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[] = {
129 .penc = {[MMU_PAGE_4K] = 0, [1 ... MMU_PAGE_COUNT - 1] = -1},
135 /* POWER4, GPUL, POWER5
137 * Support for 16Mb large pages
139 static struct mmu_psize_def mmu_psize_defaults_gp[] = {
143 .penc = {[MMU_PAGE_4K] = 0, [1 ... MMU_PAGE_COUNT - 1] = -1},
150 .penc = {[0 ... MMU_PAGE_16M - 1] = -1, [MMU_PAGE_16M] = 0,
151 [MMU_PAGE_16M + 1 ... MMU_PAGE_COUNT - 1] = -1 },
157 static unsigned long htab_convert_pte_flags(unsigned long pteflags)
159 unsigned long rflags = pteflags & 0x1fa;
161 /* _PAGE_EXEC -> NOEXEC */
162 if ((pteflags & _PAGE_EXEC) == 0)
165 /* PP bits. PAGE_USER is already PP bit 0x2, so we only
166 * need to add in 0x1 if it's a read-only user page
168 if ((pteflags & _PAGE_USER) && !((pteflags & _PAGE_RW) &&
169 (pteflags & _PAGE_DIRTY)))
173 return rflags | HPTE_R_C;
176 int htab_bolt_mapping(unsigned long vstart, unsigned long vend,
177 unsigned long pstart, unsigned long prot,
178 int psize, int ssize)
180 unsigned long vaddr, paddr;
181 unsigned int step, shift;
184 shift = mmu_psize_defs[psize].shift;
187 prot = htab_convert_pte_flags(prot);
189 DBG("htab_bolt_mapping(%lx..%lx -> %lx (%lx,%d,%d)\n",
190 vstart, vend, pstart, prot, psize, ssize);
192 for (vaddr = vstart, paddr = pstart; vaddr < vend;
193 vaddr += step, paddr += step) {
194 unsigned long hash, hpteg;
195 unsigned long vsid = get_kernel_vsid(vaddr, ssize);
196 unsigned long vpn = hpt_vpn(vaddr, vsid, ssize);
197 unsigned long tprot = prot;
200 * If we hit a bad address return error.
204 /* Make kernel text executable */
205 if (overlaps_kernel_text(vaddr, vaddr + step))
208 hash = hpt_hash(vpn, shift, ssize);
209 hpteg = ((hash & htab_hash_mask) * HPTES_PER_GROUP);
211 BUG_ON(!ppc_md.hpte_insert);
212 ret = ppc_md.hpte_insert(hpteg, vpn, paddr, tprot,
213 HPTE_V_BOLTED, psize, psize, ssize);
217 #ifdef CONFIG_DEBUG_PAGEALLOC
218 if ((paddr >> PAGE_SHIFT) < linear_map_hash_count)
219 linear_map_hash_slots[paddr >> PAGE_SHIFT] = ret | 0x80;
220 #endif /* CONFIG_DEBUG_PAGEALLOC */
222 return ret < 0 ? ret : 0;
225 #ifdef CONFIG_MEMORY_HOTPLUG
226 static int htab_remove_mapping(unsigned long vstart, unsigned long vend,
227 int psize, int ssize)
230 unsigned int step, shift;
232 shift = mmu_psize_defs[psize].shift;
235 if (!ppc_md.hpte_removebolted) {
236 printk(KERN_WARNING "Platform doesn't implement "
237 "hpte_removebolted\n");
241 for (vaddr = vstart; vaddr < vend; vaddr += step)
242 ppc_md.hpte_removebolted(vaddr, psize, ssize);
246 #endif /* CONFIG_MEMORY_HOTPLUG */
248 static int __init htab_dt_scan_seg_sizes(unsigned long node,
249 const char *uname, int depth,
252 char *type = of_get_flat_dt_prop(node, "device_type", NULL);
254 unsigned long size = 0;
256 /* We are scanning "cpu" nodes only */
257 if (type == NULL || strcmp(type, "cpu") != 0)
260 prop = (u32 *)of_get_flat_dt_prop(node, "ibm,processor-segment-sizes",
264 for (; size >= 4; size -= 4, ++prop) {
266 DBG("1T segment support detected\n");
267 cur_cpu_spec->mmu_features |= MMU_FTR_1T_SEGMENT;
271 cur_cpu_spec->mmu_features &= ~MMU_FTR_NO_SLBIE_B;
275 static void __init htab_init_seg_sizes(void)
277 of_scan_flat_dt(htab_dt_scan_seg_sizes, NULL);
280 static int __init get_idx_from_shift(unsigned int shift)
304 static int __init htab_dt_scan_page_sizes(unsigned long node,
305 const char *uname, int depth,
308 char *type = of_get_flat_dt_prop(node, "device_type", NULL);
310 unsigned long size = 0;
312 /* We are scanning "cpu" nodes only */
313 if (type == NULL || strcmp(type, "cpu") != 0)
316 prop = (u32 *)of_get_flat_dt_prop(node,
317 "ibm,segment-page-sizes", &size);
319 DBG("Page sizes from device-tree:\n");
321 cur_cpu_spec->mmu_features &= ~(MMU_FTR_16M_PAGE);
323 unsigned int base_shift = prop[0];
324 unsigned int slbenc = prop[1];
325 unsigned int lpnum = prop[2];
326 struct mmu_psize_def *def;
329 size -= 3; prop += 3;
330 base_idx = get_idx_from_shift(base_shift);
333 * skip the pte encoding also
335 prop += lpnum * 2; size -= lpnum * 2;
338 def = &mmu_psize_defs[base_idx];
339 if (base_idx == MMU_PAGE_16M)
340 cur_cpu_spec->mmu_features |= MMU_FTR_16M_PAGE;
342 def->shift = base_shift;
343 if (base_shift <= 23)
346 def->avpnm = (1 << (base_shift - 23)) - 1;
349 * We don't know for sure what's up with tlbiel, so
350 * for now we only set it for 4K and 64K pages
352 if (base_idx == MMU_PAGE_4K || base_idx == MMU_PAGE_64K)
357 while (size > 0 && lpnum) {
358 unsigned int shift = prop[0];
361 prop += 2; size -= 2;
364 idx = get_idx_from_shift(shift);
369 pr_err("Invalid penc for base_shift=%d "
370 "shift=%d\n", base_shift, shift);
372 def->penc[idx] = penc;
373 DBG(" %d: shift=%02x, sllp=%04lx, "
374 "avpnm=%08lx, tlbiel=%d, penc=%d\n",
375 idx, shift, def->sllp, def->avpnm,
376 def->tlbiel, def->penc[idx]);
384 #ifdef CONFIG_HUGETLB_PAGE
385 /* Scan for 16G memory blocks that have been set aside for huge pages
386 * and reserve those blocks for 16G huge pages.
388 static int __init htab_dt_scan_hugepage_blocks(unsigned long node,
389 const char *uname, int depth,
391 char *type = of_get_flat_dt_prop(node, "device_type", NULL);
392 unsigned long *addr_prop;
393 u32 *page_count_prop;
394 unsigned int expected_pages;
395 long unsigned int phys_addr;
396 long unsigned int block_size;
398 /* We are scanning "memory" nodes only */
399 if (type == NULL || strcmp(type, "memory") != 0)
402 /* This property is the log base 2 of the number of virtual pages that
403 * will represent this memory block. */
404 page_count_prop = of_get_flat_dt_prop(node, "ibm,expected#pages", NULL);
405 if (page_count_prop == NULL)
407 expected_pages = (1 << page_count_prop[0]);
408 addr_prop = of_get_flat_dt_prop(node, "reg", NULL);
409 if (addr_prop == NULL)
411 phys_addr = addr_prop[0];
412 block_size = addr_prop[1];
413 if (block_size != (16 * GB))
415 printk(KERN_INFO "Huge page(16GB) memory: "
416 "addr = 0x%lX size = 0x%lX pages = %d\n",
417 phys_addr, block_size, expected_pages);
418 if (phys_addr + (16 * GB) <= memblock_end_of_DRAM()) {
419 memblock_reserve(phys_addr, block_size * expected_pages);
420 add_gpage(phys_addr, block_size, expected_pages);
424 #endif /* CONFIG_HUGETLB_PAGE */
426 static void mmu_psize_set_default_penc(void)
429 for (bpsize = 0; bpsize < MMU_PAGE_COUNT; bpsize++)
430 for (apsize = 0; apsize < MMU_PAGE_COUNT; apsize++)
431 mmu_psize_defs[bpsize].penc[apsize] = -1;
434 static void __init htab_init_page_sizes(void)
438 /* se the invalid penc to -1 */
439 mmu_psize_set_default_penc();
441 /* Default to 4K pages only */
442 memcpy(mmu_psize_defs, mmu_psize_defaults_old,
443 sizeof(mmu_psize_defaults_old));
446 * Try to find the available page sizes in the device-tree
448 rc = of_scan_flat_dt(htab_dt_scan_page_sizes, NULL);
449 if (rc != 0) /* Found */
453 * Not in the device-tree, let's fallback on known size
454 * list for 16M capable GP & GR
456 if (mmu_has_feature(MMU_FTR_16M_PAGE))
457 memcpy(mmu_psize_defs, mmu_psize_defaults_gp,
458 sizeof(mmu_psize_defaults_gp));
460 #ifndef CONFIG_DEBUG_PAGEALLOC
462 * Pick a size for the linear mapping. Currently, we only support
463 * 16M, 1M and 4K which is the default
465 if (mmu_psize_defs[MMU_PAGE_16M].shift)
466 mmu_linear_psize = MMU_PAGE_16M;
467 else if (mmu_psize_defs[MMU_PAGE_1M].shift)
468 mmu_linear_psize = MMU_PAGE_1M;
469 #endif /* CONFIG_DEBUG_PAGEALLOC */
471 #ifdef CONFIG_PPC_64K_PAGES
473 * Pick a size for the ordinary pages. Default is 4K, we support
474 * 64K for user mappings and vmalloc if supported by the processor.
475 * We only use 64k for ioremap if the processor
476 * (and firmware) support cache-inhibited large pages.
477 * If not, we use 4k and set mmu_ci_restrictions so that
478 * hash_page knows to switch processes that use cache-inhibited
479 * mappings to 4k pages.
481 if (mmu_psize_defs[MMU_PAGE_64K].shift) {
482 mmu_virtual_psize = MMU_PAGE_64K;
483 mmu_vmalloc_psize = MMU_PAGE_64K;
484 if (mmu_linear_psize == MMU_PAGE_4K)
485 mmu_linear_psize = MMU_PAGE_64K;
486 if (mmu_has_feature(MMU_FTR_CI_LARGE_PAGE)) {
488 * Don't use 64k pages for ioremap on pSeries, since
489 * that would stop us accessing the HEA ethernet.
491 if (!machine_is(pseries))
492 mmu_io_psize = MMU_PAGE_64K;
494 mmu_ci_restrictions = 1;
496 #endif /* CONFIG_PPC_64K_PAGES */
498 #ifdef CONFIG_SPARSEMEM_VMEMMAP
499 /* We try to use 16M pages for vmemmap if that is supported
500 * and we have at least 1G of RAM at boot
502 if (mmu_psize_defs[MMU_PAGE_16M].shift &&
503 memblock_phys_mem_size() >= 0x40000000)
504 mmu_vmemmap_psize = MMU_PAGE_16M;
505 else if (mmu_psize_defs[MMU_PAGE_64K].shift)
506 mmu_vmemmap_psize = MMU_PAGE_64K;
508 mmu_vmemmap_psize = MMU_PAGE_4K;
509 #endif /* CONFIG_SPARSEMEM_VMEMMAP */
511 printk(KERN_DEBUG "Page orders: linear mapping = %d, "
512 "virtual = %d, io = %d"
513 #ifdef CONFIG_SPARSEMEM_VMEMMAP
517 mmu_psize_defs[mmu_linear_psize].shift,
518 mmu_psize_defs[mmu_virtual_psize].shift,
519 mmu_psize_defs[mmu_io_psize].shift
520 #ifdef CONFIG_SPARSEMEM_VMEMMAP
521 ,mmu_psize_defs[mmu_vmemmap_psize].shift
525 #ifdef CONFIG_HUGETLB_PAGE
526 /* Reserve 16G huge page memory sections for huge pages */
527 of_scan_flat_dt(htab_dt_scan_hugepage_blocks, NULL);
528 #endif /* CONFIG_HUGETLB_PAGE */
531 static int __init htab_dt_scan_pftsize(unsigned long node,
532 const char *uname, int depth,
535 char *type = of_get_flat_dt_prop(node, "device_type", NULL);
538 /* We are scanning "cpu" nodes only */
539 if (type == NULL || strcmp(type, "cpu") != 0)
542 prop = (u32 *)of_get_flat_dt_prop(node, "ibm,pft-size", NULL);
544 /* pft_size[0] is the NUMA CEC cookie */
545 ppc64_pft_size = prop[1];
551 static unsigned long __init htab_get_table_size(void)
553 unsigned long mem_size, rnd_mem_size, pteg_count, psize;
555 /* If hash size isn't already provided by the platform, we try to
556 * retrieve it from the device-tree. If it's not there neither, we
557 * calculate it now based on the total RAM size
559 if (ppc64_pft_size == 0)
560 of_scan_flat_dt(htab_dt_scan_pftsize, NULL);
562 return 1UL << ppc64_pft_size;
564 /* round mem_size up to next power of 2 */
565 mem_size = memblock_phys_mem_size();
566 rnd_mem_size = 1UL << __ilog2(mem_size);
567 if (rnd_mem_size < mem_size)
571 psize = mmu_psize_defs[mmu_virtual_psize].shift;
572 pteg_count = max(rnd_mem_size >> (psize + 1), 1UL << 11);
574 return pteg_count << 7;
577 #ifdef CONFIG_MEMORY_HOTPLUG
578 int create_section_mapping(unsigned long start, unsigned long end)
580 return htab_bolt_mapping(start, end, __pa(start),
581 pgprot_val(PAGE_KERNEL), mmu_linear_psize,
585 int remove_section_mapping(unsigned long start, unsigned long end)
587 return htab_remove_mapping(start, end, mmu_linear_psize,
590 #endif /* CONFIG_MEMORY_HOTPLUG */
592 #define FUNCTION_TEXT(A) ((*(unsigned long *)(A)))
594 static void __init htab_finish_init(void)
596 extern unsigned int *htab_call_hpte_insert1;
597 extern unsigned int *htab_call_hpte_insert2;
598 extern unsigned int *htab_call_hpte_remove;
599 extern unsigned int *htab_call_hpte_updatepp;
601 #ifdef CONFIG_PPC_HAS_HASH_64K
602 extern unsigned int *ht64_call_hpte_insert1;
603 extern unsigned int *ht64_call_hpte_insert2;
604 extern unsigned int *ht64_call_hpte_remove;
605 extern unsigned int *ht64_call_hpte_updatepp;
607 patch_branch(ht64_call_hpte_insert1,
608 FUNCTION_TEXT(ppc_md.hpte_insert),
610 patch_branch(ht64_call_hpte_insert2,
611 FUNCTION_TEXT(ppc_md.hpte_insert),
613 patch_branch(ht64_call_hpte_remove,
614 FUNCTION_TEXT(ppc_md.hpte_remove),
616 patch_branch(ht64_call_hpte_updatepp,
617 FUNCTION_TEXT(ppc_md.hpte_updatepp),
620 #endif /* CONFIG_PPC_HAS_HASH_64K */
622 patch_branch(htab_call_hpte_insert1,
623 FUNCTION_TEXT(ppc_md.hpte_insert),
625 patch_branch(htab_call_hpte_insert2,
626 FUNCTION_TEXT(ppc_md.hpte_insert),
628 patch_branch(htab_call_hpte_remove,
629 FUNCTION_TEXT(ppc_md.hpte_remove),
631 patch_branch(htab_call_hpte_updatepp,
632 FUNCTION_TEXT(ppc_md.hpte_updatepp),
636 static void __init htab_initialize(void)
639 unsigned long pteg_count;
641 unsigned long base = 0, size = 0, limit;
642 struct memblock_region *reg;
644 DBG(" -> htab_initialize()\n");
646 /* Initialize segment sizes */
647 htab_init_seg_sizes();
649 /* Initialize page sizes */
650 htab_init_page_sizes();
652 if (mmu_has_feature(MMU_FTR_1T_SEGMENT)) {
653 mmu_kernel_ssize = MMU_SEGSIZE_1T;
654 mmu_highuser_ssize = MMU_SEGSIZE_1T;
655 printk(KERN_INFO "Using 1TB segments\n");
659 * Calculate the required size of the htab. We want the number of
660 * PTEGs to equal one half the number of real pages.
662 htab_size_bytes = htab_get_table_size();
663 pteg_count = htab_size_bytes >> 7;
665 htab_hash_mask = pteg_count - 1;
667 if (firmware_has_feature(FW_FEATURE_LPAR)) {
668 /* Using a hypervisor which owns the htab */
671 #ifdef CONFIG_FA_DUMP
673 * If firmware assisted dump is active firmware preserves
674 * the contents of htab along with entire partition memory.
675 * Clear the htab if firmware assisted dump is active so
676 * that we dont end up using old mappings.
678 if (is_fadump_active() && ppc_md.hpte_clear_all)
679 ppc_md.hpte_clear_all();
682 /* Find storage for the HPT. Must be contiguous in
683 * the absolute address space. On cell we want it to be
684 * in the first 2 Gig so we can use it for IOMMU hacks.
686 if (machine_is(cell))
689 limit = MEMBLOCK_ALLOC_ANYWHERE;
691 table = memblock_alloc_base(htab_size_bytes, htab_size_bytes, limit);
693 DBG("Hash table allocated at %lx, size: %lx\n", table,
696 htab_address = __va(table);
698 /* htab absolute addr + encoded htabsize */
699 _SDR1 = table + __ilog2(pteg_count) - 11;
701 /* Initialize the HPT with no entries */
702 memset((void *)table, 0, htab_size_bytes);
705 mtspr(SPRN_SDR1, _SDR1);
708 prot = pgprot_val(PAGE_KERNEL);
710 #ifdef CONFIG_DEBUG_PAGEALLOC
711 linear_map_hash_count = memblock_end_of_DRAM() >> PAGE_SHIFT;
712 linear_map_hash_slots = __va(memblock_alloc_base(linear_map_hash_count,
714 memset(linear_map_hash_slots, 0, linear_map_hash_count);
715 #endif /* CONFIG_DEBUG_PAGEALLOC */
717 /* On U3 based machines, we need to reserve the DART area and
718 * _NOT_ map it to avoid cache paradoxes as it's remapped non
722 /* create bolted the linear mapping in the hash table */
723 for_each_memblock(memory, reg) {
724 base = (unsigned long)__va(reg->base);
727 DBG("creating mapping for region: %lx..%lx (prot: %lx)\n",
730 #ifdef CONFIG_U3_DART
731 /* Do not map the DART space. Fortunately, it will be aligned
732 * in such a way that it will not cross two memblock regions and
733 * will fit within a single 16Mb page.
734 * The DART space is assumed to be a full 16Mb region even if
735 * we only use 2Mb of that space. We will use more of it later
736 * for AGP GART. We have to use a full 16Mb large page.
738 DBG("DART base: %lx\n", dart_tablebase);
740 if (dart_tablebase != 0 && dart_tablebase >= base
741 && dart_tablebase < (base + size)) {
742 unsigned long dart_table_end = dart_tablebase + 16 * MB;
743 if (base != dart_tablebase)
744 BUG_ON(htab_bolt_mapping(base, dart_tablebase,
748 if ((base + size) > dart_table_end)
749 BUG_ON(htab_bolt_mapping(dart_tablebase+16*MB,
751 __pa(dart_table_end),
757 #endif /* CONFIG_U3_DART */
758 BUG_ON(htab_bolt_mapping(base, base + size, __pa(base),
759 prot, mmu_linear_psize, mmu_kernel_ssize));
761 memblock_set_current_limit(MEMBLOCK_ALLOC_ANYWHERE);
764 * If we have a memory_limit and we've allocated TCEs then we need to
765 * explicitly map the TCE area at the top of RAM. We also cope with the
766 * case that the TCEs start below memory_limit.
767 * tce_alloc_start/end are 16MB aligned so the mapping should work
768 * for either 4K or 16MB pages.
770 if (tce_alloc_start) {
771 tce_alloc_start = (unsigned long)__va(tce_alloc_start);
772 tce_alloc_end = (unsigned long)__va(tce_alloc_end);
774 if (base + size >= tce_alloc_start)
775 tce_alloc_start = base + size + 1;
777 BUG_ON(htab_bolt_mapping(tce_alloc_start, tce_alloc_end,
778 __pa(tce_alloc_start), prot,
779 mmu_linear_psize, mmu_kernel_ssize));
784 DBG(" <- htab_initialize()\n");
789 void __init early_init_mmu(void)
791 /* Setup initial STAB address in the PACA */
792 get_paca()->stab_real = __pa((u64)&initial_stab);
793 get_paca()->stab_addr = (u64)&initial_stab;
795 /* Initialize the MMU Hash table and create the linear mapping
796 * of memory. Has to be done before stab/slb initialization as
797 * this is currently where the page size encoding is obtained
801 /* Initialize stab / SLB management */
802 if (mmu_has_feature(MMU_FTR_SLB))
805 stab_initialize(get_paca()->stab_real);
809 void __cpuinit early_init_mmu_secondary(void)
811 /* Initialize hash table for that CPU */
812 if (!firmware_has_feature(FW_FEATURE_LPAR))
813 mtspr(SPRN_SDR1, _SDR1);
815 /* Initialize STAB/SLB. We use a virtual address as it works
816 * in real mode on pSeries.
818 if (mmu_has_feature(MMU_FTR_SLB))
821 stab_initialize(get_paca()->stab_addr);
823 #endif /* CONFIG_SMP */
826 * Called by asm hashtable.S for doing lazy icache flush
828 unsigned int hash_page_do_lazy_icache(unsigned int pp, pte_t pte, int trap)
832 if (!pfn_valid(pte_pfn(pte)))
835 page = pte_page(pte);
838 if (!test_bit(PG_arch_1, &page->flags) && !PageReserved(page)) {
840 flush_dcache_icache_page(page);
841 set_bit(PG_arch_1, &page->flags);
848 #ifdef CONFIG_PPC_MM_SLICES
849 unsigned int get_paca_psize(unsigned long addr)
852 unsigned char *hpsizes;
853 unsigned long index, mask_index;
855 if (addr < SLICE_LOW_TOP) {
856 lpsizes = get_paca()->context.low_slices_psize;
857 index = GET_LOW_SLICE_INDEX(addr);
858 return (lpsizes >> (index * 4)) & 0xF;
860 hpsizes = get_paca()->context.high_slices_psize;
861 index = GET_HIGH_SLICE_INDEX(addr);
862 mask_index = index & 0x1;
863 return (hpsizes[index >> 1] >> (mask_index * 4)) & 0xF;
867 unsigned int get_paca_psize(unsigned long addr)
869 return get_paca()->context.user_psize;
874 * Demote a segment to using 4k pages.
875 * For now this makes the whole process use 4k pages.
877 #ifdef CONFIG_PPC_64K_PAGES
878 void demote_segment_4k(struct mm_struct *mm, unsigned long addr)
880 if (get_slice_psize(mm, addr) == MMU_PAGE_4K)
882 slice_set_range_psize(mm, addr, 1, MMU_PAGE_4K);
883 #ifdef CONFIG_SPU_BASE
884 spu_flush_all_slbs(mm);
886 if (get_paca_psize(addr) != MMU_PAGE_4K) {
887 get_paca()->context = mm->context;
888 slb_flush_and_rebolt();
891 #endif /* CONFIG_PPC_64K_PAGES */
893 #ifdef CONFIG_PPC_SUBPAGE_PROT
895 * This looks up a 2-bit protection code for a 4k subpage of a 64k page.
896 * Userspace sets the subpage permissions using the subpage_prot system call.
898 * Result is 0: full permissions, _PAGE_RW: read-only,
899 * _PAGE_USER or _PAGE_USER|_PAGE_RW: no access.
901 static int subpage_protection(struct mm_struct *mm, unsigned long ea)
903 struct subpage_prot_table *spt = &mm->context.spt;
907 if (ea >= spt->maxaddr)
909 if (ea < 0x100000000) {
910 /* addresses below 4GB use spt->low_prot */
911 sbpm = spt->low_prot;
913 sbpm = spt->protptrs[ea >> SBP_L3_SHIFT];
917 sbpp = sbpm[(ea >> SBP_L2_SHIFT) & (SBP_L2_COUNT - 1)];
920 spp = sbpp[(ea >> PAGE_SHIFT) & (SBP_L1_COUNT - 1)];
922 /* extract 2-bit bitfield for this 4k subpage */
923 spp >>= 30 - 2 * ((ea >> 12) & 0xf);
925 /* turn 0,1,2,3 into combination of _PAGE_USER and _PAGE_RW */
926 spp = ((spp & 2) ? _PAGE_USER : 0) | ((spp & 1) ? _PAGE_RW : 0);
930 #else /* CONFIG_PPC_SUBPAGE_PROT */
931 static inline int subpage_protection(struct mm_struct *mm, unsigned long ea)
937 void hash_failure_debug(unsigned long ea, unsigned long access,
938 unsigned long vsid, unsigned long trap,
939 int ssize, int psize, unsigned long pte)
941 if (!printk_ratelimit())
943 pr_info("mm: Hashing failure ! EA=0x%lx access=0x%lx current=%s\n",
944 ea, access, current->comm);
945 pr_info(" trap=0x%lx vsid=0x%lx ssize=%d psize=%d pte=0x%lx\n",
946 trap, vsid, ssize, psize, pte);
951 * 1 - normal page fault
952 * -1 - critical hash insertion error
953 * -2 - access not permitted by subpage protection mechanism
955 int hash_page(unsigned long ea, unsigned long access, unsigned long trap)
959 struct mm_struct *mm;
962 const struct cpumask *tmp;
963 int rc, user_region = 0, local = 0;
966 DBG_LOW("hash_page(ea=%016lx, access=%lx, trap=%lx\n",
969 /* Get region & vsid */
970 switch (REGION_ID(ea)) {
975 DBG_LOW(" user region with no mm !\n");
978 psize = get_slice_psize(mm, ea);
979 ssize = user_segment_size(ea);
980 vsid = get_vsid(mm->context.id, ea, ssize);
982 case VMALLOC_REGION_ID:
984 vsid = get_kernel_vsid(ea, mmu_kernel_ssize);
985 if (ea < VMALLOC_END)
986 psize = mmu_vmalloc_psize;
988 psize = mmu_io_psize;
989 ssize = mmu_kernel_ssize;
993 * Send the problem up to do_page_fault
997 DBG_LOW(" mm=%p, mm->pgdir=%p, vsid=%016lx\n", mm, mm->pgd, vsid);
1001 DBG_LOW("Bad address!\n");
1009 /* Check CPU locality */
1010 tmp = cpumask_of(smp_processor_id());
1011 if (user_region && cpumask_equal(mm_cpumask(mm), tmp))
1014 #ifndef CONFIG_PPC_64K_PAGES
1015 /* If we use 4K pages and our psize is not 4K, then we might
1016 * be hitting a special driver mapping, and need to align the
1017 * address before we fetch the PTE.
1019 * It could also be a hugepage mapping, in which case this is
1020 * not necessary, but it's not harmful, either.
1022 if (psize != MMU_PAGE_4K)
1023 ea &= ~((1ul << mmu_psize_defs[psize].shift) - 1);
1024 #endif /* CONFIG_PPC_64K_PAGES */
1026 /* Get PTE and page size from page tables */
1027 ptep = find_linux_pte_or_hugepte(pgdir, ea, &hugeshift);
1028 if (ptep == NULL || !pte_present(*ptep)) {
1029 DBG_LOW(" no PTE !\n");
1033 /* Add _PAGE_PRESENT to the required access perm */
1034 access |= _PAGE_PRESENT;
1036 /* Pre-check access permissions (will be re-checked atomically
1037 * in __hash_page_XX but this pre-check is a fast path
1039 if (access & ~pte_val(*ptep)) {
1040 DBG_LOW(" no access !\n");
1044 #ifdef CONFIG_HUGETLB_PAGE
1046 return __hash_page_huge(ea, access, vsid, ptep, trap, local,
1047 ssize, hugeshift, psize);
1048 #endif /* CONFIG_HUGETLB_PAGE */
1050 #ifndef CONFIG_PPC_64K_PAGES
1051 DBG_LOW(" i-pte: %016lx\n", pte_val(*ptep));
1053 DBG_LOW(" i-pte: %016lx %016lx\n", pte_val(*ptep),
1054 pte_val(*(ptep + PTRS_PER_PTE)));
1056 /* Do actual hashing */
1057 #ifdef CONFIG_PPC_64K_PAGES
1058 /* If _PAGE_4K_PFN is set, make sure this is a 4k segment */
1059 if ((pte_val(*ptep) & _PAGE_4K_PFN) && psize == MMU_PAGE_64K) {
1060 demote_segment_4k(mm, ea);
1061 psize = MMU_PAGE_4K;
1064 /* If this PTE is non-cacheable and we have restrictions on
1065 * using non cacheable large pages, then we switch to 4k
1067 if (mmu_ci_restrictions && psize == MMU_PAGE_64K &&
1068 (pte_val(*ptep) & _PAGE_NO_CACHE)) {
1070 demote_segment_4k(mm, ea);
1071 psize = MMU_PAGE_4K;
1072 } else if (ea < VMALLOC_END) {
1074 * some driver did a non-cacheable mapping
1075 * in vmalloc space, so switch vmalloc
1078 printk(KERN_ALERT "Reducing vmalloc segment "
1079 "to 4kB pages because of "
1080 "non-cacheable mapping\n");
1081 psize = mmu_vmalloc_psize = MMU_PAGE_4K;
1082 #ifdef CONFIG_SPU_BASE
1083 spu_flush_all_slbs(mm);
1088 if (psize != get_paca_psize(ea)) {
1089 get_paca()->context = mm->context;
1090 slb_flush_and_rebolt();
1092 } else if (get_paca()->vmalloc_sllp !=
1093 mmu_psize_defs[mmu_vmalloc_psize].sllp) {
1094 get_paca()->vmalloc_sllp =
1095 mmu_psize_defs[mmu_vmalloc_psize].sllp;
1096 slb_vmalloc_update();
1098 #endif /* CONFIG_PPC_64K_PAGES */
1100 #ifdef CONFIG_PPC_HAS_HASH_64K
1101 if (psize == MMU_PAGE_64K)
1102 rc = __hash_page_64K(ea, access, vsid, ptep, trap, local, ssize);
1104 #endif /* CONFIG_PPC_HAS_HASH_64K */
1106 int spp = subpage_protection(mm, ea);
1110 rc = __hash_page_4K(ea, access, vsid, ptep, trap,
1114 /* Dump some info in case of hash insertion failure, they should
1115 * never happen so it is really useful to know if/when they do
1118 hash_failure_debug(ea, access, vsid, trap, ssize, psize,
1120 #ifndef CONFIG_PPC_64K_PAGES
1121 DBG_LOW(" o-pte: %016lx\n", pte_val(*ptep));
1123 DBG_LOW(" o-pte: %016lx %016lx\n", pte_val(*ptep),
1124 pte_val(*(ptep + PTRS_PER_PTE)));
1126 DBG_LOW(" -> rc=%d\n", rc);
1129 EXPORT_SYMBOL_GPL(hash_page);
1131 void hash_preload(struct mm_struct *mm, unsigned long ea,
1132 unsigned long access, unsigned long trap)
1137 unsigned long flags;
1138 int rc, ssize, local = 0;
1140 BUG_ON(REGION_ID(ea) != USER_REGION_ID);
1142 #ifdef CONFIG_PPC_MM_SLICES
1143 /* We only prefault standard pages for now */
1144 if (unlikely(get_slice_psize(mm, ea) != mm->context.user_psize))
1148 DBG_LOW("hash_preload(mm=%p, mm->pgdir=%p, ea=%016lx, access=%lx,"
1149 " trap=%lx\n", mm, mm->pgd, ea, access, trap);
1151 /* Get Linux PTE if available */
1155 ptep = find_linux_pte(pgdir, ea);
1159 #ifdef CONFIG_PPC_64K_PAGES
1160 /* If either _PAGE_4K_PFN or _PAGE_NO_CACHE is set (and we are on
1161 * a 64K kernel), then we don't preload, hash_page() will take
1162 * care of it once we actually try to access the page.
1163 * That way we don't have to duplicate all of the logic for segment
1164 * page size demotion here
1166 if (pte_val(*ptep) & (_PAGE_4K_PFN | _PAGE_NO_CACHE))
1168 #endif /* CONFIG_PPC_64K_PAGES */
1171 ssize = user_segment_size(ea);
1172 vsid = get_vsid(mm->context.id, ea, ssize);
1176 /* Hash doesn't like irqs */
1177 local_irq_save(flags);
1179 /* Is that local to this CPU ? */
1180 if (cpumask_equal(mm_cpumask(mm), cpumask_of(smp_processor_id())))
1184 #ifdef CONFIG_PPC_HAS_HASH_64K
1185 if (mm->context.user_psize == MMU_PAGE_64K)
1186 rc = __hash_page_64K(ea, access, vsid, ptep, trap, local, ssize);
1188 #endif /* CONFIG_PPC_HAS_HASH_64K */
1189 rc = __hash_page_4K(ea, access, vsid, ptep, trap, local, ssize,
1190 subpage_protection(mm, ea));
1192 /* Dump some info in case of hash insertion failure, they should
1193 * never happen so it is really useful to know if/when they do
1196 hash_failure_debug(ea, access, vsid, trap, ssize,
1197 mm->context.user_psize, pte_val(*ptep));
1199 local_irq_restore(flags);
1202 /* WARNING: This is called from hash_low_64.S, if you change this prototype,
1203 * do not forget to update the assembly call site !
1205 void flush_hash_page(unsigned long vpn, real_pte_t pte, int psize, int ssize,
1208 unsigned long hash, index, shift, hidx, slot;
1210 DBG_LOW("flush_hash_page(vpn=%016lx)\n", vpn);
1211 pte_iterate_hashed_subpages(pte, psize, vpn, index, shift) {
1212 hash = hpt_hash(vpn, shift, ssize);
1213 hidx = __rpte_to_hidx(pte, index);
1214 if (hidx & _PTEIDX_SECONDARY)
1216 slot = (hash & htab_hash_mask) * HPTES_PER_GROUP;
1217 slot += hidx & _PTEIDX_GROUP_IX;
1218 DBG_LOW(" sub %ld: hash=%lx, hidx=%lx\n", index, slot, hidx);
1219 ppc_md.hpte_invalidate(slot, vpn, psize, ssize, local);
1220 } pte_iterate_hashed_end();
1222 #ifdef CONFIG_PPC_TRANSACTIONAL_MEM
1223 /* Transactions are not aborted by tlbiel, only tlbie.
1224 * Without, syncing a page back to a block device w/ PIO could pick up
1225 * transactional data (bad!) so we force an abort here. Before the
1226 * sync the page will be made read-only, which will flush_hash_page.
1227 * BIG ISSUE here: if the kernel uses a page from userspace without
1228 * unmapping it first, it may see the speculated version.
1230 if (local && cpu_has_feature(CPU_FTR_TM) &&
1231 MSR_TM_ACTIVE(current->thread.regs->msr)) {
1233 tm_abort(TM_CAUSE_TLBI);
1238 void flush_hash_range(unsigned long number, int local)
1240 if (ppc_md.flush_hash_range)
1241 ppc_md.flush_hash_range(number, local);
1244 struct ppc64_tlb_batch *batch =
1245 &__get_cpu_var(ppc64_tlb_batch);
1247 for (i = 0; i < number; i++)
1248 flush_hash_page(batch->vpn[i], batch->pte[i],
1249 batch->psize, batch->ssize, local);
1254 * low_hash_fault is called when we the low level hash code failed
1255 * to instert a PTE due to an hypervisor error
1257 void low_hash_fault(struct pt_regs *regs, unsigned long address, int rc)
1259 if (user_mode(regs)) {
1260 #ifdef CONFIG_PPC_SUBPAGE_PROT
1262 _exception(SIGSEGV, regs, SEGV_ACCERR, address);
1265 _exception(SIGBUS, regs, BUS_ADRERR, address);
1267 bad_page_fault(regs, address, SIGBUS);
1270 long hpte_insert_repeating(unsigned long hash, unsigned long vpn,
1271 unsigned long pa, unsigned long rflags,
1272 unsigned long vflags, int psize, int ssize)
1274 unsigned long hpte_group;
1278 hpte_group = ((hash & htab_hash_mask) *
1279 HPTES_PER_GROUP) & ~0x7UL;
1281 /* Insert into the hash table, primary slot */
1282 slot = ppc_md.hpte_insert(hpte_group, vpn, pa, rflags, vflags,
1283 psize, psize, ssize);
1285 /* Primary is full, try the secondary */
1286 if (unlikely(slot == -1)) {
1287 hpte_group = ((~hash & htab_hash_mask) *
1288 HPTES_PER_GROUP) & ~0x7UL;
1289 slot = ppc_md.hpte_insert(hpte_group, vpn, pa, rflags,
1290 vflags | HPTE_V_SECONDARY,
1291 psize, psize, ssize);
1294 hpte_group = ((hash & htab_hash_mask) *
1295 HPTES_PER_GROUP)&~0x7UL;
1297 ppc_md.hpte_remove(hpte_group);
1305 #ifdef CONFIG_DEBUG_PAGEALLOC
1306 static void kernel_map_linear_page(unsigned long vaddr, unsigned long lmi)
1309 unsigned long vsid = get_kernel_vsid(vaddr, mmu_kernel_ssize);
1310 unsigned long vpn = hpt_vpn(vaddr, vsid, mmu_kernel_ssize);
1311 unsigned long mode = htab_convert_pte_flags(PAGE_KERNEL);
1314 hash = hpt_hash(vpn, PAGE_SHIFT, mmu_kernel_ssize);
1316 /* Don't create HPTE entries for bad address */
1320 ret = hpte_insert_repeating(hash, vpn, __pa(vaddr), mode,
1322 mmu_linear_psize, mmu_kernel_ssize);
1325 spin_lock(&linear_map_hash_lock);
1326 BUG_ON(linear_map_hash_slots[lmi] & 0x80);
1327 linear_map_hash_slots[lmi] = ret | 0x80;
1328 spin_unlock(&linear_map_hash_lock);
1331 static void kernel_unmap_linear_page(unsigned long vaddr, unsigned long lmi)
1333 unsigned long hash, hidx, slot;
1334 unsigned long vsid = get_kernel_vsid(vaddr, mmu_kernel_ssize);
1335 unsigned long vpn = hpt_vpn(vaddr, vsid, mmu_kernel_ssize);
1337 hash = hpt_hash(vpn, PAGE_SHIFT, mmu_kernel_ssize);
1338 spin_lock(&linear_map_hash_lock);
1339 BUG_ON(!(linear_map_hash_slots[lmi] & 0x80));
1340 hidx = linear_map_hash_slots[lmi] & 0x7f;
1341 linear_map_hash_slots[lmi] = 0;
1342 spin_unlock(&linear_map_hash_lock);
1343 if (hidx & _PTEIDX_SECONDARY)
1345 slot = (hash & htab_hash_mask) * HPTES_PER_GROUP;
1346 slot += hidx & _PTEIDX_GROUP_IX;
1347 ppc_md.hpte_invalidate(slot, vpn, mmu_linear_psize, mmu_kernel_ssize, 0);
1350 void kernel_map_pages(struct page *page, int numpages, int enable)
1352 unsigned long flags, vaddr, lmi;
1355 local_irq_save(flags);
1356 for (i = 0; i < numpages; i++, page++) {
1357 vaddr = (unsigned long)page_address(page);
1358 lmi = __pa(vaddr) >> PAGE_SHIFT;
1359 if (lmi >= linear_map_hash_count)
1362 kernel_map_linear_page(vaddr, lmi);
1364 kernel_unmap_linear_page(vaddr, lmi);
1366 local_irq_restore(flags);
1368 #endif /* CONFIG_DEBUG_PAGEALLOC */
1370 void setup_initial_memory_limit(phys_addr_t first_memblock_base,
1371 phys_addr_t first_memblock_size)
1373 /* We don't currently support the first MEMBLOCK not mapping 0
1374 * physical on those processors
1376 BUG_ON(first_memblock_base != 0);
1378 /* On LPAR systems, the first entry is our RMA region,
1379 * non-LPAR 64-bit hash MMU systems don't have a limitation
1380 * on real mode access, but using the first entry works well
1381 * enough. We also clamp it to 1G to avoid some funky things
1382 * such as RTAS bugs etc...
1384 ppc64_rma_size = min_t(u64, first_memblock_size, 0x40000000);
1386 /* Finally limit subsequent allocations */
1387 memblock_set_current_limit(ppc64_rma_size);