
908
909         if (tlb_type == spitfire) {
910                 int n = 0;
911
912                 for (va = start; va < end; va += 32) {
913                         spitfire_put_dcache_tag(va & 0x3fe0, 0x0);
914                         if (++n >= 512)
915                                 break;
916                 }
917         } else {
918                 start = __pa(start);
919                 end = __pa(end);
920                 for (va = start; va < end; va += 32)
921                         __asm__ __volatile__("stxa %%g0, [%0] %1\n\t"
922                                              "membar #Sync"
923                                              : /* no outputs */
924                                              : "r" (va),
925                                                "i" (ASI_DCACHE_INVALIDATE));
926         }
927 }
928 #endif /* DCACHE_ALIASING_POSSIBLE */
929
930 /* If not locked, zap it. */
931 void __flush_tlb_all(void)
932 {
933         unsigned long pstate;
934         int i;
935
936         __asm__ __volatile__("flushw\n\t"
937                              "rdpr      %%pstate, %0\n\t"
938                              "wrpr      %0, %1, %%pstate"
939                              : "=r" (pstate)
940                              : "i" (PSTATE_IE));
941         if (tlb_type == spitfire) {
942                 for (i = 0; i < 64; i++) {
943                         /* Spitfire Errata #32 workaround */
944                         /* NOTE: Always runs on spitfire, so no
945                          *       cheetah+ page size encodings.
946                          */
947                         __asm__ __volatile__("stxa      %0, [%1] %2\n\t"
948                                              "flush     %%g6"
949                                              : /* No outputs */
950                                              : "r" (0),
951                                              "r" (PRIMARY_CONTEXT), "i" (ASI_DMMU));
952
953                         if (!(spitfire_get_dtlb_data(i) & _PAGE_L)) {
954                                 __asm__ __volatile__("stxa %%g0, [%0] %1\n\t"
955                                                      "membar #Sync"
956                                                      : /* no outputs */
957                                                      : "r" (TLB_TAG_ACCESS), "i" (ASI_DMMU));
958                                 spitfire_put_dtlb_data(i, 0x0UL);
959                         }
960
961                         /* Spitfire Errata #32 workaround */
962                         /* NOTE: Always runs on spitfire, so no
963                          *       cheetah+ page size encodings.
964                          */
965                         __asm__ __volatile__("stxa      %0, [%1] %2\n\t"
966                                              "flush     %%g6"
967                                              : /* No outputs */
968                                              : "r" (0),
969                                              "r" (PRIMARY_CONTEXT), "i" (ASI_DMMU));
970
971                         if (!(spitfire_get_itlb_data(i) & _PAGE_L)) {
972                                 __asm__ __volatile__("stxa %%g0, [%0] %1\n\t"
973                                                      "membar #Sync"
974                                                      : /* no outputs */
975                                                      : "r" (TLB_TAG_ACCESS), "i" (ASI_IMMU));
976                                 spitfire_put_itlb_data(i, 0x0UL);
977                         }
978                 }
979         } else if (tlb_type == cheetah || tlb_type == cheetah_plus) {
980                 cheetah_flush_dtlb_all();
981                 cheetah_flush_itlb_all();
982         }
983         __asm__ __volatile__("wrpr      %0, 0, %%pstate"
984                              : : "r" (pstate));
985 }
986
987 /* Caller does TLB context flushing on local CPU if necessary.
988  * The caller also ensures that CTX_VALID(mm->context) is false.
989  *
990  * We must be careful about boundary cases so that we never
991  * let the user have CTX 0 (nucleus) or we ever use a CTX
992  * version of zero (and thus NO_CONTEXT would not be caught
993  * by version mis-match tests in mmu_context.h).
994  */
995 void get_new_mmu_context(struct mm_struct *mm)
996 {
997         unsigned long ctx, new_ctx;
998         unsigned long orig_pgsz_bits;
999         
1000
1001         spin_lock(&ctx_alloc_lock);
1002         orig_pgsz_bits = (mm->context.sparc64_ctx_val & CTX_PGSZ_MASK);
1003         ctx = (tlb_context_cache + 1) & CTX_NR_MASK;
1004         new_ctx = find_next_zero_bit(mmu_context_bmap, 1 << CTX_NR_BITS, ctx);
1005         if (new_ctx >= (1 << CTX_NR_BITS)) {
1006                 new_ctx = find_next_zero_bit(mmu_context_bmap, ctx, 1);
1007                 if (new_ctx >= ctx) {
1008                         int i;
1009                         new_ctx = (tlb_context_cache & CTX_VERSION_MASK) +
1010                                 CTX_FIRST_VERSION;
1011                         if (new_ctx == 1)
1012                                 new_ctx = CTX_FIRST_VERSION;
1013
1014                         /* Don't call memset, for 16 entries that's just
1015                          * plain silly...
1016                          */
1017                         mmu_context_bmap[0] = 3;
1018                         mmu_context_bmap[1] = 0;
1019                         mmu_context_bmap[2] = 0;
1020                         mmu_context_bmap[3] = 0;
1021                         for (i = 4; i < CTX_BMAP_SLOTS; i += 4) {
1022                                 mmu_context_bmap[i + 0] = 0;
1023                                 mmu_context_bmap[i + 1] = 0;
1024                                 mmu_context_bmap[i + 2] = 0;
1025                                 mmu_context_bmap[i + 3] = 0;
1026                         }
1027                         goto out;
1028                 }
1029         }
1030         mmu_context_bmap[new_ctx>>6] |= (1UL << (new_ctx & 63));
1031         new_ctx |= (tlb_context_cache & CTX_VERSION_MASK);
1032 out:
1033         tlb_context_cache = new_ctx;
1034         mm->context.sparc64_ctx_val = new_ctx | orig_pgsz_bits;
1035         spin_unlock(&ctx_alloc_lock);
1036 }
1037
1038 #ifndef CONFIG_SMP
1039 struct pgtable_cache_struct pgt_quicklists;
1040 #endif
1041
1042 /* OK, we have to color these pages. The page tables are accessed
1043  * by non-Dcache enabled mapping in the VPTE area by the dtlb_backend.S
1044  * code, as well as by PAGE_OFFSET range direct-mapped addresses by 
1045  * other parts of the kernel. By coloring, we make sure that the tlbmiss 
1046  * fast handlers do not get data from old/garbage dcache lines that 
1047  * correspond to an old/stale virtual address (user/kernel) that 
1048  * previously mapped the pagetable page while accessing vpte range 
1049  * addresses. The idea is that if the vpte color and PAGE_OFFSET range 
1050  * color is the same, then when the kernel initializes the pagetable 
1051  * using the later address range, accesses with the first address
1052  * range will see the newly initialized data rather than the garbage.
1053  */
1054 #ifdef DCACHE_ALIASING_POSSIBLE
1055 #define DC_ALIAS_SHIFT  1
1056 #else
1057 #define DC_ALIAS_SHIFT  0
1058 #endif
1059 pte_t *pte_alloc_one_kernel(struct mm_struct *mm, unsigned long address)
1060 {
1061         struct page *page;
1062         unsigned long color;
1063
1064         {
1065                 pte_t *ptep = pte_alloc_one_fast(mm, address);
1066
1067                 if (ptep)
1068                         return ptep;
1069         }
1070
1071         color = VPTE_COLOR(address);
1072         page = alloc_pages(GFP_KERNEL|__GFP_REPEAT, DC_ALIAS_SHIFT);
1073         if (page) {
1074                 unsigned long *to_free;
1075                 unsigned long paddr;
1076                 pte_t *pte;
1077
1078 #ifdef DCACHE_ALIASING_POSSIBLE
1079                 set_page_count(page, 1);
1080                 ClearPageCompound(page);
1081
1082                 set_page_count((page + 1), 1);
1083                 ClearPageCompound(page + 1);
1084 #endif
1085                 paddr = (unsigned long) page_address(page);
1086                 memset((char *)paddr, 0, (PAGE_SIZE << DC_ALIAS_SHIFT));
1087
1088                 if (!color) {
1089                         pte = (pte_t *) paddr;
1090                         to_free = (unsigned long *) (paddr + PAGE_SIZE);
1091                 } else {
1092                         pte = (pte_t *) (paddr + PAGE_SIZE);
1093                         to_free = (unsigned long *) paddr;
1094                 }
1095
1096 #ifdef DCACHE_ALIASING_POSSIBLE
1097                 /* Now free the other one up, adjust cache size. */
1098                 preempt_disable();
1099                 *to_free = (unsigned long) pte_quicklist[color ^ 0x1];
1100                 pte_quicklist[color ^ 0x1] = to_free;
1101                 pgtable_cache_size++;
1102                 preempt_enable();
1103 #endif
1104
1105                 return pte;
1106         }
1107         return NULL;
1108 }
1109
1110 void sparc_ultra_dump_itlb(void)
1111 {
1112         int slot;
1113
1114         if (tlb_type == spitfire) {
1115                 printk ("Contents of itlb: ");
1116                 for (slot = 0; slot < 14; slot++) printk ("    ");
1117                 printk ("%2x:%016lx,%016lx\n",
1118                         0,
1119                         spitfire_get_itlb_tag(0), spitfire_get_itlb_data(0));
1120                 for (slot = 1; slot < 64; slot+=3) {
1121                         printk ("%2x:%016lx,%016lx %2x:%016lx,%016lx %2x:%016lx,%016lx\n", 
1122                                 slot,
1123                                 spitfire_get_itlb_tag(slot), spitfire_get_itlb_data(slot),
1124                                 slot+1,
1125                                 spitfire_get_itlb_tag(slot+1), spitfire_get_itlb_data(slot+1),
1126                                 slot+2,
1127                                 spitfire_get_itlb_tag(slot+2), spitfire_get_itlb_data(slot+2));
1128                 }
1129         } else if (tlb_type == cheetah || tlb_type == cheetah_plus) {
1130                 printk ("Contents of itlb0:\n");
1131                 for (slot = 0; slot < 16; slot+=2) {
1132                         printk ("%2x:%016lx,%016lx %2x:%016lx,%016lx\n",
1133                                 slot,
1134                                 cheetah_get_litlb_tag(slot), cheetah_get_litlb_data(slot),
1135                                 slot+1,
1136                                 cheetah_get_litlb_tag(slot+1), cheetah_get_litlb_data(slot+1));
1137                 }
1138                 printk ("Contents of itlb2:\n");
1139                 for (slot = 0; slot < 128; slot+=2) {
1140                         printk ("%2x:%016lx,%016lx %2x:%016lx,%016lx\n",
1141                                 slot,
1142                                 cheetah_get_itlb_tag(slot), cheetah_get_itlb_data(slot),
1143                                 slot+1,
1144                                 cheetah_get_itlb_tag(slot+1), cheetah_get_itlb_data(slot+1));
1145                 }
1146         }
1147 }
1148
1149 void sparc_ultra_dump_dtlb(void)
1150 {
1151         int slot;
1152
1153         if (tlb_type == spitfire) {
1154                 printk ("Contents of dtlb: ");
1155                 for (slot = 0; slot < 14; slot++) printk ("    ");
1156                 printk ("%2x:%016lx,%016lx\n", 0,
1157                         spitfire_get_dtlb_tag(0), spitfire_get_dtlb_data(0));
1158                 for (slot = 1; slot < 64; slot+=3) {
1159                         printk ("%2x:%016lx,%016lx %2x:%016lx,%016lx %2x:%016lx,%016lx\n", 
1160                                 slot,
1161                                 spitfire_get_dtlb_tag(slot), spitfire_get_dtlb_data(slot),
1162                                 slot+1,
1163                                 spitfire_get_dtlb_tag(slot+1), spitfire_get_dtlb_data(slot+1),
1164                                 slot+2,
1165                                 spitfire_get_dtlb_tag(slot+2), spitfire_get_dtlb_data(slot+2));
1166                 }
1167         } else if (tlb_type == cheetah || tlb_type == cheetah_plus) {
1168                 printk ("Contents of dtlb0:\n");
1169                 for (slot = 0; slot < 16; slot+=2) {
1170                         printk ("%2x:%016lx,%016lx %2x:%016lx,%016lx\n",
1171                                 slot,
1172                                 cheetah_get_ldtlb_tag(slot), cheetah_get_ldtlb_data(slot),
1173                                 slot+1,
1174                                 cheetah_get_ldtlb_tag(slot+1), cheetah_get_ldtlb_data(slot+1));
1175                 }
1176                 printk ("Contents of dtlb2:\n");
1177                 for (slot = 0; slot < 512; slot+=2) {
1178                         printk ("%2x:%016lx,%016lx %2x:%016lx,%016lx\n",
1179                                 slot,
1180                                 cheetah_get_dtlb_tag(slot, 2), cheetah_get_dtlb_data(slot, 2),
1181                                 slot+1,
1182                                 cheetah_get_dtlb_tag(slot+1, 2), cheetah_get_dtlb_data(slot+1, 2));
1183                 }
1184                 if (tlb_type == cheetah_plus) {
1185                         printk ("Contents of dtlb3:\n");
1186                         for (slot = 0; slot < 512; slot+=2) {
1187                                 printk ("%2x:%016lx,%016lx %2x:%016lx,%016lx\n",
1188                                         slot,
1189                                         cheetah_get_dtlb_tag(slot, 3), cheetah_get_dtlb_data(slot, 3),
1190                                         slot+1,
1191                                         cheetah_get_dtlb_tag(slot+1, 3), cheetah_get_dtlb_data(slot+1, 3));
1192                         }
1193                 }
1194         }
1195 }
1196
1197 extern unsigned long cmdline_memory_size;
1198
1199 unsigned long __init bootmem_init(unsigned long *pages_avail)
1200 {
1201         unsigned long bootmap_size, start_pfn, end_pfn;
1202         unsigned long end_of_phys_memory = 0UL;
1203         unsigned long bootmap_pfn, bytes_avail, size;
1204         int i;
1205
1206 #ifdef CONFIG_DEBUG_BOOTMEM
1207         prom_printf("bootmem_init: Scan pavail, ");
1208 #endif
1209
1210         bytes_avail = 0UL;
1211         for (i = 0; i < pavail_ents; i++) {
1212                 end_of_phys_memory = pavail[i].phys_addr +
1213                         pavail[i].reg_size;
1214                 bytes_avail += pavail[i].reg_size;
1215                 if (cmdline_memory_size) {
1216                         if (bytes_avail > cmdline_memory_size) {
1217                                 unsigned long slack = bytes_avail - cmdline_memory_size;
1218
1219                                 bytes_avail -= slack;
1220                                 end_of_phys_memory -= slack;
1221
1222                                 pavail[i].reg_size -= slack;
1223                                 if ((long)pavail[i].reg_size <= 0L) {
1224                                         pavail[i].phys_addr = 0xdeadbeefUL;
1225                                         pavail[i].reg_size = 0UL;
1226                                         pavail_ents = i;
1227                                 } else {
1228                                         pavail[i+1].reg_size = 0Ul;
1229                                         pavail[i+1].phys_addr = 0xdeadbeefUL;
1230                                         pavail_ents = i + 1;
1231                                 }
1232                                 break;
1233                         }
1234                 }
1235         }
1236
1237         *pages_avail = bytes_avail >> PAGE_SHIFT;
1238
1239         /* Start with page aligned address of last symbol in kernel
1240          * image.  The kernel is hard mapped below PAGE_OFFSET in a
1241          * 4MB locked TLB translation.
1242          */
1243         start_pfn = PAGE_ALIGN(kern_base + kern_size) >> PAGE_SHIFT;
1244
1245         bootmap_pfn = start_pfn;
1246
1247         end_pfn = end_of_phys_memory >> PAGE_SHIFT;
1248
1249 #ifdef CONFIG_BLK_DEV_INITRD
1250         /* Now have to check initial ramdisk, so that bootmap does not overwrite it */
1251         if (sparc_ramdisk_image || sparc_ramdisk_image64) {
1252                 unsigned long ramdisk_image = sparc_ramdisk_image ?
1253                         sparc_ramdisk_image : sparc_ramdisk_image64;
1254                 if (ramdisk_image >= (unsigned long)_end - 2 * PAGE_SIZE)
1255                         ramdisk_image -= KERNBASE;
1256                 initrd_start = ramdisk_image + phys_base;
1257                 initrd_end = initrd_start + sparc_ramdisk_size;
1258                 if (initrd_end > end_of_phys_memory) {
1259                         printk(KERN_CRIT "initrd extends beyond end of memory "
1260                                          "(0x%016lx > 0x%016lx)\ndisabling initrd\n",
1261                                initrd_end, end_of_phys_memory);
1262                         initrd_start = 0;
1263                 }
1264                 if (initrd_start) {
1265                         if (initrd_start >= (start_pfn << PAGE_SHIFT) &&
1266                             initrd_start < (start_pfn << PAGE_SHIFT) + 2 * PAGE_SIZE)
1267                                 bootmap_pfn = PAGE_ALIGN (initrd_end) >> PAGE_SHIFT;
1268                 }
1269         }
1270 #endif  
1271         /* Initialize the boot-time allocator. */
1272         max_pfn = max_low_pfn = end_pfn;
1273         min_low_pfn = pfn_base;
1274
1275 #ifdef CONFIG_DEBUG_BOOTMEM
1276         prom_printf("init_bootmem(min[%lx], bootmap[%lx], max[%lx])\n",
1277                     min_low_pfn, bootmap_pfn, max_low_pfn);
1278 #endif
1279         bootmap_size = init_bootmem_node(NODE_DATA(0), bootmap_pfn, pfn_base, end_pfn);
1280
1281         /* Now register the available physical memory with the
1282          * allocator.
1283          */
1284         for (i = 0; i < pavail_ents; i++) {
1285 #ifdef CONFIG_DEBUG_BOOTMEM
1286                 prom_printf("free_bootmem(pavail:%d): base[%lx] size[%lx]\n",
1287                             i, pavail[i].phys_addr, pavail[i].reg_size);
1288 #endif
1289                 free_bootmem(pavail[i].phys_addr, pavail[i].reg_size);
1290         }
1291
1292 #ifdef CONFIG_BLK_DEV_INITRD
1293         if (initrd_start) {
1294                 size = initrd_end - initrd_start;
1295
1296                 /* Resert the initrd image area. */
1297 #ifdef CONFIG_DEBUG_BOOTMEM
1298                 prom_printf("reserve_bootmem(initrd): base[%llx] size[%lx]\n",
1299                         initrd_start, initrd_end);
1300 #endif
1301                 reserve_bootmem(initrd_start, size);
1302                 *pages_avail -= PAGE_ALIGN(size) >> PAGE_SHIFT;
1303
1304                 initrd_start += PAGE_OFFSET;
1305                 initrd_end += PAGE_OFFSET;
1306         }
1307 #endif
1308         /* Reserve the kernel text/data/bss. */
1309 #ifdef CONFIG_DEBUG_BOOTMEM
1310         prom_printf("reserve_bootmem(kernel): base[%lx] size[%lx]\n", kern_base, kern_size);
1311 #endif
1312         reserve_bootmem(kern_base, kern_size);
1313         *pages_avail -= PAGE_ALIGN(kern_size) >> PAGE_SHIFT;
1314
1315         /* Reserve the bootmem map.   We do not account for it
1316          * in pages_avail because we will release that memory
1317          * in free_all_bootmem.
1318          */
1319         size = bootmap_size;
1320 #ifdef CONFIG_DEBUG_BOOTMEM
1321         prom_printf("reserve_bootmem(bootmap): base[%lx] size[%lx]\n",
1322                     (bootmap_pfn << PAGE_SHIFT), size);
1323 #endif
1324         reserve_bootmem((bootmap_pfn << PAGE_SHIFT), size);
1325         *pages_avail -= PAGE_ALIGN(size) >> PAGE_SHIFT;
1326
1327         return end_pfn;
1328 }
1329
1330 #ifdef CONFIG_DEBUG_PAGEALLOC
1331 static unsigned long kernel_map_range(unsigned long pstart, unsigned long pend, pgprot_t prot)
1332 {
1333         unsigned long vstart = PAGE_OFFSET + pstart;
1334         unsigned long vend = PAGE_OFFSET + pend;
1335         unsigned long alloc_bytes = 0UL;
1336
1337         if ((vstart & ~PAGE_MASK) || (vend & ~PAGE_MASK)) {
1338                 prom_printf("kernel_map: Unaligned physmem[%lx:%lx]\n",
1339                             vstart, vend);
1340                 prom_halt();
1341         }
1342
1343         while (vstart < vend) {
1344                 unsigned long this_end, paddr = __pa(vstart);
1345                 pgd_t *pgd = pgd_offset_k(vstart);
1346                 pud_t *pud;
1347                 pmd_t *pmd;
1348                 pte_t *pte;
1349
1350                 pud = pud_offset(pgd, vstart);
1351                 if (pud_none(*pud)) {
1352                         pmd_t *new;
1353
1354                         new = __alloc_bootmem(PAGE_SIZE, PAGE_SIZE, PAGE_SIZE);
1355                         alloc_bytes += PAGE_SIZE;
1356                         pud_populate(&init_mm, pud, new);
1357                 }
1358
1359                 pmd = pmd_offset(pud, vstart);
1360                 if (!pmd_present(*pmd)) {
1361                         pte_t *new;
1362
1363                         new = __alloc_bootmem(PAGE_SIZE, PAGE_SIZE, PAGE_SIZE);
1364                         alloc_bytes += PAGE_SIZE;
1365                         pmd_populate_kernel(&init_mm, pmd, new);
1366                 }
1367
1368                 pte = pte_offset_kernel(pmd, vstart);
1369                 this_end = (vstart + PMD_SIZE) & PMD_MASK;
1370                 if (this_end > vend)
1371                         this_end = vend;
1372
1373                 while (vstart < this_end) {
1374                         pte_val(*pte) = (paddr | pgprot_val(prot));
1375
1376                         vstart += PAGE_SIZE;
1377                         paddr += PAGE_SIZE;
1378                         pte++;
1379                 }
1380         }
1381
1382         return alloc_bytes;
1383 }
1384
1385 static struct linux_prom64_registers pall[MAX_BANKS] __initdata;
1386 static int pall_ents __initdata;
1387
1388 extern unsigned int kvmap_linear_patch[1];
1389
1390 static void __init kernel_physical_mapping_init(void)
1391 {
1392         unsigned long i, mem_alloced = 0UL;
1393
1394         read_obp_memory("reg", &pall[0], &pall_ents);
1395
1396         for (i = 0; i < pall_ents; i++) {
1397                 unsigned long phys_start, phys_end;
1398
1399                 phys_start = pall[i].phys_addr;
1400                 phys_end = phys_start + pall[i].reg_size;
1401                 mem_alloced += kernel_map_range(phys_start, phys_end,
1402                                                 PAGE_KERNEL);
1403         }
1404
1405         printk("Allocated %ld bytes for kernel page tables.\n",
1406                mem_alloced);
1407
1408         kvmap_linear_patch[0] = 0x01000000; /* nop */
1409         flushi(&kvmap_linear_patch[0]);
1410
1411         __flush_tlb_all();
1412 }
1413
1414 void kernel_map_pages(struct page *page, int numpages, int enable)
1415 {
1416         unsigned long phys_start = page_to_pfn(page) << PAGE_SHIFT;
1417         unsigned long phys_end = phys_start + (numpages * PAGE_SIZE);
1418
1419         kernel_map_range(phys_start, phys_end,
1420                          (enable ? PAGE_KERNEL : __pgprot(0)));
1421
1422         /* we should perform an IPI and flush all tlbs,
1423          * but that can deadlock->flush only current cpu.
1424          */
1425         __flush_tlb_kernel_range(PAGE_OFFSET + phys_start,
1426                                  PAGE_OFFSET + phys_end);
1427 }
1428 #endif
1429
1430 unsigned long __init find_ecache_flush_span(unsigned long size)
1431 {
1432         int i;
1433
1434         for (i = 0; i < pavail_ents; i++) {
1435                 if (pavail[i].reg_size >= size)
1436                         return pavail[i].phys_addr;
1437         }
1438
1439         return ~0UL;
1440 }
1441
1442 /* paging_init() sets up the page tables */
1443
1444 extern void cheetah_ecache_flush_init(void);
1445
1446 static unsigned long last_valid_pfn;
1447 pgd_t swapper_pg_dir[2048];
1448
1449 void __init paging_init(void)
1450 {
1451         unsigned long end_pfn, pages_avail, shift;
1452         unsigned long real_end, i;
1453
1454         /* Find available physical memory... */
1455         read_obp_memory("available", &pavail[0], &pavail_ents);
1456
1457         phys_base = 0xffffffffffffffffUL;
1458         for (i = 0; i < pavail_ents; i++)
1459                 phys_base = min(phys_base, pavail[i].phys_addr);
1460
1461         pfn_base = phys_base >> PAGE_SHIFT;
1462
1463         kern_base = (prom_boot_mapping_phys_low >> 22UL) << 22UL;
1464         kern_size = (unsigned long)&_end - (unsigned long)KERNBASE;
1465
1466         set_bit(0, mmu_context_bmap);
1467
1468         shift = kern_base + PAGE_OFFSET - ((unsigned long)KERNBASE);
1469
1470         real_end = (unsigned long)_end;
1471         if ((real_end > ((unsigned long)KERNBASE + 0x400000)))
1472                 bigkernel = 1;
1473         if ((real_end > ((unsigned long)KERNBASE + 0x800000))) {
1474                 prom_printf("paging_init: Kernel > 8MB, too large.\n");
1475                 prom_halt();
1476         }
1477
1478         /* Set kernel pgd to upper alias so physical page computations
1479          * work.
1480          */
1481         init_mm.pgd += ((shift) / (sizeof(pgd_t)));
1482         
1483         memset(swapper_low_pmd_dir, 0, sizeof(swapper_low_pmd_dir));
1484
1485         /* Now can init the kernel/bad page tables. */
1486         pud_set(pud_offset(&swapper_pg_dir[0], 0),
1487                 swapper_low_pmd_dir + (shift / sizeof(pgd_t)));
1488         
1489         swapper_pgd_zero = pgd_val(swapper_pg_dir[0]);
1490         
1491         inherit_prom_mappings();
1492         
1493         /* Ok, we can use our TLB miss and window trap handlers safely.
1494          * We need to do a quick peek here to see if we are on StarFire
1495          * or not, so setup_tba can setup the IRQ globals correctly (it
1496          * needs to get the hard smp processor id correctly).
1497          */
1498         {
1499                 extern void setup_tba(int);
1500                 setup_tba(this_is_starfire);
1501         }
1502
1503         inherit_locked_prom_mappings(1);
1504
1505         __flush_tlb_all();
1506
1507         /* Setup bootmem... */
1508         pages_avail = 0;
1509         last_valid_pfn = end_pfn = bootmem_init(&pages_avail);
1510
1511 #ifdef CONFIG_DEBUG_PAGEALLOC
1512         kernel_physical_mapping_init();
1513 #endif
1514
1515         {
1516                 unsigned long zones_size[MAX_NR_ZONES];
1517                 unsigned long zholes_size[MAX_NR_ZONES];
1518                 unsigned long npages;
1519                 int znum;
1520
1521                 for (znum = 0; znum < MAX_NR_ZONES; znum++)
1522                         zones_size[znum] = zholes_size[znum] = 0;
1523
1524                 npages = end_pfn - pfn_base;
1525                 zones_size[ZONE_DMA] = npages;
1526                 zholes_size[ZONE_DMA] = npages - pages_avail;
1527
1528                 free_area_init_node(0, &contig_page_data, zones_size,
1529                                     phys_base >> PAGE_SHIFT, zholes_size);
1530         }
1531
1532         device_scan();
1533 }
1534
1535 static void __init taint_real_pages(void)
1536 {
1537         int i;
1538
1539         read_obp_memory("available", &pavail_rescan[0], &pavail_rescan_ents);
1540
1541         /* Find changes discovered in the physmem available rescan and
1542          * reserve the lost portions in the bootmem maps.
1543          */
1544         for (i = 0; i < pavail_ents; i++) {
1545                 unsigned long old_start, old_end;
1546
1547                 old_start = pavail[i].phys_addr;
1548                 old_end = old_start +
1549                         pavail[i].reg_size;
1550                 while (old_start < old_end) {
1551                         int n;
1552
1553                         for (n = 0; pavail_rescan_ents; n++) {
1554                                 unsigned long new_start, new_end;
1555
1556                                 new_start = pavail_rescan[n].phys_addr;
1557                                 new_end = new_start +
1558                                         pavail_rescan[n].reg_size;
1559
1560                                 if (new_start <= old_start &&
1561                                     new_end >= (old_start + PAGE_SIZE)) {
1562                                         set_bit(old_start >> 22,
1563                                                 sparc64_valid_addr_bitmap);
1564                                         goto do_next_page;
1565                                 }
1566                         }
1567                         reserve_bootmem(old_start, PAGE_SIZE);
1568
1569                 do_next_page:
1570                         old_start += PAGE_SIZE;
1571                 }
1572         }
1573 }
1574
1575 void __init mem_init(void)
1576 {
1577         unsigned long codepages, datapages, initpages;
1578         unsigned long addr, last;
1579         int i;
1580
1581         i = last_valid_pfn >> ((22 - PAGE_SHIFT) + 6);
1582         i += 1;
1583         sparc64_valid_addr_bitmap = (unsigned long *) alloc_bootmem(i << 3);
1584         if (sparc64_valid_addr_bitmap == NULL) {
1585                 prom_printf("mem_init: Cannot alloc valid_addr_bitmap.\n");
1586                 prom_halt();
1587         }
1588         memset(sparc64_valid_addr_bitmap, 0, i << 3);
1589
1590         addr = PAGE_OFFSET + kern_base;
1591         last = PAGE_ALIGN(kern_size) + addr;
1592         while (addr < last) {
1593                 set_bit(__pa(addr) >> 22, sparc64_valid_addr_bitmap);
1594                 addr += PAGE_SIZE;
1595         }
1596
1597         taint_real_pages();
1598
1599         max_mapnr = last_valid_pfn - pfn_base;
1600         high_memory = __va(last_valid_pfn << PAGE_SHIFT);
1601
1602 #ifdef CONFIG_DEBUG_BOOTMEM
1603         prom_printf("mem_init: Calling free_all_bootmem().\n");
1604 #endif
1605         totalram_pages = num_physpages = free_all_bootmem() - 1;
1606
1607         /*
1608          * Set up the zero page, mark it reserved, so that page count
1609          * is not manipulated when freeing the page from user ptes.
1610          */
1611         mem_map_zero = alloc_pages(GFP_KERNEL|__GFP_ZERO, 0);
1612         if (mem_map_zero == NULL) {
1613                 prom_printf("paging_init: Cannot alloc zero page.\n");
1614                 prom_halt();
1615         }
1616         SetPageReserved(mem_map_zero);
1617
1618         codepages = (((unsigned long) _etext) - ((unsigned long) _start));
1619         codepages = PAGE_ALIGN(codepages) >> PAGE_SHIFT;
1620         datapages = (((unsigned long) _edata) - ((unsigned long) _etext));
1621         datapages = PAGE_ALIGN(datapages) >> PAGE_SHIFT;
1622         initpages = (((unsigned long) __init_end) - ((unsigned long) __init_begin));
1623         initpages = PAGE_ALIGN(initpages) >> PAGE_SHIFT;
1624
1625         printk("Memory: %uk available (%ldk kernel code, %ldk data, %ldk init) [%016lx,%016lx]\n",
1626                nr_free_pages() << (PAGE_SHIFT-10),
1627                codepages << (PAGE_SHIFT-10),
1628                datapages << (PAGE_SHIFT-10), 
1629                initpages << (PAGE_SHIFT-10), 
1630                PAGE_OFFSET, (last_valid_pfn << PAGE_SHIFT));
1631
1632         if (tlb_type == cheetah || tlb_type == cheetah_plus)
1633                 cheetah_ecache_flush_init();
1634 }
1635
1636 void free_initmem(void)
1637 {
1638         unsigned long addr, initend;
1639
1640         /*
1641          * The init section is aligned to 8k in vmlinux.lds. Page align for >8k pagesizes.
1642          */
1643         addr = PAGE_ALIGN((unsigned long)(__init_begin));
1644         initend = (unsigned long)(__init_end) & PAGE_MASK;
1645         for (; addr < initend; addr += PAGE_SIZE) {
1646                 unsigned long page;
1647                 struct page *p;
1648
1649                 page = (addr +
1650                         ((unsigned long) __va(kern_base)) -
1651                         ((unsigned long) KERNBASE));
1652                 memset((void *)addr, 0xcc, PAGE_SIZE);
1653                 p = virt_to_page(page);
1654
1655                 ClearPageReserved(p);
1656                 set_page_count(p, 1);
1657                 __free_page(p);
1658                 num_physpages++;
1659                 totalram_pages++;
1660         }
1661 }
1662
1663 #ifdef CONFIG_BLK_DEV_INITRD
1664 void free_initrd_mem(unsigned long start, unsigned long end)
1665 {
1666         if (start < end)
1667                 printk ("Freeing initrd memory: %ldk freed\n", (end - start) >> 10);
1668         for (; start < end; start += PAGE_SIZE) {
1669                 struct page *p = virt_to_page(start);
1670
1671                 ClearPageReserved(p);
1672                 set_page_count(p, 1);
1673                 __free_page(p);
1674                 num_physpages++;
1675                 totalram_pages++;
1676         }
1677 }
1678 #endif