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/ctype.h>
31 #include <linux/cache.h>
32 #include <linux/init.h>
33 #include <linux/signal.h>
35 #include <asm/processor.h>
36 #include <asm/pgtable.h>
38 #include <asm/mmu_context.h>
40 #include <asm/types.h>
41 #include <asm/system.h>
42 #include <asm/uaccess.h>
43 #include <asm/machdep.h>
45 #include <asm/abs_addr.h>
46 #include <asm/tlbflush.h>
50 #include <asm/cacheflush.h>
51 #include <asm/cputable.h>
52 #include <asm/abs_addr.h>
53 #include <asm/sections.h>
56 #define DBG(fmt...) udbg_printf(fmt)
62 #define DBG_LOW(fmt...) udbg_printf(fmt)
64 #define DBG_LOW(fmt...)
71 * Note: pte --> Linux PTE
72 * HPTE --> PowerPC Hashed Page Table Entry
75 * htab_initialize is called with the MMU off (of course), but
76 * the kernel has been copied down to zero so it can directly
77 * reference global data. At this point it is very difficult
78 * to print debug info.
83 extern unsigned long dart_tablebase;
84 #endif /* CONFIG_U3_DART */
86 static unsigned long _SDR1;
87 struct mmu_psize_def mmu_psize_defs[MMU_PAGE_COUNT];
90 unsigned long htab_size_bytes;
91 unsigned long htab_hash_mask;
92 int mmu_linear_psize = MMU_PAGE_4K;
93 int mmu_virtual_psize = MMU_PAGE_4K;
94 int mmu_vmalloc_psize = MMU_PAGE_4K;
95 int mmu_io_psize = MMU_PAGE_4K;
96 #ifdef CONFIG_HUGETLB_PAGE
97 int mmu_huge_psize = MMU_PAGE_16M;
98 unsigned int HPAGE_SHIFT;
100 #ifdef CONFIG_PPC_64K_PAGES
101 int mmu_ci_restrictions;
104 /* There are definitions of page sizes arrays to be used when none
105 * is provided by the firmware.
108 /* Pre-POWER4 CPUs (4k pages only)
110 struct mmu_psize_def mmu_psize_defaults_old[] = {
120 /* POWER4, GPUL, POWER5
122 * Support for 16Mb large pages
124 struct mmu_psize_def mmu_psize_defaults_gp[] = {
142 int htab_bolt_mapping(unsigned long vstart, unsigned long vend,
143 unsigned long pstart, unsigned long mode, int psize)
145 unsigned long vaddr, paddr;
146 unsigned int step, shift;
147 unsigned long tmp_mode;
150 shift = mmu_psize_defs[psize].shift;
153 for (vaddr = vstart, paddr = pstart; vaddr < vend;
154 vaddr += step, paddr += step) {
155 unsigned long vpn, hash, hpteg;
156 unsigned long vsid = get_kernel_vsid(vaddr);
157 unsigned long va = (vsid << 28) | (vaddr & 0x0fffffff);
162 /* Make non-kernel text non-executable */
163 if (!in_kernel_text(vaddr))
164 tmp_mode = mode | HPTE_R_N;
166 hash = hpt_hash(va, shift);
167 hpteg = ((hash & htab_hash_mask) * HPTES_PER_GROUP);
169 DBG("htab_bolt_mapping: calling %p\n", ppc_md.hpte_insert);
171 BUG_ON(!ppc_md.hpte_insert);
172 ret = ppc_md.hpte_insert(hpteg, va, paddr,
173 tmp_mode, HPTE_V_BOLTED, psize);
178 return ret < 0 ? ret : 0;
181 static int __init htab_dt_scan_page_sizes(unsigned long node,
182 const char *uname, int depth,
185 char *type = of_get_flat_dt_prop(node, "device_type", NULL);
187 unsigned long size = 0;
189 /* We are scanning "cpu" nodes only */
190 if (type == NULL || strcmp(type, "cpu") != 0)
193 prop = (u32 *)of_get_flat_dt_prop(node,
194 "ibm,segment-page-sizes", &size);
196 DBG("Page sizes from device-tree:\n");
198 cur_cpu_spec->cpu_features &= ~(CPU_FTR_16M_PAGE);
200 unsigned int shift = prop[0];
201 unsigned int slbenc = prop[1];
202 unsigned int lpnum = prop[2];
203 unsigned int lpenc = 0;
204 struct mmu_psize_def *def;
207 size -= 3; prop += 3;
208 while(size > 0 && lpnum) {
209 if (prop[0] == shift)
211 prop += 2; size -= 2;
226 cur_cpu_spec->cpu_features |= CPU_FTR_16M_PAGE;
234 def = &mmu_psize_defs[idx];
239 def->avpnm = (1 << (shift - 23)) - 1;
242 /* We don't know for sure what's up with tlbiel, so
243 * for now we only set it for 4K and 64K pages
245 if (idx == MMU_PAGE_4K || idx == MMU_PAGE_64K)
250 DBG(" %d: shift=%02x, sllp=%04x, avpnm=%08x, "
251 "tlbiel=%d, penc=%d\n",
252 idx, shift, def->sllp, def->avpnm, def->tlbiel,
261 static void __init htab_init_page_sizes(void)
265 /* Default to 4K pages only */
266 memcpy(mmu_psize_defs, mmu_psize_defaults_old,
267 sizeof(mmu_psize_defaults_old));
270 * Try to find the available page sizes in the device-tree
272 rc = of_scan_flat_dt(htab_dt_scan_page_sizes, NULL);
273 if (rc != 0) /* Found */
277 * Not in the device-tree, let's fallback on known size
278 * list for 16M capable GP & GR
280 if (cpu_has_feature(CPU_FTR_16M_PAGE))
281 memcpy(mmu_psize_defs, mmu_psize_defaults_gp,
282 sizeof(mmu_psize_defaults_gp));
285 * Pick a size for the linear mapping. Currently, we only support
286 * 16M, 1M and 4K which is the default
288 if (mmu_psize_defs[MMU_PAGE_16M].shift)
289 mmu_linear_psize = MMU_PAGE_16M;
290 else if (mmu_psize_defs[MMU_PAGE_1M].shift)
291 mmu_linear_psize = MMU_PAGE_1M;
293 #ifdef CONFIG_PPC_64K_PAGES
295 * Pick a size for the ordinary pages. Default is 4K, we support
296 * 64K for user mappings and vmalloc if supported by the processor.
297 * We only use 64k for ioremap if the processor
298 * (and firmware) support cache-inhibited large pages.
299 * If not, we use 4k and set mmu_ci_restrictions so that
300 * hash_page knows to switch processes that use cache-inhibited
301 * mappings to 4k pages.
303 if (mmu_psize_defs[MMU_PAGE_64K].shift) {
304 mmu_virtual_psize = MMU_PAGE_64K;
305 mmu_vmalloc_psize = MMU_PAGE_64K;
306 if (cpu_has_feature(CPU_FTR_CI_LARGE_PAGE))
307 mmu_io_psize = MMU_PAGE_64K;
309 mmu_ci_restrictions = 1;
313 printk(KERN_DEBUG "Page orders: linear mapping = %d, "
314 "virtual = %d, io = %d\n",
315 mmu_psize_defs[mmu_linear_psize].shift,
316 mmu_psize_defs[mmu_virtual_psize].shift,
317 mmu_psize_defs[mmu_io_psize].shift);
319 #ifdef CONFIG_HUGETLB_PAGE
320 /* Init large page size. Currently, we pick 16M or 1M depending
321 * on what is available
323 if (mmu_psize_defs[MMU_PAGE_16M].shift)
324 mmu_huge_psize = MMU_PAGE_16M;
325 /* With 4k/4level pagetables, we can't (for now) cope with a
326 * huge page size < PMD_SIZE */
327 else if (mmu_psize_defs[MMU_PAGE_1M].shift)
328 mmu_huge_psize = MMU_PAGE_1M;
330 /* Calculate HPAGE_SHIFT and sanity check it */
331 if (mmu_psize_defs[mmu_huge_psize].shift > MIN_HUGEPTE_SHIFT &&
332 mmu_psize_defs[mmu_huge_psize].shift < SID_SHIFT)
333 HPAGE_SHIFT = mmu_psize_defs[mmu_huge_psize].shift;
335 HPAGE_SHIFT = 0; /* No huge pages dude ! */
336 #endif /* CONFIG_HUGETLB_PAGE */
339 static int __init htab_dt_scan_pftsize(unsigned long node,
340 const char *uname, int depth,
343 char *type = of_get_flat_dt_prop(node, "device_type", NULL);
346 /* We are scanning "cpu" nodes only */
347 if (type == NULL || strcmp(type, "cpu") != 0)
350 prop = (u32 *)of_get_flat_dt_prop(node, "ibm,pft-size", NULL);
352 /* pft_size[0] is the NUMA CEC cookie */
353 ppc64_pft_size = prop[1];
359 static unsigned long __init htab_get_table_size(void)
361 unsigned long mem_size, rnd_mem_size, pteg_count;
363 /* If hash size isn't already provided by the platform, we try to
364 * retrieve it from the device-tree. If it's not there neither, we
365 * calculate it now based on the total RAM size
367 if (ppc64_pft_size == 0)
368 of_scan_flat_dt(htab_dt_scan_pftsize, NULL);
370 return 1UL << ppc64_pft_size;
372 /* round mem_size up to next power of 2 */
373 mem_size = lmb_phys_mem_size();
374 rnd_mem_size = 1UL << __ilog2(mem_size);
375 if (rnd_mem_size < mem_size)
379 pteg_count = max(rnd_mem_size >> (12 + 1), 1UL << 11);
381 return pteg_count << 7;
384 #ifdef CONFIG_MEMORY_HOTPLUG
385 void create_section_mapping(unsigned long start, unsigned long end)
387 BUG_ON(htab_bolt_mapping(start, end, __pa(start),
388 _PAGE_ACCESSED | _PAGE_DIRTY | _PAGE_COHERENT | PP_RWXX,
391 #endif /* CONFIG_MEMORY_HOTPLUG */
393 static inline void make_bl(unsigned int *insn_addr, void *func)
395 unsigned long funcp = *((unsigned long *)func);
396 int offset = funcp - (unsigned long)insn_addr;
398 *insn_addr = (unsigned int)(0x48000001 | (offset & 0x03fffffc));
399 flush_icache_range((unsigned long)insn_addr, 4+
400 (unsigned long)insn_addr);
403 static void __init htab_finish_init(void)
405 extern unsigned int *htab_call_hpte_insert1;
406 extern unsigned int *htab_call_hpte_insert2;
407 extern unsigned int *htab_call_hpte_remove;
408 extern unsigned int *htab_call_hpte_updatepp;
410 #ifdef CONFIG_PPC_64K_PAGES
411 extern unsigned int *ht64_call_hpte_insert1;
412 extern unsigned int *ht64_call_hpte_insert2;
413 extern unsigned int *ht64_call_hpte_remove;
414 extern unsigned int *ht64_call_hpte_updatepp;
416 make_bl(ht64_call_hpte_insert1, ppc_md.hpte_insert);
417 make_bl(ht64_call_hpte_insert2, ppc_md.hpte_insert);
418 make_bl(ht64_call_hpte_remove, ppc_md.hpte_remove);
419 make_bl(ht64_call_hpte_updatepp, ppc_md.hpte_updatepp);
420 #endif /* CONFIG_PPC_64K_PAGES */
422 make_bl(htab_call_hpte_insert1, ppc_md.hpte_insert);
423 make_bl(htab_call_hpte_insert2, ppc_md.hpte_insert);
424 make_bl(htab_call_hpte_remove, ppc_md.hpte_remove);
425 make_bl(htab_call_hpte_updatepp, ppc_md.hpte_updatepp);
428 void __init htab_initialize(void)
431 unsigned long pteg_count;
432 unsigned long mode_rw;
433 unsigned long base = 0, size = 0;
436 extern unsigned long tce_alloc_start, tce_alloc_end;
438 DBG(" -> htab_initialize()\n");
440 /* Initialize page sizes */
441 htab_init_page_sizes();
444 * Calculate the required size of the htab. We want the number of
445 * PTEGs to equal one half the number of real pages.
447 htab_size_bytes = htab_get_table_size();
448 pteg_count = htab_size_bytes >> 7;
450 htab_hash_mask = pteg_count - 1;
452 if (firmware_has_feature(FW_FEATURE_LPAR)) {
453 /* Using a hypervisor which owns the htab */
457 /* Find storage for the HPT. Must be contiguous in
458 * the absolute address space.
460 table = lmb_alloc(htab_size_bytes, htab_size_bytes);
462 DBG("Hash table allocated at %lx, size: %lx\n", table,
465 htab_address = abs_to_virt(table);
467 /* htab absolute addr + encoded htabsize */
468 _SDR1 = table + __ilog2(pteg_count) - 11;
470 /* Initialize the HPT with no entries */
471 memset((void *)table, 0, htab_size_bytes);
474 mtspr(SPRN_SDR1, _SDR1);
477 mode_rw = _PAGE_ACCESSED | _PAGE_DIRTY | _PAGE_COHERENT | PP_RWXX;
479 /* On U3 based machines, we need to reserve the DART area and
480 * _NOT_ map it to avoid cache paradoxes as it's remapped non
484 /* create bolted the linear mapping in the hash table */
485 for (i=0; i < lmb.memory.cnt; i++) {
486 base = (unsigned long)__va(lmb.memory.region[i].base);
487 size = lmb.memory.region[i].size;
489 DBG("creating mapping for region: %lx : %lx\n", base, size);
491 #ifdef CONFIG_U3_DART
492 /* Do not map the DART space. Fortunately, it will be aligned
493 * in such a way that it will not cross two lmb regions and
494 * will fit within a single 16Mb page.
495 * The DART space is assumed to be a full 16Mb region even if
496 * we only use 2Mb of that space. We will use more of it later
497 * for AGP GART. We have to use a full 16Mb large page.
499 DBG("DART base: %lx\n", dart_tablebase);
501 if (dart_tablebase != 0 && dart_tablebase >= base
502 && dart_tablebase < (base + size)) {
503 unsigned long dart_table_end = dart_tablebase + 16 * MB;
504 if (base != dart_tablebase)
505 BUG_ON(htab_bolt_mapping(base, dart_tablebase,
508 if ((base + size) > dart_table_end)
509 BUG_ON(htab_bolt_mapping(dart_tablebase+16*MB,
511 __pa(dart_table_end),
516 #endif /* CONFIG_U3_DART */
517 BUG_ON(htab_bolt_mapping(base, base + size, __pa(base),
518 mode_rw, mmu_linear_psize));
522 * If we have a memory_limit and we've allocated TCEs then we need to
523 * explicitly map the TCE area at the top of RAM. We also cope with the
524 * case that the TCEs start below memory_limit.
525 * tce_alloc_start/end are 16MB aligned so the mapping should work
526 * for either 4K or 16MB pages.
528 if (tce_alloc_start) {
529 tce_alloc_start = (unsigned long)__va(tce_alloc_start);
530 tce_alloc_end = (unsigned long)__va(tce_alloc_end);
532 if (base + size >= tce_alloc_start)
533 tce_alloc_start = base + size + 1;
535 BUG_ON(htab_bolt_mapping(tce_alloc_start, tce_alloc_end,
536 __pa(tce_alloc_start), mode_rw,
542 DBG(" <- htab_initialize()\n");
547 void htab_initialize_secondary(void)
549 if (!firmware_has_feature(FW_FEATURE_LPAR))
550 mtspr(SPRN_SDR1, _SDR1);
554 * Called by asm hashtable.S for doing lazy icache flush
556 unsigned int hash_page_do_lazy_icache(unsigned int pp, pte_t pte, int trap)
560 if (!pfn_valid(pte_pfn(pte)))
563 page = pte_page(pte);
566 if (!test_bit(PG_arch_1, &page->flags) && !PageReserved(page)) {
568 __flush_dcache_icache(page_address(page));
569 set_bit(PG_arch_1, &page->flags);
578 * 1 - normal page fault
579 * -1 - critical hash insertion error
581 int hash_page(unsigned long ea, unsigned long access, unsigned long trap)
585 struct mm_struct *mm;
588 int rc, user_region = 0, local = 0;
591 DBG_LOW("hash_page(ea=%016lx, access=%lx, trap=%lx\n",
594 if ((ea & ~REGION_MASK) >= PGTABLE_RANGE) {
595 DBG_LOW(" out of pgtable range !\n");
599 /* Get region & vsid */
600 switch (REGION_ID(ea)) {
605 DBG_LOW(" user region with no mm !\n");
608 vsid = get_vsid(mm->context.id, ea);
609 psize = mm->context.user_psize;
611 case VMALLOC_REGION_ID:
613 vsid = get_kernel_vsid(ea);
614 if (ea < VMALLOC_END)
615 psize = mmu_vmalloc_psize;
617 psize = mmu_io_psize;
621 * Send the problem up to do_page_fault
625 DBG_LOW(" mm=%p, mm->pgdir=%p, vsid=%016lx\n", mm, mm->pgd, vsid);
632 /* Check CPU locality */
633 tmp = cpumask_of_cpu(smp_processor_id());
634 if (user_region && cpus_equal(mm->cpu_vm_mask, tmp))
637 /* Handle hugepage regions */
638 if (unlikely(in_hugepage_area(mm->context, ea))) {
639 DBG_LOW(" -> huge page !\n");
640 return hash_huge_page(mm, access, ea, vsid, local, trap);
643 /* Get PTE and page size from page tables */
644 ptep = find_linux_pte(pgdir, ea);
645 if (ptep == NULL || !pte_present(*ptep)) {
646 DBG_LOW(" no PTE !\n");
650 #ifndef CONFIG_PPC_64K_PAGES
651 DBG_LOW(" i-pte: %016lx\n", pte_val(*ptep));
653 DBG_LOW(" i-pte: %016lx %016lx\n", pte_val(*ptep),
654 pte_val(*(ptep + PTRS_PER_PTE)));
656 /* Pre-check access permissions (will be re-checked atomically
657 * in __hash_page_XX but this pre-check is a fast path
659 if (access & ~pte_val(*ptep)) {
660 DBG_LOW(" no access !\n");
664 /* Do actual hashing */
665 #ifndef CONFIG_PPC_64K_PAGES
666 rc = __hash_page_4K(ea, access, vsid, ptep, trap, local);
668 if (mmu_ci_restrictions) {
669 /* If this PTE is non-cacheable, switch to 4k */
670 if (psize == MMU_PAGE_64K &&
671 (pte_val(*ptep) & _PAGE_NO_CACHE)) {
674 mm->context.user_psize = MMU_PAGE_4K;
675 mm->context.sllp = SLB_VSID_USER |
676 mmu_psize_defs[MMU_PAGE_4K].sllp;
677 } else if (ea < VMALLOC_END) {
679 * some driver did a non-cacheable mapping
680 * in vmalloc space, so switch vmalloc
683 printk(KERN_ALERT "Reducing vmalloc segment "
684 "to 4kB pages because of "
685 "non-cacheable mapping\n");
686 psize = mmu_vmalloc_psize = MMU_PAGE_4K;
690 if (psize != get_paca()->context.user_psize) {
691 get_paca()->context = mm->context;
692 slb_flush_and_rebolt();
694 } else if (get_paca()->vmalloc_sllp !=
695 mmu_psize_defs[mmu_vmalloc_psize].sllp) {
696 get_paca()->vmalloc_sllp =
697 mmu_psize_defs[mmu_vmalloc_psize].sllp;
698 slb_flush_and_rebolt();
701 if (psize == MMU_PAGE_64K)
702 rc = __hash_page_64K(ea, access, vsid, ptep, trap, local);
704 rc = __hash_page_4K(ea, access, vsid, ptep, trap, local);
705 #endif /* CONFIG_PPC_64K_PAGES */
707 #ifndef CONFIG_PPC_64K_PAGES
708 DBG_LOW(" o-pte: %016lx\n", pte_val(*ptep));
710 DBG_LOW(" o-pte: %016lx %016lx\n", pte_val(*ptep),
711 pte_val(*(ptep + PTRS_PER_PTE)));
713 DBG_LOW(" -> rc=%d\n", rc);
716 EXPORT_SYMBOL_GPL(hash_page);
718 void hash_preload(struct mm_struct *mm, unsigned long ea,
719 unsigned long access, unsigned long trap)
728 /* We don't want huge pages prefaulted for now
730 if (unlikely(in_hugepage_area(mm->context, ea)))
733 DBG_LOW("hash_preload(mm=%p, mm->pgdir=%p, ea=%016lx, access=%lx,"
734 " trap=%lx\n", mm, mm->pgd, ea, access, trap);
736 /* Get PTE, VSID, access mask */
740 ptep = find_linux_pte(pgdir, ea);
743 vsid = get_vsid(mm->context.id, ea);
746 local_irq_save(flags);
747 mask = cpumask_of_cpu(smp_processor_id());
748 if (cpus_equal(mm->cpu_vm_mask, mask))
750 #ifndef CONFIG_PPC_64K_PAGES
751 __hash_page_4K(ea, access, vsid, ptep, trap, local);
753 if (mmu_ci_restrictions) {
754 /* If this PTE is non-cacheable, switch to 4k */
755 if (mm->context.user_psize == MMU_PAGE_64K &&
756 (pte_val(*ptep) & _PAGE_NO_CACHE)) {
757 mm->context.user_psize = MMU_PAGE_4K;
758 mm->context.sllp = SLB_VSID_USER |
759 mmu_psize_defs[MMU_PAGE_4K].sllp;
760 get_paca()->context = mm->context;
761 slb_flush_and_rebolt();
764 if (mm->context.user_psize == MMU_PAGE_64K)
765 __hash_page_64K(ea, access, vsid, ptep, trap, local);
767 __hash_page_4K(ea, access, vsid, ptep, trap, local);
768 #endif /* CONFIG_PPC_64K_PAGES */
769 local_irq_restore(flags);
772 void flush_hash_page(unsigned long va, real_pte_t pte, int psize, int local)
774 unsigned long hash, index, shift, hidx, slot;
776 DBG_LOW("flush_hash_page(va=%016x)\n", va);
777 pte_iterate_hashed_subpages(pte, psize, va, index, shift) {
778 hash = hpt_hash(va, shift);
779 hidx = __rpte_to_hidx(pte, index);
780 if (hidx & _PTEIDX_SECONDARY)
782 slot = (hash & htab_hash_mask) * HPTES_PER_GROUP;
783 slot += hidx & _PTEIDX_GROUP_IX;
784 DBG_LOW(" sub %d: hash=%x, hidx=%x\n", index, slot, hidx);
785 ppc_md.hpte_invalidate(slot, va, psize, local);
786 } pte_iterate_hashed_end();
789 void flush_hash_range(unsigned long number, int local)
791 if (ppc_md.flush_hash_range)
792 ppc_md.flush_hash_range(number, local);
795 struct ppc64_tlb_batch *batch =
796 &__get_cpu_var(ppc64_tlb_batch);
798 for (i = 0; i < number; i++)
799 flush_hash_page(batch->vaddr[i], batch->pte[i],
800 batch->psize, local);
805 * low_hash_fault is called when we the low level hash code failed
806 * to instert a PTE due to an hypervisor error
808 void low_hash_fault(struct pt_regs *regs, unsigned long address)
810 if (user_mode(regs)) {
813 info.si_signo = SIGBUS;
815 info.si_code = BUS_ADRERR;
816 info.si_addr = (void __user *)address;
817 force_sig_info(SIGBUS, &info, current);
820 bad_page_fault(regs, address, SIGBUS);