2 * PowerPC64 port by Mike Corrigan and Dave Engebretsen
3 * {mikejc|engebret}@us.ibm.com
5 * Copyright (c) 2000 Mike Corrigan <mikejc@us.ibm.com>
7 * SMP scalability work:
8 * Copyright (C) 2001 Anton Blanchard <anton@au.ibm.com>, IBM
13 * PowerPC Hashed Page Table functions
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.
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/export.h>
31 #include <linux/ctype.h>
32 #include <linux/cache.h>
33 #include <linux/init.h>
34 #include <linux/signal.h>
35 #include <linux/memblock.h>
37 #include <asm/processor.h>
38 #include <asm/pgtable.h>
40 #include <asm/mmu_context.h>
42 #include <asm/types.h>
43 #include <asm/uaccess.h>
44 #include <asm/machdep.h>
46 #include <asm/tlbflush.h>
50 #include <asm/cacheflush.h>
51 #include <asm/cputable.h>
52 #include <asm/sections.h>
55 #include <asm/code-patching.h>
56 #include <asm/fadump.h>
57 #include <asm/firmware.h>
61 #define DBG(fmt...) udbg_printf(fmt)
67 #define DBG_LOW(fmt...) udbg_printf(fmt)
69 #define DBG_LOW(fmt...)
77 * Note: pte --> Linux PTE
78 * HPTE --> PowerPC Hashed Page Table Entry
81 * htab_initialize is called with the MMU off (of course), but
82 * the kernel has been copied down to zero so it can directly
83 * reference global data. At this point it is very difficult
84 * to print debug info.
89 extern unsigned long dart_tablebase;
90 #endif /* CONFIG_U3_DART */
92 static unsigned long _SDR1;
93 struct mmu_psize_def mmu_psize_defs[MMU_PAGE_COUNT];
95 struct hash_pte *htab_address;
96 unsigned long htab_size_bytes;
97 unsigned long htab_hash_mask;
98 EXPORT_SYMBOL_GPL(htab_hash_mask);
99 int mmu_linear_psize = MMU_PAGE_4K;
100 int mmu_virtual_psize = MMU_PAGE_4K;
101 int mmu_vmalloc_psize = MMU_PAGE_4K;
102 #ifdef CONFIG_SPARSEMEM_VMEMMAP
103 int mmu_vmemmap_psize = MMU_PAGE_4K;
105 int mmu_io_psize = MMU_PAGE_4K;
106 int mmu_kernel_ssize = MMU_SEGSIZE_256M;
107 int mmu_highuser_ssize = MMU_SEGSIZE_256M;
108 u16 mmu_slb_size = 64;
109 EXPORT_SYMBOL_GPL(mmu_slb_size);
110 #ifdef CONFIG_PPC_64K_PAGES
111 int mmu_ci_restrictions;
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 */
119 /* There are definitions of page sizes arrays to be used when none
120 * is provided by the firmware.
123 /* Pre-POWER4 CPUs (4k pages only)
125 static struct mmu_psize_def mmu_psize_defaults_old[] = {
135 /* POWER4, GPUL, POWER5
137 * Support for 16Mb large pages
139 static struct mmu_psize_def mmu_psize_defaults_gp[] = {
156 static unsigned long htab_convert_pte_flags(unsigned long pteflags)
158 unsigned long rflags = pteflags & 0x1fa;
160 /* _PAGE_EXEC -> NOEXEC */
161 if ((pteflags & _PAGE_EXEC) == 0)
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
167 if ((pteflags & _PAGE_USER) && !((pteflags & _PAGE_RW) &&
168 (pteflags & _PAGE_DIRTY)))
172 return rflags | HPTE_R_C;
175 int htab_bolt_mapping(unsigned long vstart, unsigned long vend,
176 unsigned long pstart, unsigned long prot,
177 int psize, int ssize)
179 unsigned long vaddr, paddr;
180 unsigned int step, shift;
183 shift = mmu_psize_defs[psize].shift;
186 prot = htab_convert_pte_flags(prot);
188 DBG("htab_bolt_mapping(%lx..%lx -> %lx (%lx,%d,%d)\n",
189 vstart, vend, pstart, prot, psize, ssize);
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 vpn = hpt_vpn(vaddr, vsid, ssize);
196 unsigned long tprot = prot;
198 /* Make kernel text executable */
199 if (overlaps_kernel_text(vaddr, vaddr + step))
202 hash = hpt_hash(vpn, shift, ssize);
203 hpteg = ((hash & htab_hash_mask) * HPTES_PER_GROUP);
205 BUG_ON(!ppc_md.hpte_insert);
206 ret = ppc_md.hpte_insert(hpteg, vpn, paddr, tprot,
207 HPTE_V_BOLTED, psize, ssize);
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 */
216 return ret < 0 ? ret : 0;
219 #ifdef CONFIG_MEMORY_HOTPLUG
220 static int htab_remove_mapping(unsigned long vstart, unsigned long vend,
221 int psize, int ssize)
224 unsigned int step, shift;
226 shift = mmu_psize_defs[psize].shift;
229 if (!ppc_md.hpte_removebolted) {
230 printk(KERN_WARNING "Platform doesn't implement "
231 "hpte_removebolted\n");
235 for (vaddr = vstart; vaddr < vend; vaddr += step)
236 ppc_md.hpte_removebolted(vaddr, psize, ssize);
240 #endif /* CONFIG_MEMORY_HOTPLUG */
242 static int __init htab_dt_scan_seg_sizes(unsigned long node,
243 const char *uname, int depth,
246 char *type = of_get_flat_dt_prop(node, "device_type", NULL);
248 unsigned long size = 0;
250 /* We are scanning "cpu" nodes only */
251 if (type == NULL || strcmp(type, "cpu") != 0)
254 prop = (u32 *)of_get_flat_dt_prop(node, "ibm,processor-segment-sizes",
258 for (; size >= 4; size -= 4, ++prop) {
260 DBG("1T segment support detected\n");
261 cur_cpu_spec->mmu_features |= MMU_FTR_1T_SEGMENT;
265 cur_cpu_spec->mmu_features &= ~MMU_FTR_NO_SLBIE_B;
269 static void __init htab_init_seg_sizes(void)
271 of_scan_flat_dt(htab_dt_scan_seg_sizes, NULL);
274 static int __init htab_dt_scan_page_sizes(unsigned long node,
275 const char *uname, int depth,
278 char *type = of_get_flat_dt_prop(node, "device_type", NULL);
280 unsigned long size = 0;
282 /* We are scanning "cpu" nodes only */
283 if (type == NULL || strcmp(type, "cpu") != 0)
286 prop = (u32 *)of_get_flat_dt_prop(node,
287 "ibm,segment-page-sizes", &size);
289 DBG("Page sizes from device-tree:\n");
291 cur_cpu_spec->mmu_features &= ~(MMU_FTR_16M_PAGE);
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;
300 size -= 3; prop += 3;
301 while(size > 0 && lpnum) {
302 if (prop[0] == shift)
304 prop += 2; size -= 2;
319 cur_cpu_spec->mmu_features |= MMU_FTR_16M_PAGE;
327 def = &mmu_psize_defs[idx];
332 def->avpnm = (1 << (shift - 23)) - 1;
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
338 if (idx == MMU_PAGE_4K || idx == MMU_PAGE_64K)
343 DBG(" %d: shift=%02x, sllp=%04lx, avpnm=%08lx, "
344 "tlbiel=%d, penc=%d\n",
345 idx, shift, def->sllp, def->avpnm, def->tlbiel,
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.
357 static int __init htab_dt_scan_hugepage_blocks(unsigned long node,
358 const char *uname, int depth,
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;
367 /* We are scanning "memory" nodes only */
368 if (type == NULL || strcmp(type, "memory") != 0)
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)
376 expected_pages = (1 << page_count_prop[0]);
377 addr_prop = of_get_flat_dt_prop(node, "reg", NULL);
378 if (addr_prop == NULL)
380 phys_addr = addr_prop[0];
381 block_size = addr_prop[1];
382 if (block_size != (16 * GB))
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);
393 #endif /* CONFIG_HUGETLB_PAGE */
395 static void __init htab_init_page_sizes(void)
399 /* Default to 4K pages only */
400 memcpy(mmu_psize_defs, mmu_psize_defaults_old,
401 sizeof(mmu_psize_defaults_old));
404 * Try to find the available page sizes in the device-tree
406 rc = of_scan_flat_dt(htab_dt_scan_page_sizes, NULL);
407 if (rc != 0) /* Found */
411 * Not in the device-tree, let's fallback on known size
412 * list for 16M capable GP & GR
414 if (mmu_has_feature(MMU_FTR_16M_PAGE))
415 memcpy(mmu_psize_defs, mmu_psize_defaults_gp,
416 sizeof(mmu_psize_defaults_gp));
418 #ifndef CONFIG_DEBUG_PAGEALLOC
420 * Pick a size for the linear mapping. Currently, we only support
421 * 16M, 1M and 4K which is the default
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 */
429 #ifdef CONFIG_PPC_64K_PAGES
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.
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 (mmu_has_feature(MMU_FTR_CI_LARGE_PAGE)) {
446 * Don't use 64k pages for ioremap on pSeries, since
447 * that would stop us accessing the HEA ethernet.
449 if (!machine_is(pseries))
450 mmu_io_psize = MMU_PAGE_64K;
452 mmu_ci_restrictions = 1;
454 #endif /* CONFIG_PPC_64K_PAGES */
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
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;
466 mmu_vmemmap_psize = MMU_PAGE_4K;
467 #endif /* CONFIG_SPARSEMEM_VMEMMAP */
469 printk(KERN_DEBUG "Page orders: linear mapping = %d, "
470 "virtual = %d, io = %d"
471 #ifdef CONFIG_SPARSEMEM_VMEMMAP
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
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 */
489 static int __init htab_dt_scan_pftsize(unsigned long node,
490 const char *uname, int depth,
493 char *type = of_get_flat_dt_prop(node, "device_type", NULL);
496 /* We are scanning "cpu" nodes only */
497 if (type == NULL || strcmp(type, "cpu") != 0)
500 prop = (u32 *)of_get_flat_dt_prop(node, "ibm,pft-size", NULL);
502 /* pft_size[0] is the NUMA CEC cookie */
503 ppc64_pft_size = prop[1];
509 static unsigned long __init htab_get_table_size(void)
511 unsigned long mem_size, rnd_mem_size, pteg_count, psize;
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
517 if (ppc64_pft_size == 0)
518 of_scan_flat_dt(htab_dt_scan_pftsize, NULL);
520 return 1UL << ppc64_pft_size;
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)
529 psize = mmu_psize_defs[mmu_virtual_psize].shift;
530 pteg_count = max(rnd_mem_size >> (psize + 1), 1UL << 11);
532 return pteg_count << 7;
535 #ifdef CONFIG_MEMORY_HOTPLUG
536 int create_section_mapping(unsigned long start, unsigned long end)
538 return htab_bolt_mapping(start, end, __pa(start),
539 pgprot_val(PAGE_KERNEL), mmu_linear_psize,
543 int remove_section_mapping(unsigned long start, unsigned long end)
545 return htab_remove_mapping(start, end, mmu_linear_psize,
548 #endif /* CONFIG_MEMORY_HOTPLUG */
550 #define FUNCTION_TEXT(A) ((*(unsigned long *)(A)))
552 static void __init htab_finish_init(void)
554 extern unsigned int *htab_call_hpte_insert1;
555 extern unsigned int *htab_call_hpte_insert2;
556 extern unsigned int *htab_call_hpte_remove;
557 extern unsigned int *htab_call_hpte_updatepp;
559 #ifdef CONFIG_PPC_HAS_HASH_64K
560 extern unsigned int *ht64_call_hpte_insert1;
561 extern unsigned int *ht64_call_hpte_insert2;
562 extern unsigned int *ht64_call_hpte_remove;
563 extern unsigned int *ht64_call_hpte_updatepp;
565 patch_branch(ht64_call_hpte_insert1,
566 FUNCTION_TEXT(ppc_md.hpte_insert),
568 patch_branch(ht64_call_hpte_insert2,
569 FUNCTION_TEXT(ppc_md.hpte_insert),
571 patch_branch(ht64_call_hpte_remove,
572 FUNCTION_TEXT(ppc_md.hpte_remove),
574 patch_branch(ht64_call_hpte_updatepp,
575 FUNCTION_TEXT(ppc_md.hpte_updatepp),
578 #endif /* CONFIG_PPC_HAS_HASH_64K */
580 patch_branch(htab_call_hpte_insert1,
581 FUNCTION_TEXT(ppc_md.hpte_insert),
583 patch_branch(htab_call_hpte_insert2,
584 FUNCTION_TEXT(ppc_md.hpte_insert),
586 patch_branch(htab_call_hpte_remove,
587 FUNCTION_TEXT(ppc_md.hpte_remove),
589 patch_branch(htab_call_hpte_updatepp,
590 FUNCTION_TEXT(ppc_md.hpte_updatepp),
594 static void __init htab_initialize(void)
597 unsigned long pteg_count;
599 unsigned long base = 0, size = 0, limit;
600 struct memblock_region *reg;
602 DBG(" -> htab_initialize()\n");
604 /* Initialize segment sizes */
605 htab_init_seg_sizes();
607 /* Initialize page sizes */
608 htab_init_page_sizes();
610 if (mmu_has_feature(MMU_FTR_1T_SEGMENT)) {
611 mmu_kernel_ssize = MMU_SEGSIZE_1T;
612 mmu_highuser_ssize = MMU_SEGSIZE_1T;
613 printk(KERN_INFO "Using 1TB segments\n");
617 * Calculate the required size of the htab. We want the number of
618 * PTEGs to equal one half the number of real pages.
620 htab_size_bytes = htab_get_table_size();
621 pteg_count = htab_size_bytes >> 7;
623 htab_hash_mask = pteg_count - 1;
625 if (firmware_has_feature(FW_FEATURE_LPAR)) {
626 /* Using a hypervisor which owns the htab */
629 #ifdef CONFIG_FA_DUMP
631 * If firmware assisted dump is active firmware preserves
632 * the contents of htab along with entire partition memory.
633 * Clear the htab if firmware assisted dump is active so
634 * that we dont end up using old mappings.
636 if (is_fadump_active() && ppc_md.hpte_clear_all)
637 ppc_md.hpte_clear_all();
640 /* Find storage for the HPT. Must be contiguous in
641 * the absolute address space. On cell we want it to be
642 * in the first 2 Gig so we can use it for IOMMU hacks.
644 if (machine_is(cell))
647 limit = MEMBLOCK_ALLOC_ANYWHERE;
649 table = memblock_alloc_base(htab_size_bytes, htab_size_bytes, limit);
651 DBG("Hash table allocated at %lx, size: %lx\n", table,
654 htab_address = __va(table);
656 /* htab absolute addr + encoded htabsize */
657 _SDR1 = table + __ilog2(pteg_count) - 11;
659 /* Initialize the HPT with no entries */
660 memset((void *)table, 0, htab_size_bytes);
663 mtspr(SPRN_SDR1, _SDR1);
666 prot = pgprot_val(PAGE_KERNEL);
668 #ifdef CONFIG_DEBUG_PAGEALLOC
669 linear_map_hash_count = memblock_end_of_DRAM() >> PAGE_SHIFT;
670 linear_map_hash_slots = __va(memblock_alloc_base(linear_map_hash_count,
672 memset(linear_map_hash_slots, 0, linear_map_hash_count);
673 #endif /* CONFIG_DEBUG_PAGEALLOC */
675 /* On U3 based machines, we need to reserve the DART area and
676 * _NOT_ map it to avoid cache paradoxes as it's remapped non
680 /* create bolted the linear mapping in the hash table */
681 for_each_memblock(memory, reg) {
682 base = (unsigned long)__va(reg->base);
685 DBG("creating mapping for region: %lx..%lx (prot: %lx)\n",
688 #ifdef CONFIG_U3_DART
689 /* Do not map the DART space. Fortunately, it will be aligned
690 * in such a way that it will not cross two memblock regions and
691 * will fit within a single 16Mb page.
692 * The DART space is assumed to be a full 16Mb region even if
693 * we only use 2Mb of that space. We will use more of it later
694 * for AGP GART. We have to use a full 16Mb large page.
696 DBG("DART base: %lx\n", dart_tablebase);
698 if (dart_tablebase != 0 && dart_tablebase >= base
699 && dart_tablebase < (base + size)) {
700 unsigned long dart_table_end = dart_tablebase + 16 * MB;
701 if (base != dart_tablebase)
702 BUG_ON(htab_bolt_mapping(base, dart_tablebase,
706 if ((base + size) > dart_table_end)
707 BUG_ON(htab_bolt_mapping(dart_tablebase+16*MB,
709 __pa(dart_table_end),
715 #endif /* CONFIG_U3_DART */
716 BUG_ON(htab_bolt_mapping(base, base + size, __pa(base),
717 prot, mmu_linear_psize, mmu_kernel_ssize));
719 memblock_set_current_limit(MEMBLOCK_ALLOC_ANYWHERE);
722 * If we have a memory_limit and we've allocated TCEs then we need to
723 * explicitly map the TCE area at the top of RAM. We also cope with the
724 * case that the TCEs start below memory_limit.
725 * tce_alloc_start/end are 16MB aligned so the mapping should work
726 * for either 4K or 16MB pages.
728 if (tce_alloc_start) {
729 tce_alloc_start = (unsigned long)__va(tce_alloc_start);
730 tce_alloc_end = (unsigned long)__va(tce_alloc_end);
732 if (base + size >= tce_alloc_start)
733 tce_alloc_start = base + size + 1;
735 BUG_ON(htab_bolt_mapping(tce_alloc_start, tce_alloc_end,
736 __pa(tce_alloc_start), prot,
737 mmu_linear_psize, mmu_kernel_ssize));
742 DBG(" <- htab_initialize()\n");
747 void __init early_init_mmu(void)
749 /* Setup initial STAB address in the PACA */
750 get_paca()->stab_real = __pa((u64)&initial_stab);
751 get_paca()->stab_addr = (u64)&initial_stab;
753 /* Initialize the MMU Hash table and create the linear mapping
754 * of memory. Has to be done before stab/slb initialization as
755 * this is currently where the page size encoding is obtained
759 /* Initialize stab / SLB management */
760 if (mmu_has_feature(MMU_FTR_SLB))
765 void __cpuinit early_init_mmu_secondary(void)
767 /* Initialize hash table for that CPU */
768 if (!firmware_has_feature(FW_FEATURE_LPAR))
769 mtspr(SPRN_SDR1, _SDR1);
771 /* Initialize STAB/SLB. We use a virtual address as it works
772 * in real mode on pSeries.
774 if (mmu_has_feature(MMU_FTR_SLB))
777 stab_initialize(get_paca()->stab_addr);
779 #endif /* CONFIG_SMP */
782 * Called by asm hashtable.S for doing lazy icache flush
784 unsigned int hash_page_do_lazy_icache(unsigned int pp, pte_t pte, int trap)
788 if (!pfn_valid(pte_pfn(pte)))
791 page = pte_page(pte);
794 if (!test_bit(PG_arch_1, &page->flags) && !PageReserved(page)) {
796 flush_dcache_icache_page(page);
797 set_bit(PG_arch_1, &page->flags);
804 #ifdef CONFIG_PPC_MM_SLICES
805 unsigned int get_paca_psize(unsigned long addr)
808 unsigned char *hpsizes;
809 unsigned long index, mask_index;
811 if (addr < SLICE_LOW_TOP) {
812 lpsizes = get_paca()->context.low_slices_psize;
813 index = GET_LOW_SLICE_INDEX(addr);
814 return (lpsizes >> (index * 4)) & 0xF;
816 hpsizes = get_paca()->context.high_slices_psize;
817 index = GET_HIGH_SLICE_INDEX(addr);
818 mask_index = index & 0x1;
819 return (hpsizes[index >> 1] >> (mask_index * 4)) & 0xF;
823 unsigned int get_paca_psize(unsigned long addr)
825 return get_paca()->context.user_psize;
830 * Demote a segment to using 4k pages.
831 * For now this makes the whole process use 4k pages.
833 #ifdef CONFIG_PPC_64K_PAGES
834 void demote_segment_4k(struct mm_struct *mm, unsigned long addr)
836 if (get_slice_psize(mm, addr) == MMU_PAGE_4K)
838 slice_set_range_psize(mm, addr, 1, MMU_PAGE_4K);
839 #ifdef CONFIG_SPU_BASE
840 spu_flush_all_slbs(mm);
842 if (get_paca_psize(addr) != MMU_PAGE_4K) {
843 get_paca()->context = mm->context;
844 slb_flush_and_rebolt();
847 #endif /* CONFIG_PPC_64K_PAGES */
849 #ifdef CONFIG_PPC_SUBPAGE_PROT
851 * This looks up a 2-bit protection code for a 4k subpage of a 64k page.
852 * Userspace sets the subpage permissions using the subpage_prot system call.
854 * Result is 0: full permissions, _PAGE_RW: read-only,
855 * _PAGE_USER or _PAGE_USER|_PAGE_RW: no access.
857 static int subpage_protection(struct mm_struct *mm, unsigned long ea)
859 struct subpage_prot_table *spt = &mm->context.spt;
863 if (ea >= spt->maxaddr)
865 if (ea < 0x100000000) {
866 /* addresses below 4GB use spt->low_prot */
867 sbpm = spt->low_prot;
869 sbpm = spt->protptrs[ea >> SBP_L3_SHIFT];
873 sbpp = sbpm[(ea >> SBP_L2_SHIFT) & (SBP_L2_COUNT - 1)];
876 spp = sbpp[(ea >> PAGE_SHIFT) & (SBP_L1_COUNT - 1)];
878 /* extract 2-bit bitfield for this 4k subpage */
879 spp >>= 30 - 2 * ((ea >> 12) & 0xf);
881 /* turn 0,1,2,3 into combination of _PAGE_USER and _PAGE_RW */
882 spp = ((spp & 2) ? _PAGE_USER : 0) | ((spp & 1) ? _PAGE_RW : 0);
886 #else /* CONFIG_PPC_SUBPAGE_PROT */
887 static inline int subpage_protection(struct mm_struct *mm, unsigned long ea)
893 void hash_failure_debug(unsigned long ea, unsigned long access,
894 unsigned long vsid, unsigned long trap,
895 int ssize, int psize, unsigned long pte)
897 if (!printk_ratelimit())
899 pr_info("mm: Hashing failure ! EA=0x%lx access=0x%lx current=%s\n",
900 ea, access, current->comm);
901 pr_info(" trap=0x%lx vsid=0x%lx ssize=%d psize=%d pte=0x%lx\n",
902 trap, vsid, ssize, psize, pte);
907 * 1 - normal page fault
908 * -1 - critical hash insertion error
909 * -2 - access not permitted by subpage protection mechanism
911 int hash_page(unsigned long ea, unsigned long access, unsigned long trap)
915 struct mm_struct *mm;
918 const struct cpumask *tmp;
919 int rc, user_region = 0, local = 0;
922 DBG_LOW("hash_page(ea=%016lx, access=%lx, trap=%lx\n",
925 if ((ea & ~REGION_MASK) >= PGTABLE_RANGE) {
926 DBG_LOW(" out of pgtable range !\n");
930 /* Get region & vsid */
931 switch (REGION_ID(ea)) {
936 DBG_LOW(" user region with no mm !\n");
939 psize = get_slice_psize(mm, ea);
940 ssize = user_segment_size(ea);
941 vsid = get_vsid(mm->context.id, ea, ssize);
943 case VMALLOC_REGION_ID:
945 vsid = get_kernel_vsid(ea, mmu_kernel_ssize);
946 if (ea < VMALLOC_END)
947 psize = mmu_vmalloc_psize;
949 psize = mmu_io_psize;
950 ssize = mmu_kernel_ssize;
954 * Send the problem up to do_page_fault
958 DBG_LOW(" mm=%p, mm->pgdir=%p, vsid=%016lx\n", mm, mm->pgd, vsid);
965 /* Check CPU locality */
966 tmp = cpumask_of(smp_processor_id());
967 if (user_region && cpumask_equal(mm_cpumask(mm), tmp))
970 #ifndef CONFIG_PPC_64K_PAGES
971 /* If we use 4K pages and our psize is not 4K, then we might
972 * be hitting a special driver mapping, and need to align the
973 * address before we fetch the PTE.
975 * It could also be a hugepage mapping, in which case this is
976 * not necessary, but it's not harmful, either.
978 if (psize != MMU_PAGE_4K)
979 ea &= ~((1ul << mmu_psize_defs[psize].shift) - 1);
980 #endif /* CONFIG_PPC_64K_PAGES */
982 /* Get PTE and page size from page tables */
983 ptep = find_linux_pte_or_hugepte(pgdir, ea, &hugeshift);
984 if (ptep == NULL || !pte_present(*ptep)) {
985 DBG_LOW(" no PTE !\n");
989 /* Add _PAGE_PRESENT to the required access perm */
990 access |= _PAGE_PRESENT;
992 /* Pre-check access permissions (will be re-checked atomically
993 * in __hash_page_XX but this pre-check is a fast path
995 if (access & ~pte_val(*ptep)) {
996 DBG_LOW(" no access !\n");
1000 #ifdef CONFIG_HUGETLB_PAGE
1002 return __hash_page_huge(ea, access, vsid, ptep, trap, local,
1003 ssize, hugeshift, psize);
1004 #endif /* CONFIG_HUGETLB_PAGE */
1006 #ifndef CONFIG_PPC_64K_PAGES
1007 DBG_LOW(" i-pte: %016lx\n", pte_val(*ptep));
1009 DBG_LOW(" i-pte: %016lx %016lx\n", pte_val(*ptep),
1010 pte_val(*(ptep + PTRS_PER_PTE)));
1012 /* Do actual hashing */
1013 #ifdef CONFIG_PPC_64K_PAGES
1014 /* If _PAGE_4K_PFN is set, make sure this is a 4k segment */
1015 if ((pte_val(*ptep) & _PAGE_4K_PFN) && psize == MMU_PAGE_64K) {
1016 demote_segment_4k(mm, ea);
1017 psize = MMU_PAGE_4K;
1020 /* If this PTE is non-cacheable and we have restrictions on
1021 * using non cacheable large pages, then we switch to 4k
1023 if (mmu_ci_restrictions && psize == MMU_PAGE_64K &&
1024 (pte_val(*ptep) & _PAGE_NO_CACHE)) {
1026 demote_segment_4k(mm, ea);
1027 psize = MMU_PAGE_4K;
1028 } else if (ea < VMALLOC_END) {
1030 * some driver did a non-cacheable mapping
1031 * in vmalloc space, so switch vmalloc
1034 printk(KERN_ALERT "Reducing vmalloc segment "
1035 "to 4kB pages because of "
1036 "non-cacheable mapping\n");
1037 psize = mmu_vmalloc_psize = MMU_PAGE_4K;
1038 #ifdef CONFIG_SPU_BASE
1039 spu_flush_all_slbs(mm);
1044 if (psize != get_paca_psize(ea)) {
1045 get_paca()->context = mm->context;
1046 slb_flush_and_rebolt();
1048 } else if (get_paca()->vmalloc_sllp !=
1049 mmu_psize_defs[mmu_vmalloc_psize].sllp) {
1050 get_paca()->vmalloc_sllp =
1051 mmu_psize_defs[mmu_vmalloc_psize].sllp;
1052 slb_vmalloc_update();
1054 #endif /* CONFIG_PPC_64K_PAGES */
1056 #ifdef CONFIG_PPC_HAS_HASH_64K
1057 if (psize == MMU_PAGE_64K)
1058 rc = __hash_page_64K(ea, access, vsid, ptep, trap, local, ssize);
1060 #endif /* CONFIG_PPC_HAS_HASH_64K */
1062 int spp = subpage_protection(mm, ea);
1066 rc = __hash_page_4K(ea, access, vsid, ptep, trap,
1070 /* Dump some info in case of hash insertion failure, they should
1071 * never happen so it is really useful to know if/when they do
1074 hash_failure_debug(ea, access, vsid, trap, ssize, psize,
1076 #ifndef CONFIG_PPC_64K_PAGES
1077 DBG_LOW(" o-pte: %016lx\n", pte_val(*ptep));
1079 DBG_LOW(" o-pte: %016lx %016lx\n", pte_val(*ptep),
1080 pte_val(*(ptep + PTRS_PER_PTE)));
1082 DBG_LOW(" -> rc=%d\n", rc);
1085 EXPORT_SYMBOL_GPL(hash_page);
1087 void hash_preload(struct mm_struct *mm, unsigned long ea,
1088 unsigned long access, unsigned long trap)
1093 unsigned long flags;
1094 int rc, ssize, local = 0;
1096 BUG_ON(REGION_ID(ea) != USER_REGION_ID);
1098 #ifdef CONFIG_PPC_MM_SLICES
1099 /* We only prefault standard pages for now */
1100 if (unlikely(get_slice_psize(mm, ea) != mm->context.user_psize))
1104 DBG_LOW("hash_preload(mm=%p, mm->pgdir=%p, ea=%016lx, access=%lx,"
1105 " trap=%lx\n", mm, mm->pgd, ea, access, trap);
1107 /* Get Linux PTE if available */
1111 ptep = find_linux_pte(pgdir, ea);
1115 #ifdef CONFIG_PPC_64K_PAGES
1116 /* If either _PAGE_4K_PFN or _PAGE_NO_CACHE is set (and we are on
1117 * a 64K kernel), then we don't preload, hash_page() will take
1118 * care of it once we actually try to access the page.
1119 * That way we don't have to duplicate all of the logic for segment
1120 * page size demotion here
1122 if (pte_val(*ptep) & (_PAGE_4K_PFN | _PAGE_NO_CACHE))
1124 #endif /* CONFIG_PPC_64K_PAGES */
1127 ssize = user_segment_size(ea);
1128 vsid = get_vsid(mm->context.id, ea, ssize);
1130 /* Hash doesn't like irqs */
1131 local_irq_save(flags);
1133 /* Is that local to this CPU ? */
1134 if (cpumask_equal(mm_cpumask(mm), cpumask_of(smp_processor_id())))
1138 #ifdef CONFIG_PPC_HAS_HASH_64K
1139 if (mm->context.user_psize == MMU_PAGE_64K)
1140 rc = __hash_page_64K(ea, access, vsid, ptep, trap, local, ssize);
1142 #endif /* CONFIG_PPC_HAS_HASH_64K */
1143 rc = __hash_page_4K(ea, access, vsid, ptep, trap, local, ssize,
1144 subpage_protection(mm, ea));
1146 /* Dump some info in case of hash insertion failure, they should
1147 * never happen so it is really useful to know if/when they do
1150 hash_failure_debug(ea, access, vsid, trap, ssize,
1151 mm->context.user_psize, pte_val(*ptep));
1153 local_irq_restore(flags);
1156 /* WARNING: This is called from hash_low_64.S, if you change this prototype,
1157 * do not forget to update the assembly call site !
1159 void flush_hash_page(unsigned long vpn, real_pte_t pte, int psize, int ssize,
1162 unsigned long hash, index, shift, hidx, slot;
1164 DBG_LOW("flush_hash_page(vpn=%016lx)\n", vpn);
1165 pte_iterate_hashed_subpages(pte, psize, vpn, index, shift) {
1166 hash = hpt_hash(vpn, shift, ssize);
1167 hidx = __rpte_to_hidx(pte, index);
1168 if (hidx & _PTEIDX_SECONDARY)
1170 slot = (hash & htab_hash_mask) * HPTES_PER_GROUP;
1171 slot += hidx & _PTEIDX_GROUP_IX;
1172 DBG_LOW(" sub %ld: hash=%lx, hidx=%lx\n", index, slot, hidx);
1173 ppc_md.hpte_invalidate(slot, vpn, psize, ssize, local);
1174 } pte_iterate_hashed_end();
1176 #ifdef CONFIG_PPC_TRANSACTIONAL_MEM
1177 /* Transactions are not aborted by tlbiel, only tlbie.
1178 * Without, syncing a page back to a block device w/ PIO could pick up
1179 * transactional data (bad!) so we force an abort here. Before the
1180 * sync the page will be made read-only, which will flush_hash_page.
1181 * BIG ISSUE here: if the kernel uses a page from userspace without
1182 * unmapping it first, it may see the speculated version.
1184 if (local && cpu_has_feature(CPU_FTR_TM) &&
1185 MSR_TM_ACTIVE(current->thread.regs->msr)) {
1187 tm_abort(TM_CAUSE_TLBI);
1192 void flush_hash_range(unsigned long number, int local)
1194 if (ppc_md.flush_hash_range)
1195 ppc_md.flush_hash_range(number, local);
1198 struct ppc64_tlb_batch *batch =
1199 &__get_cpu_var(ppc64_tlb_batch);
1201 for (i = 0; i < number; i++)
1202 flush_hash_page(batch->vpn[i], batch->pte[i],
1203 batch->psize, batch->ssize, local);
1208 * low_hash_fault is called when we the low level hash code failed
1209 * to instert a PTE due to an hypervisor error
1211 void low_hash_fault(struct pt_regs *regs, unsigned long address, int rc)
1213 if (user_mode(regs)) {
1214 #ifdef CONFIG_PPC_SUBPAGE_PROT
1216 _exception(SIGSEGV, regs, SEGV_ACCERR, address);
1219 _exception(SIGBUS, regs, BUS_ADRERR, address);
1221 bad_page_fault(regs, address, SIGBUS);
1224 #ifdef CONFIG_DEBUG_PAGEALLOC
1225 static void kernel_map_linear_page(unsigned long vaddr, unsigned long lmi)
1227 unsigned long hash, hpteg;
1228 unsigned long vsid = get_kernel_vsid(vaddr, mmu_kernel_ssize);
1229 unsigned long vpn = hpt_vpn(vaddr, vsid, mmu_kernel_ssize);
1230 unsigned long mode = htab_convert_pte_flags(PAGE_KERNEL);
1233 hash = hpt_hash(vpn, PAGE_SHIFT, mmu_kernel_ssize);
1234 hpteg = ((hash & htab_hash_mask) * HPTES_PER_GROUP);
1236 ret = ppc_md.hpte_insert(hpteg, vpn, __pa(vaddr),
1237 mode, HPTE_V_BOLTED,
1238 mmu_linear_psize, mmu_kernel_ssize);
1240 spin_lock(&linear_map_hash_lock);
1241 BUG_ON(linear_map_hash_slots[lmi] & 0x80);
1242 linear_map_hash_slots[lmi] = ret | 0x80;
1243 spin_unlock(&linear_map_hash_lock);
1246 static void kernel_unmap_linear_page(unsigned long vaddr, unsigned long lmi)
1248 unsigned long hash, hidx, slot;
1249 unsigned long vsid = get_kernel_vsid(vaddr, mmu_kernel_ssize);
1250 unsigned long vpn = hpt_vpn(vaddr, vsid, mmu_kernel_ssize);
1252 hash = hpt_hash(vpn, PAGE_SHIFT, mmu_kernel_ssize);
1253 spin_lock(&linear_map_hash_lock);
1254 BUG_ON(!(linear_map_hash_slots[lmi] & 0x80));
1255 hidx = linear_map_hash_slots[lmi] & 0x7f;
1256 linear_map_hash_slots[lmi] = 0;
1257 spin_unlock(&linear_map_hash_lock);
1258 if (hidx & _PTEIDX_SECONDARY)
1260 slot = (hash & htab_hash_mask) * HPTES_PER_GROUP;
1261 slot += hidx & _PTEIDX_GROUP_IX;
1262 ppc_md.hpte_invalidate(slot, vpn, mmu_linear_psize, mmu_kernel_ssize, 0);
1265 void kernel_map_pages(struct page *page, int numpages, int enable)
1267 unsigned long flags, vaddr, lmi;
1270 local_irq_save(flags);
1271 for (i = 0; i < numpages; i++, page++) {
1272 vaddr = (unsigned long)page_address(page);
1273 lmi = __pa(vaddr) >> PAGE_SHIFT;
1274 if (lmi >= linear_map_hash_count)
1277 kernel_map_linear_page(vaddr, lmi);
1279 kernel_unmap_linear_page(vaddr, lmi);
1281 local_irq_restore(flags);
1283 #endif /* CONFIG_DEBUG_PAGEALLOC */
1285 void setup_initial_memory_limit(phys_addr_t first_memblock_base,
1286 phys_addr_t first_memblock_size)
1288 /* We don't currently support the first MEMBLOCK not mapping 0
1289 * physical on those processors
1291 BUG_ON(first_memblock_base != 0);
1293 /* On LPAR systems, the first entry is our RMA region,
1294 * non-LPAR 64-bit hash MMU systems don't have a limitation
1295 * on real mode access, but using the first entry works well
1296 * enough. We also clamp it to 1G to avoid some funky things
1297 * such as RTAS bugs etc...
1299 ppc64_rma_size = min_t(u64, first_memblock_size, 0x40000000);
1301 /* Finally limit subsequent allocations */
1302 memblock_set_current_limit(ppc64_rma_size);