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>
36 #include <linux/context_tracking.h>
38 #include <asm/processor.h>
39 #include <asm/pgtable.h>
41 #include <asm/mmu_context.h>
43 #include <asm/types.h>
44 #include <asm/uaccess.h>
45 #include <asm/machdep.h>
47 #include <asm/tlbflush.h>
51 #include <asm/cacheflush.h>
52 #include <asm/cputable.h>
53 #include <asm/sections.h>
56 #include <asm/code-patching.h>
57 #include <asm/fadump.h>
58 #include <asm/firmware.h>
62 #define DBG(fmt...) udbg_printf(fmt)
68 #define DBG_LOW(fmt...) udbg_printf(fmt)
70 #define DBG_LOW(fmt...)
78 * Note: pte --> Linux PTE
79 * HPTE --> PowerPC Hashed Page Table Entry
82 * htab_initialize is called with the MMU off (of course), but
83 * the kernel has been copied down to zero so it can directly
84 * reference global data. At this point it is very difficult
85 * to print debug info.
90 extern unsigned long dart_tablebase;
91 #endif /* CONFIG_U3_DART */
93 static unsigned long _SDR1;
94 struct mmu_psize_def mmu_psize_defs[MMU_PAGE_COUNT];
96 struct hash_pte *htab_address;
97 unsigned long htab_size_bytes;
98 unsigned long htab_hash_mask;
99 EXPORT_SYMBOL_GPL(htab_hash_mask);
100 int mmu_linear_psize = MMU_PAGE_4K;
101 int mmu_virtual_psize = MMU_PAGE_4K;
102 int mmu_vmalloc_psize = MMU_PAGE_4K;
103 #ifdef CONFIG_SPARSEMEM_VMEMMAP
104 int mmu_vmemmap_psize = MMU_PAGE_4K;
106 int mmu_io_psize = MMU_PAGE_4K;
107 int mmu_kernel_ssize = MMU_SEGSIZE_256M;
108 int mmu_highuser_ssize = MMU_SEGSIZE_256M;
109 u16 mmu_slb_size = 64;
110 EXPORT_SYMBOL_GPL(mmu_slb_size);
111 #ifdef CONFIG_PPC_64K_PAGES
112 int mmu_ci_restrictions;
114 #ifdef CONFIG_DEBUG_PAGEALLOC
115 static u8 *linear_map_hash_slots;
116 static unsigned long linear_map_hash_count;
117 static DEFINE_SPINLOCK(linear_map_hash_lock);
118 #endif /* CONFIG_DEBUG_PAGEALLOC */
120 /* There are definitions of page sizes arrays to be used when none
121 * is provided by the firmware.
124 /* Pre-POWER4 CPUs (4k pages only)
126 static struct mmu_psize_def mmu_psize_defaults_old[] = {
130 .penc = {[MMU_PAGE_4K] = 0, [1 ... MMU_PAGE_COUNT - 1] = -1},
136 /* POWER4, GPUL, POWER5
138 * Support for 16Mb large pages
140 static struct mmu_psize_def mmu_psize_defaults_gp[] = {
144 .penc = {[MMU_PAGE_4K] = 0, [1 ... MMU_PAGE_COUNT - 1] = -1},
151 .penc = {[0 ... MMU_PAGE_16M - 1] = -1, [MMU_PAGE_16M] = 0,
152 [MMU_PAGE_16M + 1 ... MMU_PAGE_COUNT - 1] = -1 },
158 static unsigned long htab_convert_pte_flags(unsigned long pteflags)
160 unsigned long rflags = pteflags & 0x1fa;
162 /* _PAGE_EXEC -> NOEXEC */
163 if ((pteflags & _PAGE_EXEC) == 0)
166 /* PP bits. PAGE_USER is already PP bit 0x2, so we only
167 * need to add in 0x1 if it's a read-only user page
169 if ((pteflags & _PAGE_USER) && !((pteflags & _PAGE_RW) &&
170 (pteflags & _PAGE_DIRTY)))
174 return rflags | HPTE_R_C;
177 int htab_bolt_mapping(unsigned long vstart, unsigned long vend,
178 unsigned long pstart, unsigned long prot,
179 int psize, int ssize)
181 unsigned long vaddr, paddr;
182 unsigned int step, shift;
185 shift = mmu_psize_defs[psize].shift;
188 prot = htab_convert_pte_flags(prot);
190 DBG("htab_bolt_mapping(%lx..%lx -> %lx (%lx,%d,%d)\n",
191 vstart, vend, pstart, prot, psize, ssize);
193 for (vaddr = vstart, paddr = pstart; vaddr < vend;
194 vaddr += step, paddr += step) {
195 unsigned long hash, hpteg;
196 unsigned long vsid = get_kernel_vsid(vaddr, ssize);
197 unsigned long vpn = hpt_vpn(vaddr, vsid, ssize);
198 unsigned long tprot = prot;
201 * If we hit a bad address return error.
205 /* Make kernel text executable */
206 if (overlaps_kernel_text(vaddr, vaddr + step))
209 hash = hpt_hash(vpn, shift, ssize);
210 hpteg = ((hash & htab_hash_mask) * HPTES_PER_GROUP);
212 BUG_ON(!ppc_md.hpte_insert);
213 ret = ppc_md.hpte_insert(hpteg, vpn, paddr, tprot,
214 HPTE_V_BOLTED, psize, psize, ssize);
218 #ifdef CONFIG_DEBUG_PAGEALLOC
219 if ((paddr >> PAGE_SHIFT) < linear_map_hash_count)
220 linear_map_hash_slots[paddr >> PAGE_SHIFT] = ret | 0x80;
221 #endif /* CONFIG_DEBUG_PAGEALLOC */
223 return ret < 0 ? ret : 0;
226 #ifdef CONFIG_MEMORY_HOTPLUG
227 static int htab_remove_mapping(unsigned long vstart, unsigned long vend,
228 int psize, int ssize)
231 unsigned int step, shift;
233 shift = mmu_psize_defs[psize].shift;
236 if (!ppc_md.hpte_removebolted) {
237 printk(KERN_WARNING "Platform doesn't implement "
238 "hpte_removebolted\n");
242 for (vaddr = vstart; vaddr < vend; vaddr += step)
243 ppc_md.hpte_removebolted(vaddr, psize, ssize);
247 #endif /* CONFIG_MEMORY_HOTPLUG */
249 static int __init htab_dt_scan_seg_sizes(unsigned long node,
250 const char *uname, int depth,
253 char *type = of_get_flat_dt_prop(node, "device_type", NULL);
255 unsigned long size = 0;
257 /* We are scanning "cpu" nodes only */
258 if (type == NULL || strcmp(type, "cpu") != 0)
261 prop = of_get_flat_dt_prop(node, "ibm,processor-segment-sizes", &size);
264 for (; size >= 4; size -= 4, ++prop) {
265 if (be32_to_cpu(prop[0]) == 40) {
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 = of_get_flat_dt_prop(node, "ibm,segment-page-sizes", &size);
318 pr_info("Page sizes from device-tree:\n");
320 cur_cpu_spec->mmu_features &= ~(MMU_FTR_16M_PAGE);
322 unsigned int base_shift = be32_to_cpu(prop[0]);
323 unsigned int slbenc = be32_to_cpu(prop[1]);
324 unsigned int lpnum = be32_to_cpu(prop[2]);
325 struct mmu_psize_def *def;
328 size -= 3; prop += 3;
329 base_idx = get_idx_from_shift(base_shift);
332 * skip the pte encoding also
334 prop += lpnum * 2; size -= lpnum * 2;
337 def = &mmu_psize_defs[base_idx];
338 if (base_idx == MMU_PAGE_16M)
339 cur_cpu_spec->mmu_features |= MMU_FTR_16M_PAGE;
341 def->shift = base_shift;
342 if (base_shift <= 23)
345 def->avpnm = (1 << (base_shift - 23)) - 1;
348 * We don't know for sure what's up with tlbiel, so
349 * for now we only set it for 4K and 64K pages
351 if (base_idx == MMU_PAGE_4K || base_idx == MMU_PAGE_64K)
356 while (size > 0 && lpnum) {
357 unsigned int shift = be32_to_cpu(prop[0]);
358 int penc = be32_to_cpu(prop[1]);
360 prop += 2; size -= 2;
363 idx = get_idx_from_shift(shift);
368 pr_err("Invalid penc for base_shift=%d "
369 "shift=%d\n", base_shift, shift);
371 def->penc[idx] = penc;
372 pr_info("base_shift=%d: shift=%d, sllp=0x%04lx,"
373 " avpnm=0x%08lx, tlbiel=%d, penc=%d\n",
374 base_shift, shift, def->sllp,
375 def->avpnm, def->tlbiel, def->penc[idx]);
383 #ifdef CONFIG_HUGETLB_PAGE
384 /* Scan for 16G memory blocks that have been set aside for huge pages
385 * and reserve those blocks for 16G huge pages.
387 static int __init htab_dt_scan_hugepage_blocks(unsigned long node,
388 const char *uname, int depth,
390 char *type = of_get_flat_dt_prop(node, "device_type", NULL);
392 __be32 *page_count_prop;
393 unsigned int expected_pages;
394 long unsigned int phys_addr;
395 long unsigned int block_size;
397 /* We are scanning "memory" nodes only */
398 if (type == NULL || strcmp(type, "memory") != 0)
401 /* This property is the log base 2 of the number of virtual pages that
402 * will represent this memory block. */
403 page_count_prop = of_get_flat_dt_prop(node, "ibm,expected#pages", NULL);
404 if (page_count_prop == NULL)
406 expected_pages = (1 << be32_to_cpu(page_count_prop[0]));
407 addr_prop = of_get_flat_dt_prop(node, "reg", NULL);
408 if (addr_prop == NULL)
410 phys_addr = be64_to_cpu(addr_prop[0]);
411 block_size = be64_to_cpu(addr_prop[1]);
412 if (block_size != (16 * GB))
414 printk(KERN_INFO "Huge page(16GB) memory: "
415 "addr = 0x%lX size = 0x%lX pages = %d\n",
416 phys_addr, block_size, expected_pages);
417 if (phys_addr + (16 * GB) <= memblock_end_of_DRAM()) {
418 memblock_reserve(phys_addr, block_size * expected_pages);
419 add_gpage(phys_addr, block_size, expected_pages);
423 #endif /* CONFIG_HUGETLB_PAGE */
425 static void mmu_psize_set_default_penc(void)
428 for (bpsize = 0; bpsize < MMU_PAGE_COUNT; bpsize++)
429 for (apsize = 0; apsize < MMU_PAGE_COUNT; apsize++)
430 mmu_psize_defs[bpsize].penc[apsize] = -1;
433 static void __init htab_init_page_sizes(void)
437 /* se the invalid penc to -1 */
438 mmu_psize_set_default_penc();
440 /* Default to 4K pages only */
441 memcpy(mmu_psize_defs, mmu_psize_defaults_old,
442 sizeof(mmu_psize_defaults_old));
445 * Try to find the available page sizes in the device-tree
447 rc = of_scan_flat_dt(htab_dt_scan_page_sizes, NULL);
448 if (rc != 0) /* Found */
452 * Not in the device-tree, let's fallback on known size
453 * list for 16M capable GP & GR
455 if (mmu_has_feature(MMU_FTR_16M_PAGE))
456 memcpy(mmu_psize_defs, mmu_psize_defaults_gp,
457 sizeof(mmu_psize_defaults_gp));
459 #ifndef CONFIG_DEBUG_PAGEALLOC
461 * Pick a size for the linear mapping. Currently, we only support
462 * 16M, 1M and 4K which is the default
464 if (mmu_psize_defs[MMU_PAGE_16M].shift)
465 mmu_linear_psize = MMU_PAGE_16M;
466 else if (mmu_psize_defs[MMU_PAGE_1M].shift)
467 mmu_linear_psize = MMU_PAGE_1M;
468 #endif /* CONFIG_DEBUG_PAGEALLOC */
470 #ifdef CONFIG_PPC_64K_PAGES
472 * Pick a size for the ordinary pages. Default is 4K, we support
473 * 64K for user mappings and vmalloc if supported by the processor.
474 * We only use 64k for ioremap if the processor
475 * (and firmware) support cache-inhibited large pages.
476 * If not, we use 4k and set mmu_ci_restrictions so that
477 * hash_page knows to switch processes that use cache-inhibited
478 * mappings to 4k pages.
480 if (mmu_psize_defs[MMU_PAGE_64K].shift) {
481 mmu_virtual_psize = MMU_PAGE_64K;
482 mmu_vmalloc_psize = MMU_PAGE_64K;
483 if (mmu_linear_psize == MMU_PAGE_4K)
484 mmu_linear_psize = MMU_PAGE_64K;
485 if (mmu_has_feature(MMU_FTR_CI_LARGE_PAGE)) {
487 * Don't use 64k pages for ioremap on pSeries, since
488 * that would stop us accessing the HEA ethernet.
490 if (!machine_is(pseries))
491 mmu_io_psize = MMU_PAGE_64K;
493 mmu_ci_restrictions = 1;
495 #endif /* CONFIG_PPC_64K_PAGES */
497 #ifdef CONFIG_SPARSEMEM_VMEMMAP
498 /* We try to use 16M pages for vmemmap if that is supported
499 * and we have at least 1G of RAM at boot
501 if (mmu_psize_defs[MMU_PAGE_16M].shift &&
502 memblock_phys_mem_size() >= 0x40000000)
503 mmu_vmemmap_psize = MMU_PAGE_16M;
504 else if (mmu_psize_defs[MMU_PAGE_64K].shift)
505 mmu_vmemmap_psize = MMU_PAGE_64K;
507 mmu_vmemmap_psize = MMU_PAGE_4K;
508 #endif /* CONFIG_SPARSEMEM_VMEMMAP */
510 printk(KERN_DEBUG "Page orders: linear mapping = %d, "
511 "virtual = %d, io = %d"
512 #ifdef CONFIG_SPARSEMEM_VMEMMAP
516 mmu_psize_defs[mmu_linear_psize].shift,
517 mmu_psize_defs[mmu_virtual_psize].shift,
518 mmu_psize_defs[mmu_io_psize].shift
519 #ifdef CONFIG_SPARSEMEM_VMEMMAP
520 ,mmu_psize_defs[mmu_vmemmap_psize].shift
524 #ifdef CONFIG_HUGETLB_PAGE
525 /* Reserve 16G huge page memory sections for huge pages */
526 of_scan_flat_dt(htab_dt_scan_hugepage_blocks, NULL);
527 #endif /* CONFIG_HUGETLB_PAGE */
530 static int __init htab_dt_scan_pftsize(unsigned long node,
531 const char *uname, int depth,
534 char *type = of_get_flat_dt_prop(node, "device_type", NULL);
537 /* We are scanning "cpu" nodes only */
538 if (type == NULL || strcmp(type, "cpu") != 0)
541 prop = of_get_flat_dt_prop(node, "ibm,pft-size", NULL);
543 /* pft_size[0] is the NUMA CEC cookie */
544 ppc64_pft_size = be32_to_cpu(prop[1]);
550 static unsigned long __init htab_get_table_size(void)
552 unsigned long mem_size, rnd_mem_size, pteg_count, psize;
554 /* If hash size isn't already provided by the platform, we try to
555 * retrieve it from the device-tree. If it's not there neither, we
556 * calculate it now based on the total RAM size
558 if (ppc64_pft_size == 0)
559 of_scan_flat_dt(htab_dt_scan_pftsize, NULL);
561 return 1UL << ppc64_pft_size;
563 /* round mem_size up to next power of 2 */
564 mem_size = memblock_phys_mem_size();
565 rnd_mem_size = 1UL << __ilog2(mem_size);
566 if (rnd_mem_size < mem_size)
570 psize = mmu_psize_defs[mmu_virtual_psize].shift;
571 pteg_count = max(rnd_mem_size >> (psize + 1), 1UL << 11);
573 return pteg_count << 7;
576 #ifdef CONFIG_MEMORY_HOTPLUG
577 int create_section_mapping(unsigned long start, unsigned long end)
579 return htab_bolt_mapping(start, end, __pa(start),
580 pgprot_val(PAGE_KERNEL), mmu_linear_psize,
584 int remove_section_mapping(unsigned long start, unsigned long end)
586 return htab_remove_mapping(start, end, mmu_linear_psize,
589 #endif /* CONFIG_MEMORY_HOTPLUG */
591 #define FUNCTION_TEXT(A) ((*(unsigned long *)(A)))
593 static void __init htab_finish_init(void)
595 extern unsigned int *htab_call_hpte_insert1;
596 extern unsigned int *htab_call_hpte_insert2;
597 extern unsigned int *htab_call_hpte_remove;
598 extern unsigned int *htab_call_hpte_updatepp;
600 #ifdef CONFIG_PPC_HAS_HASH_64K
601 extern unsigned int *ht64_call_hpte_insert1;
602 extern unsigned int *ht64_call_hpte_insert2;
603 extern unsigned int *ht64_call_hpte_remove;
604 extern unsigned int *ht64_call_hpte_updatepp;
606 patch_branch(ht64_call_hpte_insert1,
607 FUNCTION_TEXT(ppc_md.hpte_insert),
609 patch_branch(ht64_call_hpte_insert2,
610 FUNCTION_TEXT(ppc_md.hpte_insert),
612 patch_branch(ht64_call_hpte_remove,
613 FUNCTION_TEXT(ppc_md.hpte_remove),
615 patch_branch(ht64_call_hpte_updatepp,
616 FUNCTION_TEXT(ppc_md.hpte_updatepp),
619 #endif /* CONFIG_PPC_HAS_HASH_64K */
621 patch_branch(htab_call_hpte_insert1,
622 FUNCTION_TEXT(ppc_md.hpte_insert),
624 patch_branch(htab_call_hpte_insert2,
625 FUNCTION_TEXT(ppc_md.hpte_insert),
627 patch_branch(htab_call_hpte_remove,
628 FUNCTION_TEXT(ppc_md.hpte_remove),
630 patch_branch(htab_call_hpte_updatepp,
631 FUNCTION_TEXT(ppc_md.hpte_updatepp),
635 static void __init htab_initialize(void)
638 unsigned long pteg_count;
640 unsigned long base = 0, size = 0, limit;
641 struct memblock_region *reg;
643 DBG(" -> htab_initialize()\n");
645 /* Initialize segment sizes */
646 htab_init_seg_sizes();
648 /* Initialize page sizes */
649 htab_init_page_sizes();
651 if (mmu_has_feature(MMU_FTR_1T_SEGMENT)) {
652 mmu_kernel_ssize = MMU_SEGSIZE_1T;
653 mmu_highuser_ssize = MMU_SEGSIZE_1T;
654 printk(KERN_INFO "Using 1TB segments\n");
658 * Calculate the required size of the htab. We want the number of
659 * PTEGs to equal one half the number of real pages.
661 htab_size_bytes = htab_get_table_size();
662 pteg_count = htab_size_bytes >> 7;
664 htab_hash_mask = pteg_count - 1;
666 if (firmware_has_feature(FW_FEATURE_LPAR)) {
667 /* Using a hypervisor which owns the htab */
670 #ifdef CONFIG_FA_DUMP
672 * If firmware assisted dump is active firmware preserves
673 * the contents of htab along with entire partition memory.
674 * Clear the htab if firmware assisted dump is active so
675 * that we dont end up using old mappings.
677 if (is_fadump_active() && ppc_md.hpte_clear_all)
678 ppc_md.hpte_clear_all();
681 /* Find storage for the HPT. Must be contiguous in
682 * the absolute address space. On cell we want it to be
683 * in the first 2 Gig so we can use it for IOMMU hacks.
685 if (machine_is(cell))
688 limit = MEMBLOCK_ALLOC_ANYWHERE;
690 table = memblock_alloc_base(htab_size_bytes, htab_size_bytes, limit);
692 DBG("Hash table allocated at %lx, size: %lx\n", table,
695 htab_address = __va(table);
697 /* htab absolute addr + encoded htabsize */
698 _SDR1 = table + __ilog2(pteg_count) - 11;
700 /* Initialize the HPT with no entries */
701 memset((void *)table, 0, htab_size_bytes);
704 mtspr(SPRN_SDR1, _SDR1);
707 prot = pgprot_val(PAGE_KERNEL);
709 #ifdef CONFIG_DEBUG_PAGEALLOC
710 linear_map_hash_count = memblock_end_of_DRAM() >> PAGE_SHIFT;
711 linear_map_hash_slots = __va(memblock_alloc_base(linear_map_hash_count,
713 memset(linear_map_hash_slots, 0, linear_map_hash_count);
714 #endif /* CONFIG_DEBUG_PAGEALLOC */
716 /* On U3 based machines, we need to reserve the DART area and
717 * _NOT_ map it to avoid cache paradoxes as it's remapped non
721 /* create bolted the linear mapping in the hash table */
722 for_each_memblock(memory, reg) {
723 base = (unsigned long)__va(reg->base);
726 DBG("creating mapping for region: %lx..%lx (prot: %lx)\n",
729 #ifdef CONFIG_U3_DART
730 /* Do not map the DART space. Fortunately, it will be aligned
731 * in such a way that it will not cross two memblock regions and
732 * will fit within a single 16Mb page.
733 * The DART space is assumed to be a full 16Mb region even if
734 * we only use 2Mb of that space. We will use more of it later
735 * for AGP GART. We have to use a full 16Mb large page.
737 DBG("DART base: %lx\n", dart_tablebase);
739 if (dart_tablebase != 0 && dart_tablebase >= base
740 && dart_tablebase < (base + size)) {
741 unsigned long dart_table_end = dart_tablebase + 16 * MB;
742 if (base != dart_tablebase)
743 BUG_ON(htab_bolt_mapping(base, dart_tablebase,
747 if ((base + size) > dart_table_end)
748 BUG_ON(htab_bolt_mapping(dart_tablebase+16*MB,
750 __pa(dart_table_end),
756 #endif /* CONFIG_U3_DART */
757 BUG_ON(htab_bolt_mapping(base, base + size, __pa(base),
758 prot, mmu_linear_psize, mmu_kernel_ssize));
760 memblock_set_current_limit(MEMBLOCK_ALLOC_ANYWHERE);
763 * If we have a memory_limit and we've allocated TCEs then we need to
764 * explicitly map the TCE area at the top of RAM. We also cope with the
765 * case that the TCEs start below memory_limit.
766 * tce_alloc_start/end are 16MB aligned so the mapping should work
767 * for either 4K or 16MB pages.
769 if (tce_alloc_start) {
770 tce_alloc_start = (unsigned long)__va(tce_alloc_start);
771 tce_alloc_end = (unsigned long)__va(tce_alloc_end);
773 if (base + size >= tce_alloc_start)
774 tce_alloc_start = base + size + 1;
776 BUG_ON(htab_bolt_mapping(tce_alloc_start, tce_alloc_end,
777 __pa(tce_alloc_start), prot,
778 mmu_linear_psize, mmu_kernel_ssize));
783 DBG(" <- htab_initialize()\n");
788 void __init early_init_mmu(void)
790 /* Setup initial STAB address in the PACA */
791 get_paca()->stab_real = __pa((u64)&initial_stab);
792 get_paca()->stab_addr = (u64)&initial_stab;
794 /* Initialize the MMU Hash table and create the linear mapping
795 * of memory. Has to be done before stab/slb initialization as
796 * this is currently where the page size encoding is obtained
800 /* Initialize stab / SLB management */
801 if (mmu_has_feature(MMU_FTR_SLB))
804 stab_initialize(get_paca()->stab_real);
808 void early_init_mmu_secondary(void)
810 /* Initialize hash table for that CPU */
811 if (!firmware_has_feature(FW_FEATURE_LPAR))
812 mtspr(SPRN_SDR1, _SDR1);
814 /* Initialize STAB/SLB. We use a virtual address as it works
815 * in real mode on pSeries.
817 if (mmu_has_feature(MMU_FTR_SLB))
820 stab_initialize(get_paca()->stab_addr);
822 #endif /* CONFIG_SMP */
825 * Called by asm hashtable.S for doing lazy icache flush
827 unsigned int hash_page_do_lazy_icache(unsigned int pp, pte_t pte, int trap)
831 if (!pfn_valid(pte_pfn(pte)))
834 page = pte_page(pte);
837 if (!test_bit(PG_arch_1, &page->flags) && !PageReserved(page)) {
839 flush_dcache_icache_page(page);
840 set_bit(PG_arch_1, &page->flags);
847 #ifdef CONFIG_PPC_MM_SLICES
848 unsigned int get_paca_psize(unsigned long addr)
851 unsigned char *hpsizes;
852 unsigned long index, mask_index;
854 if (addr < SLICE_LOW_TOP) {
855 lpsizes = get_paca()->context.low_slices_psize;
856 index = GET_LOW_SLICE_INDEX(addr);
857 return (lpsizes >> (index * 4)) & 0xF;
859 hpsizes = get_paca()->context.high_slices_psize;
860 index = GET_HIGH_SLICE_INDEX(addr);
861 mask_index = index & 0x1;
862 return (hpsizes[index >> 1] >> (mask_index * 4)) & 0xF;
866 unsigned int get_paca_psize(unsigned long addr)
868 return get_paca()->context.user_psize;
873 * Demote a segment to using 4k pages.
874 * For now this makes the whole process use 4k pages.
876 #ifdef CONFIG_PPC_64K_PAGES
877 void demote_segment_4k(struct mm_struct *mm, unsigned long addr)
879 if (get_slice_psize(mm, addr) == MMU_PAGE_4K)
881 slice_set_range_psize(mm, addr, 1, MMU_PAGE_4K);
882 #ifdef CONFIG_SPU_BASE
883 spu_flush_all_slbs(mm);
885 if (get_paca_psize(addr) != MMU_PAGE_4K) {
886 get_paca()->context = mm->context;
887 slb_flush_and_rebolt();
890 #endif /* CONFIG_PPC_64K_PAGES */
892 #ifdef CONFIG_PPC_SUBPAGE_PROT
894 * This looks up a 2-bit protection code for a 4k subpage of a 64k page.
895 * Userspace sets the subpage permissions using the subpage_prot system call.
897 * Result is 0: full permissions, _PAGE_RW: read-only,
898 * _PAGE_USER or _PAGE_USER|_PAGE_RW: no access.
900 static int subpage_protection(struct mm_struct *mm, unsigned long ea)
902 struct subpage_prot_table *spt = &mm->context.spt;
906 if (ea >= spt->maxaddr)
908 if (ea < 0x100000000UL) {
909 /* addresses below 4GB use spt->low_prot */
910 sbpm = spt->low_prot;
912 sbpm = spt->protptrs[ea >> SBP_L3_SHIFT];
916 sbpp = sbpm[(ea >> SBP_L2_SHIFT) & (SBP_L2_COUNT - 1)];
919 spp = sbpp[(ea >> PAGE_SHIFT) & (SBP_L1_COUNT - 1)];
921 /* extract 2-bit bitfield for this 4k subpage */
922 spp >>= 30 - 2 * ((ea >> 12) & 0xf);
924 /* turn 0,1,2,3 into combination of _PAGE_USER and _PAGE_RW */
925 spp = ((spp & 2) ? _PAGE_USER : 0) | ((spp & 1) ? _PAGE_RW : 0);
929 #else /* CONFIG_PPC_SUBPAGE_PROT */
930 static inline int subpage_protection(struct mm_struct *mm, unsigned long ea)
936 void hash_failure_debug(unsigned long ea, unsigned long access,
937 unsigned long vsid, unsigned long trap,
938 int ssize, int psize, int lpsize, unsigned long pte)
940 if (!printk_ratelimit())
942 pr_info("mm: Hashing failure ! EA=0x%lx access=0x%lx current=%s\n",
943 ea, access, current->comm);
944 pr_info(" trap=0x%lx vsid=0x%lx ssize=%d base psize=%d psize %d pte=0x%lx\n",
945 trap, vsid, ssize, psize, lpsize, pte);
950 * 1 - normal page fault
951 * -1 - critical hash insertion error
952 * -2 - access not permitted by subpage protection mechanism
954 int hash_page(unsigned long ea, unsigned long access, unsigned long trap)
956 enum ctx_state prev_state = exception_enter();
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");
979 psize = get_slice_psize(mm, ea);
980 ssize = user_segment_size(ea);
981 vsid = get_vsid(mm->context.id, ea, ssize);
983 case VMALLOC_REGION_ID:
985 vsid = get_kernel_vsid(ea, mmu_kernel_ssize);
986 if (ea < VMALLOC_END)
987 psize = mmu_vmalloc_psize;
989 psize = mmu_io_psize;
990 ssize = mmu_kernel_ssize;
994 * Send the problem up to do_page_fault
999 DBG_LOW(" mm=%p, mm->pgdir=%p, vsid=%016lx\n", mm, mm->pgd, vsid);
1003 DBG_LOW("Bad address!\n");
1009 if (pgdir == NULL) {
1014 /* Check CPU locality */
1015 tmp = cpumask_of(smp_processor_id());
1016 if (user_region && cpumask_equal(mm_cpumask(mm), tmp))
1019 #ifndef CONFIG_PPC_64K_PAGES
1020 /* If we use 4K pages and our psize is not 4K, then we might
1021 * be hitting a special driver mapping, and need to align the
1022 * address before we fetch the PTE.
1024 * It could also be a hugepage mapping, in which case this is
1025 * not necessary, but it's not harmful, either.
1027 if (psize != MMU_PAGE_4K)
1028 ea &= ~((1ul << mmu_psize_defs[psize].shift) - 1);
1029 #endif /* CONFIG_PPC_64K_PAGES */
1031 /* Get PTE and page size from page tables */
1032 ptep = find_linux_pte_or_hugepte(pgdir, ea, &hugeshift);
1033 if (ptep == NULL || !pte_present(*ptep)) {
1034 DBG_LOW(" no PTE !\n");
1039 /* Add _PAGE_PRESENT to the required access perm */
1040 access |= _PAGE_PRESENT;
1042 /* Pre-check access permissions (will be re-checked atomically
1043 * in __hash_page_XX but this pre-check is a fast path
1045 if (access & ~pte_val(*ptep)) {
1046 DBG_LOW(" no access !\n");
1052 if (pmd_trans_huge(*(pmd_t *)ptep))
1053 rc = __hash_page_thp(ea, access, vsid, (pmd_t *)ptep,
1054 trap, local, ssize, psize);
1055 #ifdef CONFIG_HUGETLB_PAGE
1057 rc = __hash_page_huge(ea, access, vsid, ptep, trap,
1058 local, ssize, hugeshift, psize);
1062 * if we have hugeshift, and is not transhuge with
1063 * hugetlb disabled, something is really wrong.
1072 #ifndef CONFIG_PPC_64K_PAGES
1073 DBG_LOW(" i-pte: %016lx\n", pte_val(*ptep));
1075 DBG_LOW(" i-pte: %016lx %016lx\n", pte_val(*ptep),
1076 pte_val(*(ptep + PTRS_PER_PTE)));
1078 /* Do actual hashing */
1079 #ifdef CONFIG_PPC_64K_PAGES
1080 /* If _PAGE_4K_PFN is set, make sure this is a 4k segment */
1081 if ((pte_val(*ptep) & _PAGE_4K_PFN) && psize == MMU_PAGE_64K) {
1082 demote_segment_4k(mm, ea);
1083 psize = MMU_PAGE_4K;
1086 /* If this PTE is non-cacheable and we have restrictions on
1087 * using non cacheable large pages, then we switch to 4k
1089 if (mmu_ci_restrictions && psize == MMU_PAGE_64K &&
1090 (pte_val(*ptep) & _PAGE_NO_CACHE)) {
1092 demote_segment_4k(mm, ea);
1093 psize = MMU_PAGE_4K;
1094 } else if (ea < VMALLOC_END) {
1096 * some driver did a non-cacheable mapping
1097 * in vmalloc space, so switch vmalloc
1100 printk(KERN_ALERT "Reducing vmalloc segment "
1101 "to 4kB pages because of "
1102 "non-cacheable mapping\n");
1103 psize = mmu_vmalloc_psize = MMU_PAGE_4K;
1104 #ifdef CONFIG_SPU_BASE
1105 spu_flush_all_slbs(mm);
1110 if (psize != get_paca_psize(ea)) {
1111 get_paca()->context = mm->context;
1112 slb_flush_and_rebolt();
1114 } else if (get_paca()->vmalloc_sllp !=
1115 mmu_psize_defs[mmu_vmalloc_psize].sllp) {
1116 get_paca()->vmalloc_sllp =
1117 mmu_psize_defs[mmu_vmalloc_psize].sllp;
1118 slb_vmalloc_update();
1120 #endif /* CONFIG_PPC_64K_PAGES */
1122 #ifdef CONFIG_PPC_HAS_HASH_64K
1123 if (psize == MMU_PAGE_64K)
1124 rc = __hash_page_64K(ea, access, vsid, ptep, trap, local, ssize);
1126 #endif /* CONFIG_PPC_HAS_HASH_64K */
1128 int spp = subpage_protection(mm, ea);
1132 rc = __hash_page_4K(ea, access, vsid, ptep, trap,
1136 /* Dump some info in case of hash insertion failure, they should
1137 * never happen so it is really useful to know if/when they do
1140 hash_failure_debug(ea, access, vsid, trap, ssize, psize,
1141 psize, pte_val(*ptep));
1142 #ifndef CONFIG_PPC_64K_PAGES
1143 DBG_LOW(" o-pte: %016lx\n", pte_val(*ptep));
1145 DBG_LOW(" o-pte: %016lx %016lx\n", pte_val(*ptep),
1146 pte_val(*(ptep + PTRS_PER_PTE)));
1148 DBG_LOW(" -> rc=%d\n", rc);
1151 exception_exit(prev_state);
1154 EXPORT_SYMBOL_GPL(hash_page);
1156 void hash_preload(struct mm_struct *mm, unsigned long ea,
1157 unsigned long access, unsigned long trap)
1163 unsigned long flags;
1164 int rc, ssize, local = 0;
1166 BUG_ON(REGION_ID(ea) != USER_REGION_ID);
1168 #ifdef CONFIG_PPC_MM_SLICES
1169 /* We only prefault standard pages for now */
1170 if (unlikely(get_slice_psize(mm, ea) != mm->context.user_psize))
1174 DBG_LOW("hash_preload(mm=%p, mm->pgdir=%p, ea=%016lx, access=%lx,"
1175 " trap=%lx\n", mm, mm->pgd, ea, access, trap);
1177 /* Get Linux PTE if available */
1183 ssize = user_segment_size(ea);
1184 vsid = get_vsid(mm->context.id, ea, ssize);
1188 * Hash doesn't like irqs. Walking linux page table with irq disabled
1189 * saves us from holding multiple locks.
1191 local_irq_save(flags);
1194 * THP pages use update_mmu_cache_pmd. We don't do
1195 * hash preload there. Hence can ignore THP here
1197 ptep = find_linux_pte_or_hugepte(pgdir, ea, &hugepage_shift);
1201 WARN_ON(hugepage_shift);
1202 #ifdef CONFIG_PPC_64K_PAGES
1203 /* If either _PAGE_4K_PFN or _PAGE_NO_CACHE is set (and we are on
1204 * a 64K kernel), then we don't preload, hash_page() will take
1205 * care of it once we actually try to access the page.
1206 * That way we don't have to duplicate all of the logic for segment
1207 * page size demotion here
1209 if (pte_val(*ptep) & (_PAGE_4K_PFN | _PAGE_NO_CACHE))
1211 #endif /* CONFIG_PPC_64K_PAGES */
1213 /* Is that local to this CPU ? */
1214 if (cpumask_equal(mm_cpumask(mm), cpumask_of(smp_processor_id())))
1218 #ifdef CONFIG_PPC_HAS_HASH_64K
1219 if (mm->context.user_psize == MMU_PAGE_64K)
1220 rc = __hash_page_64K(ea, access, vsid, ptep, trap, local, ssize);
1222 #endif /* CONFIG_PPC_HAS_HASH_64K */
1223 rc = __hash_page_4K(ea, access, vsid, ptep, trap, local, ssize,
1224 subpage_protection(mm, ea));
1226 /* Dump some info in case of hash insertion failure, they should
1227 * never happen so it is really useful to know if/when they do
1230 hash_failure_debug(ea, access, vsid, trap, ssize,
1231 mm->context.user_psize,
1232 mm->context.user_psize,
1235 local_irq_restore(flags);
1238 /* WARNING: This is called from hash_low_64.S, if you change this prototype,
1239 * do not forget to update the assembly call site !
1241 void flush_hash_page(unsigned long vpn, real_pte_t pte, int psize, int ssize,
1244 unsigned long hash, index, shift, hidx, slot;
1246 DBG_LOW("flush_hash_page(vpn=%016lx)\n", vpn);
1247 pte_iterate_hashed_subpages(pte, psize, vpn, index, shift) {
1248 hash = hpt_hash(vpn, shift, ssize);
1249 hidx = __rpte_to_hidx(pte, index);
1250 if (hidx & _PTEIDX_SECONDARY)
1252 slot = (hash & htab_hash_mask) * HPTES_PER_GROUP;
1253 slot += hidx & _PTEIDX_GROUP_IX;
1254 DBG_LOW(" sub %ld: hash=%lx, hidx=%lx\n", index, slot, hidx);
1256 * We use same base page size and actual psize, because we don't
1257 * use these functions for hugepage
1259 ppc_md.hpte_invalidate(slot, vpn, psize, psize, ssize, local);
1260 } pte_iterate_hashed_end();
1262 #ifdef CONFIG_PPC_TRANSACTIONAL_MEM
1263 /* Transactions are not aborted by tlbiel, only tlbie.
1264 * Without, syncing a page back to a block device w/ PIO could pick up
1265 * transactional data (bad!) so we force an abort here. Before the
1266 * sync the page will be made read-only, which will flush_hash_page.
1267 * BIG ISSUE here: if the kernel uses a page from userspace without
1268 * unmapping it first, it may see the speculated version.
1270 if (local && cpu_has_feature(CPU_FTR_TM) &&
1271 current->thread.regs &&
1272 MSR_TM_ACTIVE(current->thread.regs->msr)) {
1274 tm_abort(TM_CAUSE_TLBI);
1279 void flush_hash_range(unsigned long number, int local)
1281 if (ppc_md.flush_hash_range)
1282 ppc_md.flush_hash_range(number, local);
1285 struct ppc64_tlb_batch *batch =
1286 &__get_cpu_var(ppc64_tlb_batch);
1288 for (i = 0; i < number; i++)
1289 flush_hash_page(batch->vpn[i], batch->pte[i],
1290 batch->psize, batch->ssize, local);
1295 * low_hash_fault is called when we the low level hash code failed
1296 * to instert a PTE due to an hypervisor error
1298 void low_hash_fault(struct pt_regs *regs, unsigned long address, int rc)
1300 enum ctx_state prev_state = exception_enter();
1302 if (user_mode(regs)) {
1303 #ifdef CONFIG_PPC_SUBPAGE_PROT
1305 _exception(SIGSEGV, regs, SEGV_ACCERR, address);
1308 _exception(SIGBUS, regs, BUS_ADRERR, address);
1310 bad_page_fault(regs, address, SIGBUS);
1312 exception_exit(prev_state);
1315 long hpte_insert_repeating(unsigned long hash, unsigned long vpn,
1316 unsigned long pa, unsigned long rflags,
1317 unsigned long vflags, int psize, int ssize)
1319 unsigned long hpte_group;
1323 hpte_group = ((hash & htab_hash_mask) *
1324 HPTES_PER_GROUP) & ~0x7UL;
1326 /* Insert into the hash table, primary slot */
1327 slot = ppc_md.hpte_insert(hpte_group, vpn, pa, rflags, vflags,
1328 psize, psize, ssize);
1330 /* Primary is full, try the secondary */
1331 if (unlikely(slot == -1)) {
1332 hpte_group = ((~hash & htab_hash_mask) *
1333 HPTES_PER_GROUP) & ~0x7UL;
1334 slot = ppc_md.hpte_insert(hpte_group, vpn, pa, rflags,
1335 vflags | HPTE_V_SECONDARY,
1336 psize, psize, ssize);
1339 hpte_group = ((hash & htab_hash_mask) *
1340 HPTES_PER_GROUP)&~0x7UL;
1342 ppc_md.hpte_remove(hpte_group);
1350 #ifdef CONFIG_DEBUG_PAGEALLOC
1351 static void kernel_map_linear_page(unsigned long vaddr, unsigned long lmi)
1354 unsigned long vsid = get_kernel_vsid(vaddr, mmu_kernel_ssize);
1355 unsigned long vpn = hpt_vpn(vaddr, vsid, mmu_kernel_ssize);
1356 unsigned long mode = htab_convert_pte_flags(PAGE_KERNEL);
1359 hash = hpt_hash(vpn, PAGE_SHIFT, mmu_kernel_ssize);
1361 /* Don't create HPTE entries for bad address */
1365 ret = hpte_insert_repeating(hash, vpn, __pa(vaddr), mode,
1367 mmu_linear_psize, mmu_kernel_ssize);
1370 spin_lock(&linear_map_hash_lock);
1371 BUG_ON(linear_map_hash_slots[lmi] & 0x80);
1372 linear_map_hash_slots[lmi] = ret | 0x80;
1373 spin_unlock(&linear_map_hash_lock);
1376 static void kernel_unmap_linear_page(unsigned long vaddr, unsigned long lmi)
1378 unsigned long hash, hidx, slot;
1379 unsigned long vsid = get_kernel_vsid(vaddr, mmu_kernel_ssize);
1380 unsigned long vpn = hpt_vpn(vaddr, vsid, mmu_kernel_ssize);
1382 hash = hpt_hash(vpn, PAGE_SHIFT, mmu_kernel_ssize);
1383 spin_lock(&linear_map_hash_lock);
1384 BUG_ON(!(linear_map_hash_slots[lmi] & 0x80));
1385 hidx = linear_map_hash_slots[lmi] & 0x7f;
1386 linear_map_hash_slots[lmi] = 0;
1387 spin_unlock(&linear_map_hash_lock);
1388 if (hidx & _PTEIDX_SECONDARY)
1390 slot = (hash & htab_hash_mask) * HPTES_PER_GROUP;
1391 slot += hidx & _PTEIDX_GROUP_IX;
1392 ppc_md.hpte_invalidate(slot, vpn, mmu_linear_psize, mmu_linear_psize,
1393 mmu_kernel_ssize, 0);
1396 void kernel_map_pages(struct page *page, int numpages, int enable)
1398 unsigned long flags, vaddr, lmi;
1401 local_irq_save(flags);
1402 for (i = 0; i < numpages; i++, page++) {
1403 vaddr = (unsigned long)page_address(page);
1404 lmi = __pa(vaddr) >> PAGE_SHIFT;
1405 if (lmi >= linear_map_hash_count)
1408 kernel_map_linear_page(vaddr, lmi);
1410 kernel_unmap_linear_page(vaddr, lmi);
1412 local_irq_restore(flags);
1414 #endif /* CONFIG_DEBUG_PAGEALLOC */
1416 void setup_initial_memory_limit(phys_addr_t first_memblock_base,
1417 phys_addr_t first_memblock_size)
1419 /* We don't currently support the first MEMBLOCK not mapping 0
1420 * physical on those processors
1422 BUG_ON(first_memblock_base != 0);
1424 /* On LPAR systems, the first entry is our RMA region,
1425 * non-LPAR 64-bit hash MMU systems don't have a limitation
1426 * on real mode access, but using the first entry works well
1427 * enough. We also clamp it to 1G to avoid some funky things
1428 * such as RTAS bugs etc...
1430 ppc64_rma_size = min_t(u64, first_memblock_size, 0x40000000);
1432 /* Finally limit subsequent allocations */
1433 memblock_set_current_limit(ppc64_rma_size);