x86: Do not try to disable hpet if it hasn't been initialized before
[sfrench/cifs-2.6.git] / arch / powerpc / mm / hash_utils_64.c
1 /*
2  * PowerPC64 port by Mike Corrigan and Dave Engebretsen
3  *   {mikejc|engebret}@us.ibm.com
4  *
5  *    Copyright (c) 2000 Mike Corrigan <mikejc@us.ibm.com>
6  *
7  * SMP scalability work:
8  *    Copyright (C) 2001 Anton Blanchard <anton@au.ibm.com>, IBM
9  * 
10  *    Module name: htab.c
11  *
12  *    Description:
13  *      PowerPC Hashed Page Table functions
14  *
15  * This program is free software; you can redistribute it and/or
16  * modify it under the terms of the GNU General Public License
17  * as published by the Free Software Foundation; either version
18  * 2 of the License, or (at your option) any later version.
19  */
20
21 #undef DEBUG
22 #undef DEBUG_LOW
23
24 #include <linux/spinlock.h>
25 #include <linux/errno.h>
26 #include <linux/sched.h>
27 #include <linux/proc_fs.h>
28 #include <linux/stat.h>
29 #include <linux/sysctl.h>
30 #include <linux/ctype.h>
31 #include <linux/cache.h>
32 #include <linux/init.h>
33 #include <linux/signal.h>
34 #include <linux/memblock.h>
35
36 #include <asm/processor.h>
37 #include <asm/pgtable.h>
38 #include <asm/mmu.h>
39 #include <asm/mmu_context.h>
40 #include <asm/page.h>
41 #include <asm/types.h>
42 #include <asm/system.h>
43 #include <asm/uaccess.h>
44 #include <asm/machdep.h>
45 #include <asm/prom.h>
46 #include <asm/abs_addr.h>
47 #include <asm/tlbflush.h>
48 #include <asm/io.h>
49 #include <asm/eeh.h>
50 #include <asm/tlb.h>
51 #include <asm/cacheflush.h>
52 #include <asm/cputable.h>
53 #include <asm/sections.h>
54 #include <asm/spu.h>
55 #include <asm/udbg.h>
56
57 #ifdef DEBUG
58 #define DBG(fmt...) udbg_printf(fmt)
59 #else
60 #define DBG(fmt...)
61 #endif
62
63 #ifdef DEBUG_LOW
64 #define DBG_LOW(fmt...) udbg_printf(fmt)
65 #else
66 #define DBG_LOW(fmt...)
67 #endif
68
69 #define KB (1024)
70 #define MB (1024*KB)
71 #define GB (1024L*MB)
72
73 /*
74  * Note:  pte   --> Linux PTE
75  *        HPTE  --> PowerPC Hashed Page Table Entry
76  *
77  * Execution context:
78  *   htab_initialize is called with the MMU off (of course), but
79  *   the kernel has been copied down to zero so it can directly
80  *   reference global data.  At this point it is very difficult
81  *   to print debug info.
82  *
83  */
84
85 #ifdef CONFIG_U3_DART
86 extern unsigned long dart_tablebase;
87 #endif /* CONFIG_U3_DART */
88
89 static unsigned long _SDR1;
90 struct mmu_psize_def mmu_psize_defs[MMU_PAGE_COUNT];
91
92 struct hash_pte *htab_address;
93 unsigned long htab_size_bytes;
94 unsigned long htab_hash_mask;
95 EXPORT_SYMBOL_GPL(htab_hash_mask);
96 int mmu_linear_psize = MMU_PAGE_4K;
97 int mmu_virtual_psize = MMU_PAGE_4K;
98 int mmu_vmalloc_psize = MMU_PAGE_4K;
99 #ifdef CONFIG_SPARSEMEM_VMEMMAP
100 int mmu_vmemmap_psize = MMU_PAGE_4K;
101 #endif
102 int mmu_io_psize = MMU_PAGE_4K;
103 int mmu_kernel_ssize = MMU_SEGSIZE_256M;
104 int mmu_highuser_ssize = MMU_SEGSIZE_256M;
105 u16 mmu_slb_size = 64;
106 EXPORT_SYMBOL_GPL(mmu_slb_size);
107 #ifdef CONFIG_HUGETLB_PAGE
108 unsigned int HPAGE_SHIFT;
109 #endif
110 #ifdef CONFIG_PPC_64K_PAGES
111 int mmu_ci_restrictions;
112 #endif
113 #ifdef CONFIG_DEBUG_PAGEALLOC
114 static u8 *linear_map_hash_slots;
115 static unsigned long linear_map_hash_count;
116 static DEFINE_SPINLOCK(linear_map_hash_lock);
117 #endif /* CONFIG_DEBUG_PAGEALLOC */
118
119 /* There are definitions of page sizes arrays to be used when none
120  * is provided by the firmware.
121  */
122
123 /* Pre-POWER4 CPUs (4k pages only)
124  */
125 static struct mmu_psize_def mmu_psize_defaults_old[] = {
126         [MMU_PAGE_4K] = {
127                 .shift  = 12,
128                 .sllp   = 0,
129                 .penc   = 0,
130                 .avpnm  = 0,
131                 .tlbiel = 0,
132         },
133 };
134
135 /* POWER4, GPUL, POWER5
136  *
137  * Support for 16Mb large pages
138  */
139 static struct mmu_psize_def mmu_psize_defaults_gp[] = {
140         [MMU_PAGE_4K] = {
141                 .shift  = 12,
142                 .sllp   = 0,
143                 .penc   = 0,
144                 .avpnm  = 0,
145                 .tlbiel = 1,
146         },
147         [MMU_PAGE_16M] = {
148                 .shift  = 24,
149                 .sllp   = SLB_VSID_L,
150                 .penc   = 0,
151                 .avpnm  = 0x1UL,
152                 .tlbiel = 0,
153         },
154 };
155
156 static unsigned long htab_convert_pte_flags(unsigned long pteflags)
157 {
158         unsigned long rflags = pteflags & 0x1fa;
159
160         /* _PAGE_EXEC -> NOEXEC */
161         if ((pteflags & _PAGE_EXEC) == 0)
162                 rflags |= HPTE_R_N;
163
164         /* PP bits. PAGE_USER is already PP bit 0x2, so we only
165          * need to add in 0x1 if it's a read-only user page
166          */
167         if ((pteflags & _PAGE_USER) && !((pteflags & _PAGE_RW) &&
168                                          (pteflags & _PAGE_DIRTY)))
169                 rflags |= 1;
170
171         /* Always add C */
172         return rflags | HPTE_R_C;
173 }
174
175 int htab_bolt_mapping(unsigned long vstart, unsigned long vend,
176                       unsigned long pstart, unsigned long prot,
177                       int psize, int ssize)
178 {
179         unsigned long vaddr, paddr;
180         unsigned int step, shift;
181         int ret = 0;
182
183         shift = mmu_psize_defs[psize].shift;
184         step = 1 << shift;
185
186         prot = htab_convert_pte_flags(prot);
187
188         DBG("htab_bolt_mapping(%lx..%lx -> %lx (%lx,%d,%d)\n",
189             vstart, vend, pstart, prot, psize, ssize);
190
191         for (vaddr = vstart, paddr = pstart; vaddr < vend;
192              vaddr += step, paddr += step) {
193                 unsigned long hash, hpteg;
194                 unsigned long vsid = get_kernel_vsid(vaddr, ssize);
195                 unsigned long va = hpt_va(vaddr, vsid, ssize);
196                 unsigned long tprot = prot;
197
198                 /* Make kernel text executable */
199                 if (overlaps_kernel_text(vaddr, vaddr + step))
200                         tprot &= ~HPTE_R_N;
201
202                 hash = hpt_hash(va, shift, ssize);
203                 hpteg = ((hash & htab_hash_mask) * HPTES_PER_GROUP);
204
205                 BUG_ON(!ppc_md.hpte_insert);
206                 ret = ppc_md.hpte_insert(hpteg, va, paddr, tprot,
207                                          HPTE_V_BOLTED, psize, ssize);
208
209                 if (ret < 0)
210                         break;
211 #ifdef CONFIG_DEBUG_PAGEALLOC
212                 if ((paddr >> PAGE_SHIFT) < linear_map_hash_count)
213                         linear_map_hash_slots[paddr >> PAGE_SHIFT] = ret | 0x80;
214 #endif /* CONFIG_DEBUG_PAGEALLOC */
215         }
216         return ret < 0 ? ret : 0;
217 }
218
219 #ifdef CONFIG_MEMORY_HOTPLUG
220 static int htab_remove_mapping(unsigned long vstart, unsigned long vend,
221                       int psize, int ssize)
222 {
223         unsigned long vaddr;
224         unsigned int step, shift;
225
226         shift = mmu_psize_defs[psize].shift;
227         step = 1 << shift;
228
229         if (!ppc_md.hpte_removebolted) {
230                 printk(KERN_WARNING "Platform doesn't implement "
231                                 "hpte_removebolted\n");
232                 return -EINVAL;
233         }
234
235         for (vaddr = vstart; vaddr < vend; vaddr += step)
236                 ppc_md.hpte_removebolted(vaddr, psize, ssize);
237
238         return 0;
239 }
240 #endif /* CONFIG_MEMORY_HOTPLUG */
241
242 static int __init htab_dt_scan_seg_sizes(unsigned long node,
243                                          const char *uname, int depth,
244                                          void *data)
245 {
246         char *type = of_get_flat_dt_prop(node, "device_type", NULL);
247         u32 *prop;
248         unsigned long size = 0;
249
250         /* We are scanning "cpu" nodes only */
251         if (type == NULL || strcmp(type, "cpu") != 0)
252                 return 0;
253
254         prop = (u32 *)of_get_flat_dt_prop(node, "ibm,processor-segment-sizes",
255                                           &size);
256         if (prop == NULL)
257                 return 0;
258         for (; size >= 4; size -= 4, ++prop) {
259                 if (prop[0] == 40) {
260                         DBG("1T segment support detected\n");
261                         cur_cpu_spec->cpu_features |= CPU_FTR_1T_SEGMENT;
262                         return 1;
263                 }
264         }
265         cur_cpu_spec->cpu_features &= ~CPU_FTR_NO_SLBIE_B;
266         return 0;
267 }
268
269 static void __init htab_init_seg_sizes(void)
270 {
271         of_scan_flat_dt(htab_dt_scan_seg_sizes, NULL);
272 }
273
274 static int __init htab_dt_scan_page_sizes(unsigned long node,
275                                           const char *uname, int depth,
276                                           void *data)
277 {
278         char *type = of_get_flat_dt_prop(node, "device_type", NULL);
279         u32 *prop;
280         unsigned long size = 0;
281
282         /* We are scanning "cpu" nodes only */
283         if (type == NULL || strcmp(type, "cpu") != 0)
284                 return 0;
285
286         prop = (u32 *)of_get_flat_dt_prop(node,
287                                           "ibm,segment-page-sizes", &size);
288         if (prop != NULL) {
289                 DBG("Page sizes from device-tree:\n");
290                 size /= 4;
291                 cur_cpu_spec->cpu_features &= ~(CPU_FTR_16M_PAGE);
292                 while(size > 0) {
293                         unsigned int shift = prop[0];
294                         unsigned int slbenc = prop[1];
295                         unsigned int lpnum = prop[2];
296                         unsigned int lpenc = 0;
297                         struct mmu_psize_def *def;
298                         int idx = -1;
299
300                         size -= 3; prop += 3;
301                         while(size > 0 && lpnum) {
302                                 if (prop[0] == shift)
303                                         lpenc = prop[1];
304                                 prop += 2; size -= 2;
305                                 lpnum--;
306                         }
307                         switch(shift) {
308                         case 0xc:
309                                 idx = MMU_PAGE_4K;
310                                 break;
311                         case 0x10:
312                                 idx = MMU_PAGE_64K;
313                                 break;
314                         case 0x14:
315                                 idx = MMU_PAGE_1M;
316                                 break;
317                         case 0x18:
318                                 idx = MMU_PAGE_16M;
319                                 cur_cpu_spec->cpu_features |= CPU_FTR_16M_PAGE;
320                                 break;
321                         case 0x22:
322                                 idx = MMU_PAGE_16G;
323                                 break;
324                         }
325                         if (idx < 0)
326                                 continue;
327                         def = &mmu_psize_defs[idx];
328                         def->shift = shift;
329                         if (shift <= 23)
330                                 def->avpnm = 0;
331                         else
332                                 def->avpnm = (1 << (shift - 23)) - 1;
333                         def->sllp = slbenc;
334                         def->penc = lpenc;
335                         /* We don't know for sure what's up with tlbiel, so
336                          * for now we only set it for 4K and 64K pages
337                          */
338                         if (idx == MMU_PAGE_4K || idx == MMU_PAGE_64K)
339                                 def->tlbiel = 1;
340                         else
341                                 def->tlbiel = 0;
342
343                         DBG(" %d: shift=%02x, sllp=%04lx, avpnm=%08lx, "
344                             "tlbiel=%d, penc=%d\n",
345                             idx, shift, def->sllp, def->avpnm, def->tlbiel,
346                             def->penc);
347                 }
348                 return 1;
349         }
350         return 0;
351 }
352
353 #ifdef CONFIG_HUGETLB_PAGE
354 /* Scan for 16G memory blocks that have been set aside for huge pages
355  * and reserve those blocks for 16G huge pages.
356  */
357 static int __init htab_dt_scan_hugepage_blocks(unsigned long node,
358                                         const char *uname, int depth,
359                                         void *data) {
360         char *type = of_get_flat_dt_prop(node, "device_type", NULL);
361         unsigned long *addr_prop;
362         u32 *page_count_prop;
363         unsigned int expected_pages;
364         long unsigned int phys_addr;
365         long unsigned int block_size;
366
367         /* We are scanning "memory" nodes only */
368         if (type == NULL || strcmp(type, "memory") != 0)
369                 return 0;
370
371         /* This property is the log base 2 of the number of virtual pages that
372          * will represent this memory block. */
373         page_count_prop = of_get_flat_dt_prop(node, "ibm,expected#pages", NULL);
374         if (page_count_prop == NULL)
375                 return 0;
376         expected_pages = (1 << page_count_prop[0]);
377         addr_prop = of_get_flat_dt_prop(node, "reg", NULL);
378         if (addr_prop == NULL)
379                 return 0;
380         phys_addr = addr_prop[0];
381         block_size = addr_prop[1];
382         if (block_size != (16 * GB))
383                 return 0;
384         printk(KERN_INFO "Huge page(16GB) memory: "
385                         "addr = 0x%lX size = 0x%lX pages = %d\n",
386                         phys_addr, block_size, expected_pages);
387         if (phys_addr + (16 * GB) <= memblock_end_of_DRAM()) {
388                 memblock_reserve(phys_addr, block_size * expected_pages);
389                 add_gpage(phys_addr, block_size, expected_pages);
390         }
391         return 0;
392 }
393 #endif /* CONFIG_HUGETLB_PAGE */
394
395 static void __init htab_init_page_sizes(void)
396 {
397         int rc;
398
399         /* Default to 4K pages only */
400         memcpy(mmu_psize_defs, mmu_psize_defaults_old,
401                sizeof(mmu_psize_defaults_old));
402
403         /*
404          * Try to find the available page sizes in the device-tree
405          */
406         rc = of_scan_flat_dt(htab_dt_scan_page_sizes, NULL);
407         if (rc != 0)  /* Found */
408                 goto found;
409
410         /*
411          * Not in the device-tree, let's fallback on known size
412          * list for 16M capable GP & GR
413          */
414         if (cpu_has_feature(CPU_FTR_16M_PAGE))
415                 memcpy(mmu_psize_defs, mmu_psize_defaults_gp,
416                        sizeof(mmu_psize_defaults_gp));
417  found:
418 #ifndef CONFIG_DEBUG_PAGEALLOC
419         /*
420          * Pick a size for the linear mapping. Currently, we only support
421          * 16M, 1M and 4K which is the default
422          */
423         if (mmu_psize_defs[MMU_PAGE_16M].shift)
424                 mmu_linear_psize = MMU_PAGE_16M;
425         else if (mmu_psize_defs[MMU_PAGE_1M].shift)
426                 mmu_linear_psize = MMU_PAGE_1M;
427 #endif /* CONFIG_DEBUG_PAGEALLOC */
428
429 #ifdef CONFIG_PPC_64K_PAGES
430         /*
431          * Pick a size for the ordinary pages. Default is 4K, we support
432          * 64K for user mappings and vmalloc if supported by the processor.
433          * We only use 64k for ioremap if the processor
434          * (and firmware) support cache-inhibited large pages.
435          * If not, we use 4k and set mmu_ci_restrictions so that
436          * hash_page knows to switch processes that use cache-inhibited
437          * mappings to 4k pages.
438          */
439         if (mmu_psize_defs[MMU_PAGE_64K].shift) {
440                 mmu_virtual_psize = MMU_PAGE_64K;
441                 mmu_vmalloc_psize = MMU_PAGE_64K;
442                 if (mmu_linear_psize == MMU_PAGE_4K)
443                         mmu_linear_psize = MMU_PAGE_64K;
444                 if (cpu_has_feature(CPU_FTR_CI_LARGE_PAGE)) {
445                         /*
446                          * Don't use 64k pages for ioremap on pSeries, since
447                          * that would stop us accessing the HEA ethernet.
448                          */
449                         if (!machine_is(pseries))
450                                 mmu_io_psize = MMU_PAGE_64K;
451                 } else
452                         mmu_ci_restrictions = 1;
453         }
454 #endif /* CONFIG_PPC_64K_PAGES */
455
456 #ifdef CONFIG_SPARSEMEM_VMEMMAP
457         /* We try to use 16M pages for vmemmap if that is supported
458          * and we have at least 1G of RAM at boot
459          */
460         if (mmu_psize_defs[MMU_PAGE_16M].shift &&
461             memblock_phys_mem_size() >= 0x40000000)
462                 mmu_vmemmap_psize = MMU_PAGE_16M;
463         else if (mmu_psize_defs[MMU_PAGE_64K].shift)
464                 mmu_vmemmap_psize = MMU_PAGE_64K;
465         else
466                 mmu_vmemmap_psize = MMU_PAGE_4K;
467 #endif /* CONFIG_SPARSEMEM_VMEMMAP */
468
469         printk(KERN_DEBUG "Page orders: linear mapping = %d, "
470                "virtual = %d, io = %d"
471 #ifdef CONFIG_SPARSEMEM_VMEMMAP
472                ", vmemmap = %d"
473 #endif
474                "\n",
475                mmu_psize_defs[mmu_linear_psize].shift,
476                mmu_psize_defs[mmu_virtual_psize].shift,
477                mmu_psize_defs[mmu_io_psize].shift
478 #ifdef CONFIG_SPARSEMEM_VMEMMAP
479                ,mmu_psize_defs[mmu_vmemmap_psize].shift
480 #endif
481                );
482
483 #ifdef CONFIG_HUGETLB_PAGE
484         /* Reserve 16G huge page memory sections for huge pages */
485         of_scan_flat_dt(htab_dt_scan_hugepage_blocks, NULL);
486 #endif /* CONFIG_HUGETLB_PAGE */
487 }
488
489 static int __init htab_dt_scan_pftsize(unsigned long node,
490                                        const char *uname, int depth,
491                                        void *data)
492 {
493         char *type = of_get_flat_dt_prop(node, "device_type", NULL);
494         u32 *prop;
495
496         /* We are scanning "cpu" nodes only */
497         if (type == NULL || strcmp(type, "cpu") != 0)
498                 return 0;
499
500         prop = (u32 *)of_get_flat_dt_prop(node, "ibm,pft-size", NULL);
501         if (prop != NULL) {
502                 /* pft_size[0] is the NUMA CEC cookie */
503                 ppc64_pft_size = prop[1];
504                 return 1;
505         }
506         return 0;
507 }
508
509 static unsigned long __init htab_get_table_size(void)
510 {
511         unsigned long mem_size, rnd_mem_size, pteg_count, psize;
512
513         /* If hash size isn't already provided by the platform, we try to
514          * retrieve it from the device-tree. If it's not there neither, we
515          * calculate it now based on the total RAM size
516          */
517         if (ppc64_pft_size == 0)
518                 of_scan_flat_dt(htab_dt_scan_pftsize, NULL);
519         if (ppc64_pft_size)
520                 return 1UL << ppc64_pft_size;
521
522         /* round mem_size up to next power of 2 */
523         mem_size = memblock_phys_mem_size();
524         rnd_mem_size = 1UL << __ilog2(mem_size);
525         if (rnd_mem_size < mem_size)
526                 rnd_mem_size <<= 1;
527
528         /* # pages / 2 */
529         psize = mmu_psize_defs[mmu_virtual_psize].shift;
530         pteg_count = max(rnd_mem_size >> (psize + 1), 1UL << 11);
531
532         return pteg_count << 7;
533 }
534
535 #ifdef CONFIG_MEMORY_HOTPLUG
536 void create_section_mapping(unsigned long start, unsigned long end)
537 {
538         BUG_ON(htab_bolt_mapping(start, end, __pa(start),
539                                  pgprot_val(PAGE_KERNEL), mmu_linear_psize,
540                                  mmu_kernel_ssize));
541 }
542
543 int remove_section_mapping(unsigned long start, unsigned long end)
544 {
545         return htab_remove_mapping(start, end, mmu_linear_psize,
546                         mmu_kernel_ssize);
547 }
548 #endif /* CONFIG_MEMORY_HOTPLUG */
549
550 static inline void make_bl(unsigned int *insn_addr, void *func)
551 {
552         unsigned long funcp = *((unsigned long *)func);
553         int offset = funcp - (unsigned long)insn_addr;
554
555         *insn_addr = (unsigned int)(0x48000001 | (offset & 0x03fffffc));
556         flush_icache_range((unsigned long)insn_addr, 4+
557                            (unsigned long)insn_addr);
558 }
559
560 static void __init htab_finish_init(void)
561 {
562         extern unsigned int *htab_call_hpte_insert1;
563         extern unsigned int *htab_call_hpte_insert2;
564         extern unsigned int *htab_call_hpte_remove;
565         extern unsigned int *htab_call_hpte_updatepp;
566
567 #ifdef CONFIG_PPC_HAS_HASH_64K
568         extern unsigned int *ht64_call_hpte_insert1;
569         extern unsigned int *ht64_call_hpte_insert2;
570         extern unsigned int *ht64_call_hpte_remove;
571         extern unsigned int *ht64_call_hpte_updatepp;
572
573         make_bl(ht64_call_hpte_insert1, ppc_md.hpte_insert);
574         make_bl(ht64_call_hpte_insert2, ppc_md.hpte_insert);
575         make_bl(ht64_call_hpte_remove, ppc_md.hpte_remove);
576         make_bl(ht64_call_hpte_updatepp, ppc_md.hpte_updatepp);
577 #endif /* CONFIG_PPC_HAS_HASH_64K */
578
579         make_bl(htab_call_hpte_insert1, ppc_md.hpte_insert);
580         make_bl(htab_call_hpte_insert2, ppc_md.hpte_insert);
581         make_bl(htab_call_hpte_remove, ppc_md.hpte_remove);
582         make_bl(htab_call_hpte_updatepp, ppc_md.hpte_updatepp);
583 }
584
585 static void __init htab_initialize(void)
586 {
587         unsigned long table;
588         unsigned long pteg_count;
589         unsigned long prot;
590         unsigned long base = 0, size = 0, limit;
591         int i;
592
593         DBG(" -> htab_initialize()\n");
594
595         /* Initialize segment sizes */
596         htab_init_seg_sizes();
597
598         /* Initialize page sizes */
599         htab_init_page_sizes();
600
601         if (cpu_has_feature(CPU_FTR_1T_SEGMENT)) {
602                 mmu_kernel_ssize = MMU_SEGSIZE_1T;
603                 mmu_highuser_ssize = MMU_SEGSIZE_1T;
604                 printk(KERN_INFO "Using 1TB segments\n");
605         }
606
607         /*
608          * Calculate the required size of the htab.  We want the number of
609          * PTEGs to equal one half the number of real pages.
610          */ 
611         htab_size_bytes = htab_get_table_size();
612         pteg_count = htab_size_bytes >> 7;
613
614         htab_hash_mask = pteg_count - 1;
615
616         if (firmware_has_feature(FW_FEATURE_LPAR)) {
617                 /* Using a hypervisor which owns the htab */
618                 htab_address = NULL;
619                 _SDR1 = 0; 
620         } else {
621                 /* Find storage for the HPT.  Must be contiguous in
622                  * the absolute address space. On cell we want it to be
623                  * in the first 2 Gig so we can use it for IOMMU hacks.
624                  */
625                 if (machine_is(cell))
626                         limit = 0x80000000;
627                 else
628                         limit = 0;
629
630                 table = memblock_alloc_base(htab_size_bytes, htab_size_bytes, limit);
631
632                 DBG("Hash table allocated at %lx, size: %lx\n", table,
633                     htab_size_bytes);
634
635                 htab_address = abs_to_virt(table);
636
637                 /* htab absolute addr + encoded htabsize */
638                 _SDR1 = table + __ilog2(pteg_count) - 11;
639
640                 /* Initialize the HPT with no entries */
641                 memset((void *)table, 0, htab_size_bytes);
642
643                 /* Set SDR1 */
644                 mtspr(SPRN_SDR1, _SDR1);
645         }
646
647         prot = pgprot_val(PAGE_KERNEL);
648
649 #ifdef CONFIG_DEBUG_PAGEALLOC
650         linear_map_hash_count = memblock_end_of_DRAM() >> PAGE_SHIFT;
651         linear_map_hash_slots = __va(memblock_alloc_base(linear_map_hash_count,
652                                                     1, memblock.rmo_size));
653         memset(linear_map_hash_slots, 0, linear_map_hash_count);
654 #endif /* CONFIG_DEBUG_PAGEALLOC */
655
656         /* On U3 based machines, we need to reserve the DART area and
657          * _NOT_ map it to avoid cache paradoxes as it's remapped non
658          * cacheable later on
659          */
660
661         /* create bolted the linear mapping in the hash table */
662         for (i=0; i < memblock.memory.cnt; i++) {
663                 base = (unsigned long)__va(memblock.memory.region[i].base);
664                 size = memblock.memory.region[i].size;
665
666                 DBG("creating mapping for region: %lx..%lx (prot: %lx)\n",
667                     base, size, prot);
668
669 #ifdef CONFIG_U3_DART
670                 /* Do not map the DART space. Fortunately, it will be aligned
671                  * in such a way that it will not cross two memblock regions and
672                  * will fit within a single 16Mb page.
673                  * The DART space is assumed to be a full 16Mb region even if
674                  * we only use 2Mb of that space. We will use more of it later
675                  * for AGP GART. We have to use a full 16Mb large page.
676                  */
677                 DBG("DART base: %lx\n", dart_tablebase);
678
679                 if (dart_tablebase != 0 && dart_tablebase >= base
680                     && dart_tablebase < (base + size)) {
681                         unsigned long dart_table_end = dart_tablebase + 16 * MB;
682                         if (base != dart_tablebase)
683                                 BUG_ON(htab_bolt_mapping(base, dart_tablebase,
684                                                         __pa(base), prot,
685                                                         mmu_linear_psize,
686                                                         mmu_kernel_ssize));
687                         if ((base + size) > dart_table_end)
688                                 BUG_ON(htab_bolt_mapping(dart_tablebase+16*MB,
689                                                         base + size,
690                                                         __pa(dart_table_end),
691                                                          prot,
692                                                          mmu_linear_psize,
693                                                          mmu_kernel_ssize));
694                         continue;
695                 }
696 #endif /* CONFIG_U3_DART */
697                 BUG_ON(htab_bolt_mapping(base, base + size, __pa(base),
698                                 prot, mmu_linear_psize, mmu_kernel_ssize));
699        }
700
701         /*
702          * If we have a memory_limit and we've allocated TCEs then we need to
703          * explicitly map the TCE area at the top of RAM. We also cope with the
704          * case that the TCEs start below memory_limit.
705          * tce_alloc_start/end are 16MB aligned so the mapping should work
706          * for either 4K or 16MB pages.
707          */
708         if (tce_alloc_start) {
709                 tce_alloc_start = (unsigned long)__va(tce_alloc_start);
710                 tce_alloc_end = (unsigned long)__va(tce_alloc_end);
711
712                 if (base + size >= tce_alloc_start)
713                         tce_alloc_start = base + size + 1;
714
715                 BUG_ON(htab_bolt_mapping(tce_alloc_start, tce_alloc_end,
716                                          __pa(tce_alloc_start), prot,
717                                          mmu_linear_psize, mmu_kernel_ssize));
718         }
719
720         htab_finish_init();
721
722         DBG(" <- htab_initialize()\n");
723 }
724 #undef KB
725 #undef MB
726
727 void __init early_init_mmu(void)
728 {
729         /* Setup initial STAB address in the PACA */
730         get_paca()->stab_real = __pa((u64)&initial_stab);
731         get_paca()->stab_addr = (u64)&initial_stab;
732
733         /* Initialize the MMU Hash table and create the linear mapping
734          * of memory. Has to be done before stab/slb initialization as
735          * this is currently where the page size encoding is obtained
736          */
737         htab_initialize();
738
739         /* Initialize stab / SLB management except on iSeries
740          */
741         if (cpu_has_feature(CPU_FTR_SLB))
742                 slb_initialize();
743         else if (!firmware_has_feature(FW_FEATURE_ISERIES))
744                 stab_initialize(get_paca()->stab_real);
745 }
746
747 #ifdef CONFIG_SMP
748 void __cpuinit early_init_mmu_secondary(void)
749 {
750         /* Initialize hash table for that CPU */
751         if (!firmware_has_feature(FW_FEATURE_LPAR))
752                 mtspr(SPRN_SDR1, _SDR1);
753
754         /* Initialize STAB/SLB. We use a virtual address as it works
755          * in real mode on pSeries and we want a virutal address on
756          * iSeries anyway
757          */
758         if (cpu_has_feature(CPU_FTR_SLB))
759                 slb_initialize();
760         else
761                 stab_initialize(get_paca()->stab_addr);
762 }
763 #endif /* CONFIG_SMP */
764
765 /*
766  * Called by asm hashtable.S for doing lazy icache flush
767  */
768 unsigned int hash_page_do_lazy_icache(unsigned int pp, pte_t pte, int trap)
769 {
770         struct page *page;
771
772         if (!pfn_valid(pte_pfn(pte)))
773                 return pp;
774
775         page = pte_page(pte);
776
777         /* page is dirty */
778         if (!test_bit(PG_arch_1, &page->flags) && !PageReserved(page)) {
779                 if (trap == 0x400) {
780                         flush_dcache_icache_page(page);
781                         set_bit(PG_arch_1, &page->flags);
782                 } else
783                         pp |= HPTE_R_N;
784         }
785         return pp;
786 }
787
788 #ifdef CONFIG_PPC_MM_SLICES
789 unsigned int get_paca_psize(unsigned long addr)
790 {
791         unsigned long index, slices;
792
793         if (addr < SLICE_LOW_TOP) {
794                 slices = get_paca()->context.low_slices_psize;
795                 index = GET_LOW_SLICE_INDEX(addr);
796         } else {
797                 slices = get_paca()->context.high_slices_psize;
798                 index = GET_HIGH_SLICE_INDEX(addr);
799         }
800         return (slices >> (index * 4)) & 0xF;
801 }
802
803 #else
804 unsigned int get_paca_psize(unsigned long addr)
805 {
806         return get_paca()->context.user_psize;
807 }
808 #endif
809
810 /*
811  * Demote a segment to using 4k pages.
812  * For now this makes the whole process use 4k pages.
813  */
814 #ifdef CONFIG_PPC_64K_PAGES
815 void demote_segment_4k(struct mm_struct *mm, unsigned long addr)
816 {
817         if (get_slice_psize(mm, addr) == MMU_PAGE_4K)
818                 return;
819         slice_set_range_psize(mm, addr, 1, MMU_PAGE_4K);
820 #ifdef CONFIG_SPU_BASE
821         spu_flush_all_slbs(mm);
822 #endif
823         if (get_paca_psize(addr) != MMU_PAGE_4K) {
824                 get_paca()->context = mm->context;
825                 slb_flush_and_rebolt();
826         }
827 }
828 #endif /* CONFIG_PPC_64K_PAGES */
829
830 #ifdef CONFIG_PPC_SUBPAGE_PROT
831 /*
832  * This looks up a 2-bit protection code for a 4k subpage of a 64k page.
833  * Userspace sets the subpage permissions using the subpage_prot system call.
834  *
835  * Result is 0: full permissions, _PAGE_RW: read-only,
836  * _PAGE_USER or _PAGE_USER|_PAGE_RW: no access.
837  */
838 static int subpage_protection(struct mm_struct *mm, unsigned long ea)
839 {
840         struct subpage_prot_table *spt = &mm->context.spt;
841         u32 spp = 0;
842         u32 **sbpm, *sbpp;
843
844         if (ea >= spt->maxaddr)
845                 return 0;
846         if (ea < 0x100000000) {
847                 /* addresses below 4GB use spt->low_prot */
848                 sbpm = spt->low_prot;
849         } else {
850                 sbpm = spt->protptrs[ea >> SBP_L3_SHIFT];
851                 if (!sbpm)
852                         return 0;
853         }
854         sbpp = sbpm[(ea >> SBP_L2_SHIFT) & (SBP_L2_COUNT - 1)];
855         if (!sbpp)
856                 return 0;
857         spp = sbpp[(ea >> PAGE_SHIFT) & (SBP_L1_COUNT - 1)];
858
859         /* extract 2-bit bitfield for this 4k subpage */
860         spp >>= 30 - 2 * ((ea >> 12) & 0xf);
861
862         /* turn 0,1,2,3 into combination of _PAGE_USER and _PAGE_RW */
863         spp = ((spp & 2) ? _PAGE_USER : 0) | ((spp & 1) ? _PAGE_RW : 0);
864         return spp;
865 }
866
867 #else /* CONFIG_PPC_SUBPAGE_PROT */
868 static inline int subpage_protection(struct mm_struct *mm, unsigned long ea)
869 {
870         return 0;
871 }
872 #endif
873
874 /* Result code is:
875  *  0 - handled
876  *  1 - normal page fault
877  * -1 - critical hash insertion error
878  * -2 - access not permitted by subpage protection mechanism
879  */
880 int hash_page(unsigned long ea, unsigned long access, unsigned long trap)
881 {
882         pgd_t *pgdir;
883         unsigned long vsid;
884         struct mm_struct *mm;
885         pte_t *ptep;
886         unsigned hugeshift;
887         const struct cpumask *tmp;
888         int rc, user_region = 0, local = 0;
889         int psize, ssize;
890
891         DBG_LOW("hash_page(ea=%016lx, access=%lx, trap=%lx\n",
892                 ea, access, trap);
893
894         if ((ea & ~REGION_MASK) >= PGTABLE_RANGE) {
895                 DBG_LOW(" out of pgtable range !\n");
896                 return 1;
897         }
898
899         /* Get region & vsid */
900         switch (REGION_ID(ea)) {
901         case USER_REGION_ID:
902                 user_region = 1;
903                 mm = current->mm;
904                 if (! mm) {
905                         DBG_LOW(" user region with no mm !\n");
906                         return 1;
907                 }
908                 psize = get_slice_psize(mm, ea);
909                 ssize = user_segment_size(ea);
910                 vsid = get_vsid(mm->context.id, ea, ssize);
911                 break;
912         case VMALLOC_REGION_ID:
913                 mm = &init_mm;
914                 vsid = get_kernel_vsid(ea, mmu_kernel_ssize);
915                 if (ea < VMALLOC_END)
916                         psize = mmu_vmalloc_psize;
917                 else
918                         psize = mmu_io_psize;
919                 ssize = mmu_kernel_ssize;
920                 break;
921         default:
922                 /* Not a valid range
923                  * Send the problem up to do_page_fault 
924                  */
925                 return 1;
926         }
927         DBG_LOW(" mm=%p, mm->pgdir=%p, vsid=%016lx\n", mm, mm->pgd, vsid);
928
929         /* Get pgdir */
930         pgdir = mm->pgd;
931         if (pgdir == NULL)
932                 return 1;
933
934         /* Check CPU locality */
935         tmp = cpumask_of(smp_processor_id());
936         if (user_region && cpumask_equal(mm_cpumask(mm), tmp))
937                 local = 1;
938
939 #ifndef CONFIG_PPC_64K_PAGES
940         /* If we use 4K pages and our psize is not 4K, then we might
941          * be hitting a special driver mapping, and need to align the
942          * address before we fetch the PTE.
943          *
944          * It could also be a hugepage mapping, in which case this is
945          * not necessary, but it's not harmful, either.
946          */
947         if (psize != MMU_PAGE_4K)
948                 ea &= ~((1ul << mmu_psize_defs[psize].shift) - 1);
949 #endif /* CONFIG_PPC_64K_PAGES */
950
951         /* Get PTE and page size from page tables */
952         ptep = find_linux_pte_or_hugepte(pgdir, ea, &hugeshift);
953         if (ptep == NULL || !pte_present(*ptep)) {
954                 DBG_LOW(" no PTE !\n");
955                 return 1;
956         }
957
958 #ifdef CONFIG_HUGETLB_PAGE
959         if (hugeshift)
960                 return __hash_page_huge(ea, access, vsid, ptep, trap, local,
961                                         ssize, hugeshift, psize);
962 #endif /* CONFIG_HUGETLB_PAGE */
963
964 #ifndef CONFIG_PPC_64K_PAGES
965         DBG_LOW(" i-pte: %016lx\n", pte_val(*ptep));
966 #else
967         DBG_LOW(" i-pte: %016lx %016lx\n", pte_val(*ptep),
968                 pte_val(*(ptep + PTRS_PER_PTE)));
969 #endif
970         /* Pre-check access permissions (will be re-checked atomically
971          * in __hash_page_XX but this pre-check is a fast path
972          */
973         if (access & ~pte_val(*ptep)) {
974                 DBG_LOW(" no access !\n");
975                 return 1;
976         }
977
978         /* Do actual hashing */
979 #ifdef CONFIG_PPC_64K_PAGES
980         /* If _PAGE_4K_PFN is set, make sure this is a 4k segment */
981         if ((pte_val(*ptep) & _PAGE_4K_PFN) && psize == MMU_PAGE_64K) {
982                 demote_segment_4k(mm, ea);
983                 psize = MMU_PAGE_4K;
984         }
985
986         /* If this PTE is non-cacheable and we have restrictions on
987          * using non cacheable large pages, then we switch to 4k
988          */
989         if (mmu_ci_restrictions && psize == MMU_PAGE_64K &&
990             (pte_val(*ptep) & _PAGE_NO_CACHE)) {
991                 if (user_region) {
992                         demote_segment_4k(mm, ea);
993                         psize = MMU_PAGE_4K;
994                 } else if (ea < VMALLOC_END) {
995                         /*
996                          * some driver did a non-cacheable mapping
997                          * in vmalloc space, so switch vmalloc
998                          * to 4k pages
999                          */
1000                         printk(KERN_ALERT "Reducing vmalloc segment "
1001                                "to 4kB pages because of "
1002                                "non-cacheable mapping\n");
1003                         psize = mmu_vmalloc_psize = MMU_PAGE_4K;
1004 #ifdef CONFIG_SPU_BASE
1005                         spu_flush_all_slbs(mm);
1006 #endif
1007                 }
1008         }
1009         if (user_region) {
1010                 if (psize != get_paca_psize(ea)) {
1011                         get_paca()->context = mm->context;
1012                         slb_flush_and_rebolt();
1013                 }
1014         } else if (get_paca()->vmalloc_sllp !=
1015                    mmu_psize_defs[mmu_vmalloc_psize].sllp) {
1016                 get_paca()->vmalloc_sllp =
1017                         mmu_psize_defs[mmu_vmalloc_psize].sllp;
1018                 slb_vmalloc_update();
1019         }
1020 #endif /* CONFIG_PPC_64K_PAGES */
1021
1022 #ifdef CONFIG_PPC_HAS_HASH_64K
1023         if (psize == MMU_PAGE_64K)
1024                 rc = __hash_page_64K(ea, access, vsid, ptep, trap, local, ssize);
1025         else
1026 #endif /* CONFIG_PPC_HAS_HASH_64K */
1027         {
1028                 int spp = subpage_protection(mm, ea);
1029                 if (access & spp)
1030                         rc = -2;
1031                 else
1032                         rc = __hash_page_4K(ea, access, vsid, ptep, trap,
1033                                             local, ssize, spp);
1034         }
1035
1036 #ifndef CONFIG_PPC_64K_PAGES
1037         DBG_LOW(" o-pte: %016lx\n", pte_val(*ptep));
1038 #else
1039         DBG_LOW(" o-pte: %016lx %016lx\n", pte_val(*ptep),
1040                 pte_val(*(ptep + PTRS_PER_PTE)));
1041 #endif
1042         DBG_LOW(" -> rc=%d\n", rc);
1043         return rc;
1044 }
1045 EXPORT_SYMBOL_GPL(hash_page);
1046
1047 void hash_preload(struct mm_struct *mm, unsigned long ea,
1048                   unsigned long access, unsigned long trap)
1049 {
1050         unsigned long vsid;
1051         void *pgdir;
1052         pte_t *ptep;
1053         unsigned long flags;
1054         int local = 0;
1055         int ssize;
1056
1057         BUG_ON(REGION_ID(ea) != USER_REGION_ID);
1058
1059 #ifdef CONFIG_PPC_MM_SLICES
1060         /* We only prefault standard pages for now */
1061         if (unlikely(get_slice_psize(mm, ea) != mm->context.user_psize))
1062                 return;
1063 #endif
1064
1065         DBG_LOW("hash_preload(mm=%p, mm->pgdir=%p, ea=%016lx, access=%lx,"
1066                 " trap=%lx\n", mm, mm->pgd, ea, access, trap);
1067
1068         /* Get Linux PTE if available */
1069         pgdir = mm->pgd;
1070         if (pgdir == NULL)
1071                 return;
1072         ptep = find_linux_pte(pgdir, ea);
1073         if (!ptep)
1074                 return;
1075
1076 #ifdef CONFIG_PPC_64K_PAGES
1077         /* If either _PAGE_4K_PFN or _PAGE_NO_CACHE is set (and we are on
1078          * a 64K kernel), then we don't preload, hash_page() will take
1079          * care of it once we actually try to access the page.
1080          * That way we don't have to duplicate all of the logic for segment
1081          * page size demotion here
1082          */
1083         if (pte_val(*ptep) & (_PAGE_4K_PFN | _PAGE_NO_CACHE))
1084                 return;
1085 #endif /* CONFIG_PPC_64K_PAGES */
1086
1087         /* Get VSID */
1088         ssize = user_segment_size(ea);
1089         vsid = get_vsid(mm->context.id, ea, ssize);
1090
1091         /* Hash doesn't like irqs */
1092         local_irq_save(flags);
1093
1094         /* Is that local to this CPU ? */
1095         if (cpumask_equal(mm_cpumask(mm), cpumask_of(smp_processor_id())))
1096                 local = 1;
1097
1098         /* Hash it in */
1099 #ifdef CONFIG_PPC_HAS_HASH_64K
1100         if (mm->context.user_psize == MMU_PAGE_64K)
1101                 __hash_page_64K(ea, access, vsid, ptep, trap, local, ssize);
1102         else
1103 #endif /* CONFIG_PPC_HAS_HASH_64K */
1104                 __hash_page_4K(ea, access, vsid, ptep, trap, local, ssize,
1105                                subpage_protection(pgdir, ea));
1106
1107         local_irq_restore(flags);
1108 }
1109
1110 /* WARNING: This is called from hash_low_64.S, if you change this prototype,
1111  *          do not forget to update the assembly call site !
1112  */
1113 void flush_hash_page(unsigned long va, real_pte_t pte, int psize, int ssize,
1114                      int local)
1115 {
1116         unsigned long hash, index, shift, hidx, slot;
1117
1118         DBG_LOW("flush_hash_page(va=%016lx)\n", va);
1119         pte_iterate_hashed_subpages(pte, psize, va, index, shift) {
1120                 hash = hpt_hash(va, shift, ssize);
1121                 hidx = __rpte_to_hidx(pte, index);
1122                 if (hidx & _PTEIDX_SECONDARY)
1123                         hash = ~hash;
1124                 slot = (hash & htab_hash_mask) * HPTES_PER_GROUP;
1125                 slot += hidx & _PTEIDX_GROUP_IX;
1126                 DBG_LOW(" sub %ld: hash=%lx, hidx=%lx\n", index, slot, hidx);
1127                 ppc_md.hpte_invalidate(slot, va, psize, ssize, local);
1128         } pte_iterate_hashed_end();
1129 }
1130
1131 void flush_hash_range(unsigned long number, int local)
1132 {
1133         if (ppc_md.flush_hash_range)
1134                 ppc_md.flush_hash_range(number, local);
1135         else {
1136                 int i;
1137                 struct ppc64_tlb_batch *batch =
1138                         &__get_cpu_var(ppc64_tlb_batch);
1139
1140                 for (i = 0; i < number; i++)
1141                         flush_hash_page(batch->vaddr[i], batch->pte[i],
1142                                         batch->psize, batch->ssize, local);
1143         }
1144 }
1145
1146 /*
1147  * low_hash_fault is called when we the low level hash code failed
1148  * to instert a PTE due to an hypervisor error
1149  */
1150 void low_hash_fault(struct pt_regs *regs, unsigned long address, int rc)
1151 {
1152         if (user_mode(regs)) {
1153 #ifdef CONFIG_PPC_SUBPAGE_PROT
1154                 if (rc == -2)
1155                         _exception(SIGSEGV, regs, SEGV_ACCERR, address);
1156                 else
1157 #endif
1158                         _exception(SIGBUS, regs, BUS_ADRERR, address);
1159         } else
1160                 bad_page_fault(regs, address, SIGBUS);
1161 }
1162
1163 #ifdef CONFIG_DEBUG_PAGEALLOC
1164 static void kernel_map_linear_page(unsigned long vaddr, unsigned long lmi)
1165 {
1166         unsigned long hash, hpteg;
1167         unsigned long vsid = get_kernel_vsid(vaddr, mmu_kernel_ssize);
1168         unsigned long va = hpt_va(vaddr, vsid, mmu_kernel_ssize);
1169         unsigned long mode = htab_convert_pte_flags(PAGE_KERNEL);
1170         int ret;
1171
1172         hash = hpt_hash(va, PAGE_SHIFT, mmu_kernel_ssize);
1173         hpteg = ((hash & htab_hash_mask) * HPTES_PER_GROUP);
1174
1175         ret = ppc_md.hpte_insert(hpteg, va, __pa(vaddr),
1176                                  mode, HPTE_V_BOLTED,
1177                                  mmu_linear_psize, mmu_kernel_ssize);
1178         BUG_ON (ret < 0);
1179         spin_lock(&linear_map_hash_lock);
1180         BUG_ON(linear_map_hash_slots[lmi] & 0x80);
1181         linear_map_hash_slots[lmi] = ret | 0x80;
1182         spin_unlock(&linear_map_hash_lock);
1183 }
1184
1185 static void kernel_unmap_linear_page(unsigned long vaddr, unsigned long lmi)
1186 {
1187         unsigned long hash, hidx, slot;
1188         unsigned long vsid = get_kernel_vsid(vaddr, mmu_kernel_ssize);
1189         unsigned long va = hpt_va(vaddr, vsid, mmu_kernel_ssize);
1190
1191         hash = hpt_hash(va, PAGE_SHIFT, mmu_kernel_ssize);
1192         spin_lock(&linear_map_hash_lock);
1193         BUG_ON(!(linear_map_hash_slots[lmi] & 0x80));
1194         hidx = linear_map_hash_slots[lmi] & 0x7f;
1195         linear_map_hash_slots[lmi] = 0;
1196         spin_unlock(&linear_map_hash_lock);
1197         if (hidx & _PTEIDX_SECONDARY)
1198                 hash = ~hash;
1199         slot = (hash & htab_hash_mask) * HPTES_PER_GROUP;
1200         slot += hidx & _PTEIDX_GROUP_IX;
1201         ppc_md.hpte_invalidate(slot, va, mmu_linear_psize, mmu_kernel_ssize, 0);
1202 }
1203
1204 void kernel_map_pages(struct page *page, int numpages, int enable)
1205 {
1206         unsigned long flags, vaddr, lmi;
1207         int i;
1208
1209         local_irq_save(flags);
1210         for (i = 0; i < numpages; i++, page++) {
1211                 vaddr = (unsigned long)page_address(page);
1212                 lmi = __pa(vaddr) >> PAGE_SHIFT;
1213                 if (lmi >= linear_map_hash_count)
1214                         continue;
1215                 if (enable)
1216                         kernel_map_linear_page(vaddr, lmi);
1217                 else
1218                         kernel_unmap_linear_page(vaddr, lmi);
1219         }
1220         local_irq_restore(flags);
1221 }
1222 #endif /* CONFIG_DEBUG_PAGEALLOC */