From 1da177e4c3f41524e886b7f1b8a0c1fc7321cac2 Mon Sep 17 00:00:00 2001 From: Linus Torvalds Date: Sat, 16 Apr 2005 15:20:36 -0700 Subject: Linux-2.6.12-rc2 Initial git repository build. I'm not bothering with the full history, even though we have it. We can create a separate "historical" git archive of that later if we want to, and in the meantime it's about 3.2GB when imported into git - space that would just make the early git days unnecessarily complicated, when we don't have a lot of good infrastructure for it. Let it rip! --- arch/sh64/mm/Makefile | 44 ++ arch/sh64/mm/cache.c | 1041 ++++++++++++++++++++++++++++++++++++++++++++ arch/sh64/mm/extable.c | 81 ++++ arch/sh64/mm/fault.c | 601 +++++++++++++++++++++++++ arch/sh64/mm/hugetlbpage.c | 264 +++++++++++ arch/sh64/mm/init.c | 196 +++++++++ arch/sh64/mm/ioremap.c | 469 ++++++++++++++++++++ arch/sh64/mm/tlb.c | 166 +++++++ arch/sh64/mm/tlbmiss.c | 280 ++++++++++++ 9 files changed, 3142 insertions(+) create mode 100644 arch/sh64/mm/Makefile create mode 100644 arch/sh64/mm/cache.c create mode 100644 arch/sh64/mm/extable.c create mode 100644 arch/sh64/mm/fault.c create mode 100644 arch/sh64/mm/hugetlbpage.c create mode 100644 arch/sh64/mm/init.c create mode 100644 arch/sh64/mm/ioremap.c create mode 100644 arch/sh64/mm/tlb.c create mode 100644 arch/sh64/mm/tlbmiss.c (limited to 'arch/sh64/mm') diff --git a/arch/sh64/mm/Makefile b/arch/sh64/mm/Makefile new file mode 100644 index 00000000000..ff19378ac90 --- /dev/null +++ b/arch/sh64/mm/Makefile @@ -0,0 +1,44 @@ +# +# This file is subject to the terms and conditions of the GNU General Public +# License. See the file "COPYING" in the main directory of this archive +# for more details. +# +# Copyright (C) 2000, 2001 Paolo Alberelli +# Copyright (C) 2003, 2004 Paul Mundt +# +# Makefile for the sh64-specific parts of the Linux memory manager. +# +# Note! Dependencies are done automagically by 'make dep', which also +# removes any old dependencies. DON'T put your own dependencies here +# unless it's something special (ie not a .c file). +# + +obj-y := init.o fault.o ioremap.o extable.o cache.o tlbmiss.o tlb.o + +obj-$(CONFIG_HUGETLB_PAGE) += hugetlbpage.o + +# Special flags for tlbmiss.o. This puts restrictions on the number of +# caller-save registers that the compiler can target when building this file. +# This is required because the code is called from a context in entry.S where +# very few registers have been saved in the exception handler (for speed +# reasons). +# The caller save registers that have been saved and which can be used are +# r2,r3,r4,r5 : argument passing +# r15, r18 : SP and LINK +# tr0-4 : allow all caller-save TR's. The compiler seems to be able to make +# use of them, so it's probably beneficial to performance to save them +# and have them available for it. +# +# The resources not listed below are callee save, i.e. the compiler is free to +# use any of them and will spill them to the stack itself. + +CFLAGS_tlbmiss.o += -ffixed-r7 \ + -ffixed-r8 -ffixed-r9 -ffixed-r10 -ffixed-r11 -ffixed-r12 \ + -ffixed-r13 -ffixed-r14 -ffixed-r16 -ffixed-r17 -ffixed-r19 \ + -ffixed-r20 -ffixed-r21 -ffixed-r22 -ffixed-r23 \ + -ffixed-r24 -ffixed-r25 -ffixed-r26 -ffixed-r27 \ + -ffixed-r36 -ffixed-r37 -ffixed-r38 -ffixed-r39 -ffixed-r40 \ + -ffixed-r41 -ffixed-r42 -ffixed-r43 \ + -ffixed-r60 -ffixed-r61 -ffixed-r62 \ + -fomit-frame-pointer + diff --git a/arch/sh64/mm/cache.c b/arch/sh64/mm/cache.c new file mode 100644 index 00000000000..3b87e25ea77 --- /dev/null +++ b/arch/sh64/mm/cache.c @@ -0,0 +1,1041 @@ +/* + * This file is subject to the terms and conditions of the GNU General Public + * License. See the file "COPYING" in the main directory of this archive + * for more details. + * + * arch/sh64/mm/cache.c + * + * Original version Copyright (C) 2000, 2001 Paolo Alberelli + * Second version Copyright (C) benedict.gaster@superh.com 2002 + * Third version Copyright Richard.Curnow@superh.com 2003 + * Hacks to third version Copyright (C) 2003 Paul Mundt + */ + +/****************************************************************************/ + +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include /* for flush_itlb_range */ + +#include + +/* This function is in entry.S */ +extern unsigned long switch_and_save_asid(unsigned long new_asid); + +/* Wired TLB entry for the D-cache */ +static unsigned long long dtlb_cache_slot; + +/** + * sh64_cache_init() + * + * This is pretty much just a straightforward clone of the SH + * detect_cpu_and_cache_system(). + * + * This function is responsible for setting up all of the cache + * info dynamically as well as taking care of CPU probing and + * setting up the relevant subtype data. + * + * FIXME: For the time being, we only really support the SH5-101 + * out of the box, and don't support dynamic probing for things + * like the SH5-103 or even cut2 of the SH5-101. Implement this + * later! + */ +int __init sh64_cache_init(void) +{ + /* + * First, setup some sane values for the I-cache. + */ + cpu_data->icache.ways = 4; + cpu_data->icache.sets = 256; + cpu_data->icache.linesz = L1_CACHE_BYTES; + + /* + * FIXME: This can probably be cleaned up a bit as well.. for example, + * do we really need the way shift _and_ the way_step_shift ?? Judging + * by the existing code, I would guess no.. is there any valid reason + * why we need to be tracking this around? + */ + cpu_data->icache.way_shift = 13; + cpu_data->icache.entry_shift = 5; + cpu_data->icache.set_shift = 4; + cpu_data->icache.way_step_shift = 16; + cpu_data->icache.asid_shift = 2; + + /* + * way offset = cache size / associativity, so just don't factor in + * associativity in the first place.. + */ + cpu_data->icache.way_ofs = cpu_data->icache.sets * + cpu_data->icache.linesz; + + cpu_data->icache.asid_mask = 0x3fc; + cpu_data->icache.idx_mask = 0x1fe0; + cpu_data->icache.epn_mask = 0xffffe000; + cpu_data->icache.flags = 0; + + /* + * Next, setup some sane values for the D-cache. + * + * On the SH5, these are pretty consistent with the I-cache settings, + * so we just copy over the existing definitions.. these can be fixed + * up later, especially if we add runtime CPU probing. + * + * Though in the meantime it saves us from having to duplicate all of + * the above definitions.. + */ + cpu_data->dcache = cpu_data->icache; + + /* + * Setup any cache-related flags here + */ +#if defined(CONFIG_DCACHE_WRITE_THROUGH) + set_bit(SH_CACHE_MODE_WT, &(cpu_data->dcache.flags)); +#elif defined(CONFIG_DCACHE_WRITE_BACK) + set_bit(SH_CACHE_MODE_WB, &(cpu_data->dcache.flags)); +#endif + + /* + * We also need to reserve a slot for the D-cache in the DTLB, so we + * do this now .. + */ + dtlb_cache_slot = sh64_get_wired_dtlb_entry(); + + return 0; +} + +#ifdef CONFIG_DCACHE_DISABLED +#define sh64_dcache_purge_all() do { } while (0) +#define sh64_dcache_purge_coloured_phy_page(paddr, eaddr) do { } while (0) +#define sh64_dcache_purge_user_range(mm, start, end) do { } while (0) +#define sh64_dcache_purge_phy_page(paddr) do { } while (0) +#define sh64_dcache_purge_virt_page(mm, eaddr) do { } while (0) +#define sh64_dcache_purge_kernel_range(start, end) do { } while (0) +#define sh64_dcache_wback_current_user_range(start, end) do { } while (0) +#endif + +/*##########################################################################*/ + +/* From here onwards, a rewrite of the implementation, + by Richard.Curnow@superh.com. + + The major changes in this compared to the old version are; + 1. use more selective purging through OCBP instead of using ALLOCO to purge + by natural replacement. This avoids purging out unrelated cache lines + that happen to be in the same set. + 2. exploit the APIs copy_user_page and clear_user_page better + 3. be more selective about I-cache purging, in particular use invalidate_all + more sparingly. + + */ + +/*########################################################################## + SUPPORT FUNCTIONS + ##########################################################################*/ + +/****************************************************************************/ +/* The following group of functions deal with mapping and unmapping a temporary + page into the DTLB slot that have been set aside for our exclusive use. */ +/* In order to accomplish this, we use the generic interface for adding and + removing a wired slot entry as defined in arch/sh64/mm/tlb.c */ +/****************************************************************************/ + +static unsigned long slot_own_flags; + +static inline void sh64_setup_dtlb_cache_slot(unsigned long eaddr, unsigned long asid, unsigned long paddr) +{ + local_irq_save(slot_own_flags); + sh64_setup_tlb_slot(dtlb_cache_slot, eaddr, asid, paddr); +} + +static inline void sh64_teardown_dtlb_cache_slot(void) +{ + sh64_teardown_tlb_slot(dtlb_cache_slot); + local_irq_restore(slot_own_flags); +} + +/****************************************************************************/ + +#ifndef CONFIG_ICACHE_DISABLED + +static void __inline__ sh64_icache_inv_all(void) +{ + unsigned long long addr, flag, data; + unsigned int flags; + + addr=ICCR0; + flag=ICCR0_ICI; + data=0; + + /* Make this a critical section for safety (probably not strictly necessary.) */ + local_irq_save(flags); + + /* Without %1 it gets unexplicably wrong */ + asm volatile("getcfg %3, 0, %0\n\t" + "or %0, %2, %0\n\t" + "putcfg %3, 0, %0\n\t" + "synci" + : "=&r" (data) + : "0" (data), "r" (flag), "r" (addr)); + + local_irq_restore(flags); +} + +static void sh64_icache_inv_kernel_range(unsigned long start, unsigned long end) +{ + /* Invalidate range of addresses [start,end] from the I-cache, where + * the addresses lie in the kernel superpage. */ + + unsigned long long ullend, addr, aligned_start; +#if (NEFF == 32) + aligned_start = (unsigned long long)(signed long long)(signed long) start; +#else +#error "NEFF != 32" +#endif + aligned_start &= L1_CACHE_ALIGN_MASK; + addr = aligned_start; +#if (NEFF == 32) + ullend = (unsigned long long) (signed long long) (signed long) end; +#else +#error "NEFF != 32" +#endif + while (addr <= ullend) { + asm __volatile__ ("icbi %0, 0" : : "r" (addr)); + addr += L1_CACHE_BYTES; + } +} + +static void sh64_icache_inv_user_page(struct vm_area_struct *vma, unsigned long eaddr) +{ + /* If we get called, we know that vma->vm_flags contains VM_EXEC. + Also, eaddr is page-aligned. */ + + unsigned long long addr, end_addr; + unsigned long flags = 0; + unsigned long running_asid, vma_asid; + addr = eaddr; + end_addr = addr + PAGE_SIZE; + + /* Check whether we can use the current ASID for the I-cache + invalidation. For example, if we're called via + access_process_vm->flush_cache_page->here, (e.g. when reading from + /proc), 'running_asid' will be that of the reader, not of the + victim. + + Also, note the risk that we might get pre-empted between the ASID + compare and blocking IRQs, and before we regain control, the + pid->ASID mapping changes. However, the whole cache will get + invalidated when the mapping is renewed, so the worst that can + happen is that the loop below ends up invalidating somebody else's + cache entries. + */ + + running_asid = get_asid(); + vma_asid = (vma->vm_mm->context & MMU_CONTEXT_ASID_MASK); + if (running_asid != vma_asid) { + local_irq_save(flags); + switch_and_save_asid(vma_asid); + } + while (addr < end_addr) { + /* Worth unrolling a little */ + asm __volatile__("icbi %0, 0" : : "r" (addr)); + asm __volatile__("icbi %0, 32" : : "r" (addr)); + asm __volatile__("icbi %0, 64" : : "r" (addr)); + asm __volatile__("icbi %0, 96" : : "r" (addr)); + addr += 128; + } + if (running_asid != vma_asid) { + switch_and_save_asid(running_asid); + local_irq_restore(flags); + } +} + +/****************************************************************************/ + +static void sh64_icache_inv_user_page_range(struct mm_struct *mm, + unsigned long start, unsigned long end) +{ + /* Used for invalidating big chunks of I-cache, i.e. assume the range + is whole pages. If 'start' or 'end' is not page aligned, the code + is conservative and invalidates to the ends of the enclosing pages. + This is functionally OK, just a performance loss. */ + + /* See the comments below in sh64_dcache_purge_user_range() regarding + the choice of algorithm. However, for the I-cache option (2) isn't + available because there are no physical tags so aliases can't be + resolved. The icbi instruction has to be used through the user + mapping. Because icbi is cheaper than ocbp on a cache hit, it + would be cheaper to use the selective code for a large range than is + possible with the D-cache. Just assume 64 for now as a working + figure. + */ + + int n_pages; + + if (!mm) return; + + n_pages = ((end - start) >> PAGE_SHIFT); + if (n_pages >= 64) { + sh64_icache_inv_all(); + } else { + unsigned long aligned_start; + unsigned long eaddr; + unsigned long after_last_page_start; + unsigned long mm_asid, current_asid; + unsigned long long flags = 0ULL; + + mm_asid = mm->context & MMU_CONTEXT_ASID_MASK; + current_asid = get_asid(); + + if (mm_asid != current_asid) { + /* Switch ASID and run the invalidate loop under cli */ + local_irq_save(flags); + switch_and_save_asid(mm_asid); + } + + aligned_start = start & PAGE_MASK; + after_last_page_start = PAGE_SIZE + ((end - 1) & PAGE_MASK); + + while (aligned_start < after_last_page_start) { + struct vm_area_struct *vma; + unsigned long vma_end; + vma = find_vma(mm, aligned_start); + if (!vma || (aligned_start <= vma->vm_end)) { + /* Avoid getting stuck in an error condition */ + aligned_start += PAGE_SIZE; + continue; + } + vma_end = vma->vm_end; + if (vma->vm_flags & VM_EXEC) { + /* Executable */ + eaddr = aligned_start; + while (eaddr < vma_end) { + sh64_icache_inv_user_page(vma, eaddr); + eaddr += PAGE_SIZE; + } + } + aligned_start = vma->vm_end; /* Skip to start of next region */ + } + if (mm_asid != current_asid) { + switch_and_save_asid(current_asid); + local_irq_restore(flags); + } + } +} + +static void sh64_icache_inv_user_small_range(struct mm_struct *mm, + unsigned long start, int len) +{ + + /* Invalidate a small range of user context I-cache, not necessarily + page (or even cache-line) aligned. */ + + unsigned long long eaddr = start; + unsigned long long eaddr_end = start + len; + unsigned long current_asid, mm_asid; + unsigned long long flags; + unsigned long long epage_start; + + /* Since this is used inside ptrace, the ASID in the mm context + typically won't match current_asid. We'll have to switch ASID to do + this. For safety, and given that the range will be small, do all + this under cli. + + Note, there is a hazard that the ASID in mm->context is no longer + actually associated with mm, i.e. if the mm->context has started a + new cycle since mm was last active. However, this is just a + performance issue: all that happens is that we invalidate lines + belonging to another mm, so the owning process has to refill them + when that mm goes live again. mm itself can't have any cache + entries because there will have been a flush_cache_all when the new + mm->context cycle started. */ + + /* Align to start of cache line. Otherwise, suppose len==8 and start + was at 32N+28 : the last 4 bytes wouldn't get invalidated. */ + eaddr = start & L1_CACHE_ALIGN_MASK; + eaddr_end = start + len; + + local_irq_save(flags); + mm_asid = mm->context & MMU_CONTEXT_ASID_MASK; + current_asid = switch_and_save_asid(mm_asid); + + epage_start = eaddr & PAGE_MASK; + + while (eaddr < eaddr_end) + { + asm __volatile__("icbi %0, 0" : : "r" (eaddr)); + eaddr += L1_CACHE_BYTES; + } + switch_and_save_asid(current_asid); + local_irq_restore(flags); +} + +static void sh64_icache_inv_current_user_range(unsigned long start, unsigned long end) +{ + /* The icbi instruction never raises ITLBMISS. i.e. if there's not a + cache hit on the virtual tag the instruction ends there, without a + TLB lookup. */ + + unsigned long long aligned_start; + unsigned long long ull_end; + unsigned long long addr; + + ull_end = end; + + /* Just invalidate over the range using the natural addresses. TLB + miss handling will be OK (TBC). Since it's for the current process, + either we're already in the right ASID context, or the ASIDs have + been recycled since we were last active in which case we might just + invalidate another processes I-cache entries : no worries, just a + performance drop for him. */ + aligned_start = start & L1_CACHE_ALIGN_MASK; + addr = aligned_start; + while (addr < ull_end) { + asm __volatile__ ("icbi %0, 0" : : "r" (addr)); + asm __volatile__ ("nop"); + asm __volatile__ ("nop"); + addr += L1_CACHE_BYTES; + } +} + +#endif /* !CONFIG_ICACHE_DISABLED */ + +/****************************************************************************/ + +#ifndef CONFIG_DCACHE_DISABLED + +/* Buffer used as the target of alloco instructions to purge data from cache + sets by natural eviction. -- RPC */ +#define DUMMY_ALLOCO_AREA_SIZE L1_CACHE_SIZE_BYTES + (1024 * 4) +static unsigned char dummy_alloco_area[DUMMY_ALLOCO_AREA_SIZE] __cacheline_aligned = { 0, }; + +/****************************************************************************/ + +static void __inline__ sh64_dcache_purge_sets(int sets_to_purge_base, int n_sets) +{ + /* Purge all ways in a particular block of sets, specified by the base + set number and number of sets. Can handle wrap-around, if that's + needed. */ + + int dummy_buffer_base_set; + unsigned long long eaddr, eaddr0, eaddr1; + int j; + int set_offset; + + dummy_buffer_base_set = ((int)&dummy_alloco_area & cpu_data->dcache.idx_mask) >> cpu_data->dcache.entry_shift; + set_offset = sets_to_purge_base - dummy_buffer_base_set; + + for (j=0; jdcache.sets - 1); + eaddr0 = (unsigned long long)dummy_alloco_area + (set_offset << cpu_data->dcache.entry_shift); + + /* Do one alloco which hits the required set per cache way. For + write-back mode, this will purge the #ways resident lines. There's + little point unrolling this loop because the allocos stall more if + they're too close together. */ + eaddr1 = eaddr0 + cpu_data->dcache.way_ofs * cpu_data->dcache.ways; + for (eaddr=eaddr0; eaddrdcache.way_ofs) { + asm __volatile__ ("alloco %0, 0" : : "r" (eaddr)); + asm __volatile__ ("synco"); /* TAKum03020 */ + } + + eaddr1 = eaddr0 + cpu_data->dcache.way_ofs * cpu_data->dcache.ways; + for (eaddr=eaddr0; eaddrdcache.way_ofs) { + /* Load from each address. Required because alloco is a NOP if + the cache is write-through. Write-through is a config option. */ + if (test_bit(SH_CACHE_MODE_WT, &(cpu_data->dcache.flags))) + *(volatile unsigned char *)(int)eaddr; + } + } + + /* Don't use OCBI to invalidate the lines. That costs cycles directly. + If the dummy block is just left resident, it will naturally get + evicted as required. */ + + return; +} + +/****************************************************************************/ + +static void sh64_dcache_purge_all(void) +{ + /* Purge the entire contents of the dcache. The most efficient way to + achieve this is to use alloco instructions on a region of unused + memory equal in size to the cache, thereby causing the current + contents to be discarded by natural eviction. The alternative, + namely reading every tag, setting up a mapping for the corresponding + page and doing an OCBP for the line, would be much more expensive. + */ + + sh64_dcache_purge_sets(0, cpu_data->dcache.sets); + + return; + +} + +/****************************************************************************/ + +static void sh64_dcache_purge_kernel_range(unsigned long start, unsigned long end) +{ + /* Purge the range of addresses [start,end] from the D-cache. The + addresses lie in the superpage mapping. There's no harm if we + overpurge at either end - just a small performance loss. */ + unsigned long long ullend, addr, aligned_start; +#if (NEFF == 32) + aligned_start = (unsigned long long)(signed long long)(signed long) start; +#else +#error "NEFF != 32" +#endif + aligned_start &= L1_CACHE_ALIGN_MASK; + addr = aligned_start; +#if (NEFF == 32) + ullend = (unsigned long long) (signed long long) (signed long) end; +#else +#error "NEFF != 32" +#endif + while (addr <= ullend) { + asm __volatile__ ("ocbp %0, 0" : : "r" (addr)); + addr += L1_CACHE_BYTES; + } + return; +} + +/* Assumes this address (+ (2**n_synbits) pages up from it) aren't used for + anything else in the kernel */ +#define MAGIC_PAGE0_START 0xffffffffec000000ULL + +static void sh64_dcache_purge_coloured_phy_page(unsigned long paddr, unsigned long eaddr) +{ + /* Purge the physical page 'paddr' from the cache. It's known that any + cache lines requiring attention have the same page colour as the the + address 'eaddr'. + + This relies on the fact that the D-cache matches on physical tags + when no virtual tag matches. So we create an alias for the original + page and purge through that. (Alternatively, we could have done + this by switching ASID to match the original mapping and purged + through that, but that involves ASID switching cost + probably a + TLBMISS + refill anyway.) + */ + + unsigned long long magic_page_start; + unsigned long long magic_eaddr, magic_eaddr_end; + + magic_page_start = MAGIC_PAGE0_START + (eaddr & CACHE_OC_SYN_MASK); + + /* As long as the kernel is not pre-emptible, this doesn't need to be + under cli/sti. */ + + sh64_setup_dtlb_cache_slot(magic_page_start, get_asid(), paddr); + + magic_eaddr = magic_page_start; + magic_eaddr_end = magic_eaddr + PAGE_SIZE; + while (magic_eaddr < magic_eaddr_end) { + /* Little point in unrolling this loop - the OCBPs are blocking + and won't go any quicker (i.e. the loop overhead is parallel + to part of the OCBP execution.) */ + asm __volatile__ ("ocbp %0, 0" : : "r" (magic_eaddr)); + magic_eaddr += L1_CACHE_BYTES; + } + + sh64_teardown_dtlb_cache_slot(); +} + +/****************************************************************************/ + +static void sh64_dcache_purge_phy_page(unsigned long paddr) +{ + /* Pure a page given its physical start address, by creating a + temporary 1 page mapping and purging across that. Even if we know + the virtual address (& vma or mm) of the page, the method here is + more elegant because it avoids issues of coping with page faults on + the purge instructions (i.e. no special-case code required in the + critical path in the TLB miss handling). */ + + unsigned long long eaddr_start, eaddr, eaddr_end; + int i; + + /* As long as the kernel is not pre-emptible, this doesn't need to be + under cli/sti. */ + + eaddr_start = MAGIC_PAGE0_START; + for (i=0; i < (1 << CACHE_OC_N_SYNBITS); i++) { + sh64_setup_dtlb_cache_slot(eaddr_start, get_asid(), paddr); + + eaddr = eaddr_start; + eaddr_end = eaddr + PAGE_SIZE; + while (eaddr < eaddr_end) { + asm __volatile__ ("ocbp %0, 0" : : "r" (eaddr)); + eaddr += L1_CACHE_BYTES; + } + + sh64_teardown_dtlb_cache_slot(); + eaddr_start += PAGE_SIZE; + } +} + +static void sh64_dcache_purge_user_page(struct mm_struct *mm, unsigned long eaddr) +{ + pgd_t *pgd; + pmd_t *pmd; + pte_t *pte; + pte_t entry; + unsigned long paddr; + + /* NOTE : all the callers of this have mm->page_table_lock held, so the + following page table traversal is safe even on SMP/pre-emptible. */ + + if (!mm) return; /* No way to find physical address of page */ + pgd = pgd_offset(mm, eaddr); + if (pgd_bad(*pgd)) return; + + pmd = pmd_offset(pgd, eaddr); + if (pmd_none(*pmd) || pmd_bad(*pmd)) return; + + pte = pte_offset_kernel(pmd, eaddr); + entry = *pte; + if (pte_none(entry) || !pte_present(entry)) return; + + paddr = pte_val(entry) & PAGE_MASK; + + sh64_dcache_purge_coloured_phy_page(paddr, eaddr); + +} +/****************************************************************************/ + +static void sh64_dcache_purge_user_range(struct mm_struct *mm, + unsigned long start, unsigned long end) +{ + /* There are at least 5 choices for the implementation of this, with + pros (+), cons(-), comments(*): + + 1. ocbp each line in the range through the original user's ASID + + no lines spuriously evicted + - tlbmiss handling (must either handle faults on demand => extra + special-case code in tlbmiss critical path), or map the page in + advance (=> flush_tlb_range in advance to avoid multiple hits) + - ASID switching + - expensive for large ranges + + 2. temporarily map each page in the range to a special effective + address and ocbp through the temporary mapping; relies on the + fact that SH-5 OCB* always do TLB lookup and match on ptags (they + never look at the etags) + + no spurious evictions + - expensive for large ranges + * surely cheaper than (1) + + 3. walk all the lines in the cache, check the tags, if a match + occurs create a page mapping to ocbp the line through + + no spurious evictions + - tag inspection overhead + - (especially for small ranges) + - potential cost of setting up/tearing down page mapping for + every line that matches the range + * cost partly independent of range size + + 4. walk all the lines in the cache, check the tags, if a match + occurs use 4 * alloco to purge the line (+3 other probably + innocent victims) by natural eviction + + no tlb mapping overheads + - spurious evictions + - tag inspection overhead + + 5. implement like flush_cache_all + + no tag inspection overhead + - spurious evictions + - bad for small ranges + + (1) can be ruled out as more expensive than (2). (2) appears best + for small ranges. The choice between (3), (4) and (5) for large + ranges and the range size for the large/small boundary need + benchmarking to determine. + + For now use approach (2) for small ranges and (5) for large ones. + + */ + + int n_pages; + + n_pages = ((end - start) >> PAGE_SHIFT); + if (n_pages >= 64) { +#if 1 + sh64_dcache_purge_all(); +#else + unsigned long long set, way; + unsigned long mm_asid = mm->context & MMU_CONTEXT_ASID_MASK; + for (set = 0; set < cpu_data->dcache.sets; set++) { + unsigned long long set_base_config_addr = CACHE_OC_ADDRESS_ARRAY + (set << cpu_data->dcache.set_shift); + for (way = 0; way < cpu_data->dcache.ways; way++) { + unsigned long long config_addr = set_base_config_addr + (way << cpu_data->dcache.way_step_shift); + unsigned long long tag0; + unsigned long line_valid; + + asm __volatile__("getcfg %1, 0, %0" : "=r" (tag0) : "r" (config_addr)); + line_valid = tag0 & SH_CACHE_VALID; + if (line_valid) { + unsigned long cache_asid; + unsigned long epn; + + cache_asid = (tag0 & cpu_data->dcache.asid_mask) >> cpu_data->dcache.asid_shift; + /* The next line needs some + explanation. The virtual tags + encode bits [31:13] of the virtual + address, bit [12] of the 'tag' being + implied by the cache set index. */ + epn = (tag0 & cpu_data->dcache.epn_mask) | ((set & 0x80) << cpu_data->dcache.entry_shift); + + if ((cache_asid == mm_asid) && (start <= epn) && (epn < end)) { + /* TODO : could optimise this + call by batching multiple + adjacent sets together. */ + sh64_dcache_purge_sets(set, 1); + break; /* Don't waste time inspecting other ways for this set */ + } + } + } + } +#endif + } else { + /* 'Small' range */ + unsigned long aligned_start; + unsigned long eaddr; + unsigned long last_page_start; + + aligned_start = start & PAGE_MASK; + /* 'end' is 1 byte beyond the end of the range */ + last_page_start = (end - 1) & PAGE_MASK; + + eaddr = aligned_start; + while (eaddr <= last_page_start) { + sh64_dcache_purge_user_page(mm, eaddr); + eaddr += PAGE_SIZE; + } + } + return; +} + +static void sh64_dcache_wback_current_user_range(unsigned long start, unsigned long end) +{ + unsigned long long aligned_start; + unsigned long long ull_end; + unsigned long long addr; + + ull_end = end; + + /* Just wback over the range using the natural addresses. TLB miss + handling will be OK (TBC) : the range has just been written to by + the signal frame setup code, so the PTEs must exist. + + Note, if we have CONFIG_PREEMPT and get preempted inside this loop, + it doesn't matter, even if the pid->ASID mapping changes whilst + we're away. In that case the cache will have been flushed when the + mapping was renewed. So the writebacks below will be nugatory (and + we'll doubtless have to fault the TLB entry/ies in again with the + new ASID), but it's a rare case. + */ + aligned_start = start & L1_CACHE_ALIGN_MASK; + addr = aligned_start; + while (addr < ull_end) { + asm __volatile__ ("ocbwb %0, 0" : : "r" (addr)); + addr += L1_CACHE_BYTES; + } +} + +/****************************************************************************/ + +/* These *MUST* lie in an area of virtual address space that's otherwise unused. */ +#define UNIQUE_EADDR_START 0xe0000000UL +#define UNIQUE_EADDR_END 0xe8000000UL + +static unsigned long sh64_make_unique_eaddr(unsigned long user_eaddr, unsigned long paddr) +{ + /* Given a physical address paddr, and a user virtual address + user_eaddr which will eventually be mapped to it, create a one-off + kernel-private eaddr mapped to the same paddr. This is used for + creating special destination pages for copy_user_page and + clear_user_page */ + + static unsigned long current_pointer = UNIQUE_EADDR_START; + unsigned long coloured_pointer; + + if (current_pointer == UNIQUE_EADDR_END) { + sh64_dcache_purge_all(); + current_pointer = UNIQUE_EADDR_START; + } + + coloured_pointer = (current_pointer & ~CACHE_OC_SYN_MASK) | (user_eaddr & CACHE_OC_SYN_MASK); + sh64_setup_dtlb_cache_slot(coloured_pointer, get_asid(), paddr); + + current_pointer += (PAGE_SIZE << CACHE_OC_N_SYNBITS); + + return coloured_pointer; +} + +/****************************************************************************/ + +static void sh64_copy_user_page_coloured(void *to, void *from, unsigned long address) +{ + void *coloured_to; + + /* Discard any existing cache entries of the wrong colour. These are + present quite often, if the kernel has recently used the page + internally, then given it up, then it's been allocated to the user. + */ + sh64_dcache_purge_coloured_phy_page(__pa(to), (unsigned long) to); + + coloured_to = (void *) sh64_make_unique_eaddr(address, __pa(to)); + sh64_page_copy(from, coloured_to); + + sh64_teardown_dtlb_cache_slot(); +} + +static void sh64_clear_user_page_coloured(void *to, unsigned long address) +{ + void *coloured_to; + + /* Discard any existing kernel-originated lines of the wrong colour (as + above) */ + sh64_dcache_purge_coloured_phy_page(__pa(to), (unsigned long) to); + + coloured_to = (void *) sh64_make_unique_eaddr(address, __pa(to)); + sh64_page_clear(coloured_to); + + sh64_teardown_dtlb_cache_slot(); +} + +#endif /* !CONFIG_DCACHE_DISABLED */ + +/****************************************************************************/ + +/*########################################################################## + EXTERNALLY CALLABLE API. + ##########################################################################*/ + +/* These functions are described in Documentation/cachetlb.txt. + Each one of these functions varies in behaviour depending on whether the + I-cache and/or D-cache are configured out. + + Note that the Linux term 'flush' corresponds to what is termed 'purge' in + the sh/sh64 jargon for the D-cache, i.e. write back dirty data then + invalidate the cache lines, and 'invalidate' for the I-cache. + */ + +#undef FLUSH_TRACE + +void flush_cache_all(void) +{ + /* Invalidate the entire contents of both caches, after writing back to + memory any dirty data from the D-cache. */ + sh64_dcache_purge_all(); + sh64_icache_inv_all(); +} + +/****************************************************************************/ + +void flush_cache_mm(struct mm_struct *mm) +{ + /* Invalidate an entire user-address space from both caches, after + writing back dirty data (e.g. for shared mmap etc). */ + + /* This could be coded selectively by inspecting all the tags then + doing 4*alloco on any set containing a match (as for + flush_cache_range), but fork/exit/execve (where this is called from) + are expensive anyway. */ + + /* Have to do a purge here, despite the comments re I-cache below. + There could be odd-coloured dirty data associated with the mm still + in the cache - if this gets written out through natural eviction + after the kernel has reused the page there will be chaos. + */ + + sh64_dcache_purge_all(); + + /* The mm being torn down won't ever be active again, so any Icache + lines tagged with its ASID won't be visible for the rest of the + lifetime of this ASID cycle. Before the ASID gets reused, there + will be a flush_cache_all. Hence we don't need to touch the + I-cache. This is similar to the lack of action needed in + flush_tlb_mm - see fault.c. */ +} + +/****************************************************************************/ + +void flush_cache_range(struct vm_area_struct *vma, unsigned long start, + unsigned long end) +{ + struct mm_struct *mm = vma->vm_mm; + + /* Invalidate (from both caches) the range [start,end) of virtual + addresses from the user address space specified by mm, after writing + back any dirty data. + + Note(1), 'end' is 1 byte beyond the end of the range to flush. + + Note(2), this is called with mm->page_table_lock held.*/ + + sh64_dcache_purge_user_range(mm, start, end); + sh64_icache_inv_user_page_range(mm, start, end); +} + +/****************************************************************************/ + +void flush_cache_page(struct vm_area_struct *vma, unsigned long eaddr, unsigned long pfn) +{ + /* Invalidate any entries in either cache for the vma within the user + address space vma->vm_mm for the page starting at virtual address + 'eaddr'. This seems to be used primarily in breaking COW. Note, + the I-cache must be searched too in case the page in question is + both writable and being executed from (e.g. stack trampolines.) + + Note(1), this is called with mm->page_table_lock held. + */ + + sh64_dcache_purge_phy_page(pfn << PAGE_SHIFT); + + if (vma->vm_flags & VM_EXEC) { + sh64_icache_inv_user_page(vma, eaddr); + } +} + +/****************************************************************************/ + +#ifndef CONFIG_DCACHE_DISABLED + +void copy_user_page(void *to, void *from, unsigned long address, struct page *page) +{ + /* 'from' and 'to' are kernel virtual addresses (within the superpage + mapping of the physical RAM). 'address' is the user virtual address + where the copy 'to' will be mapped after. This allows a custom + mapping to be used to ensure that the new copy is placed in the + right cache sets for the user to see it without having to bounce it + out via memory. Note however : the call to flush_page_to_ram in + (generic)/mm/memory.c:(break_cow) undoes all this good work in that one + very important case! + + TBD : can we guarantee that on every call, any cache entries for + 'from' are in the same colour sets as 'address' also? i.e. is this + always used just to deal with COW? (I suspect not). */ + + /* There are two possibilities here for when the page 'from' was last accessed: + * by the kernel : this is OK, no purge required. + * by the/a user (e.g. for break_COW) : need to purge. + + If the potential user mapping at 'address' is the same colour as + 'from' there is no need to purge any cache lines from the 'from' + page mapped into cache sets of colour 'address'. (The copy will be + accessing the page through 'from'). + */ + + if (((address ^ (unsigned long) from) & CACHE_OC_SYN_MASK) != 0) { + sh64_dcache_purge_coloured_phy_page(__pa(from), address); + } + + if (((address ^ (unsigned long) to) & CACHE_OC_SYN_MASK) == 0) { + /* No synonym problem on destination */ + sh64_page_copy(from, to); + } else { + sh64_copy_user_page_coloured(to, from, address); + } + + /* Note, don't need to flush 'from' page from the cache again - it's + done anyway by the generic code */ +} + +void clear_user_page(void *to, unsigned long address, struct page *page) +{ + /* 'to' is a kernel virtual address (within the superpage + mapping of the physical RAM). 'address' is the user virtual address + where the 'to' page will be mapped after. This allows a custom + mapping to be used to ensure that the new copy is placed in the + right cache sets for the user to see it without having to bounce it + out via memory. + */ + + if (((address ^ (unsigned long) to) & CACHE_OC_SYN_MASK) == 0) { + /* No synonym problem on destination */ + sh64_page_clear(to); + } else { + sh64_clear_user_page_coloured(to, address); + } +} + +#endif /* !CONFIG_DCACHE_DISABLED */ + +/****************************************************************************/ + +void flush_dcache_page(struct page *page) +{ + sh64_dcache_purge_phy_page(page_to_phys(page)); + wmb(); +} + +/****************************************************************************/ + +void flush_icache_range(unsigned long start, unsigned long end) +{ + /* Flush the range [start,end] of kernel virtual adddress space from + the I-cache. The corresponding range must be purged from the + D-cache also because the SH-5 doesn't have cache snooping between + the caches. The addresses will be visible through the superpage + mapping, therefore it's guaranteed that there no cache entries for + the range in cache sets of the wrong colour. + + Primarily used for cohering the I-cache after a module has + been loaded. */ + + /* We also make sure to purge the same range from the D-cache since + flush_page_to_ram() won't be doing this for us! */ + + sh64_dcache_purge_kernel_range(start, end); + wmb(); + sh64_icache_inv_kernel_range(start, end); +} + +/****************************************************************************/ + +void flush_icache_user_range(struct vm_area_struct *vma, + struct page *page, unsigned long addr, int len) +{ + /* Flush the range of user (defined by vma->vm_mm) address space + starting at 'addr' for 'len' bytes from the cache. The range does + not straddle a page boundary, the unique physical page containing + the range is 'page'. This seems to be used mainly for invalidating + an address range following a poke into the program text through the + ptrace() call from another process (e.g. for BRK instruction + insertion). */ + + sh64_dcache_purge_coloured_phy_page(page_to_phys(page), addr); + mb(); + + if (vma->vm_flags & VM_EXEC) { + sh64_icache_inv_user_small_range(vma->vm_mm, addr, len); + } +} + +/*########################################################################## + ARCH/SH64 PRIVATE CALLABLE API. + ##########################################################################*/ + +void flush_cache_sigtramp(unsigned long start, unsigned long end) +{ + /* For the address range [start,end), write back the data from the + D-cache and invalidate the corresponding region of the I-cache for + the current process. Used to flush signal trampolines on the stack + to make them executable. */ + + sh64_dcache_wback_current_user_range(start, end); + wmb(); + sh64_icache_inv_current_user_range(start, end); +} + diff --git a/arch/sh64/mm/extable.c b/arch/sh64/mm/extable.c new file mode 100644 index 00000000000..9da50e28b3f --- /dev/null +++ b/arch/sh64/mm/extable.c @@ -0,0 +1,81 @@ +/* + * This file is subject to the terms and conditions of the GNU General Public + * License. See the file "COPYING" in the main directory of this archive + * for more details. + * + * arch/sh64/mm/extable.c + * + * Copyright (C) 2003 Richard Curnow + * Copyright (C) 2003, 2004 Paul Mundt + * + * Cloned from the 2.5 SH version.. + */ +#include +#include +#include +#include + +extern unsigned long copy_user_memcpy, copy_user_memcpy_end; +extern void __copy_user_fixup(void); + +static const struct exception_table_entry __copy_user_fixup_ex = { + .fixup = (unsigned long)&__copy_user_fixup, +}; + +/* Some functions that may trap due to a bad user-mode address have too many loads + and stores in them to make it at all practical to label each one and put them all in + the main exception table. + + In particular, the fast memcpy routine is like this. It's fix-up is just to fall back + to a slow byte-at-a-time copy, which is handled the conventional way. So it's functionally + OK to just handle any trap occurring in the fast memcpy with that fixup. */ +static const struct exception_table_entry *check_exception_ranges(unsigned long addr) +{ + if ((addr >= (unsigned long)©_user_memcpy) && + (addr <= (unsigned long)©_user_memcpy_end)) + return &__copy_user_fixup_ex; + + return NULL; +} + +/* Simple binary search */ +const struct exception_table_entry * +search_extable(const struct exception_table_entry *first, + const struct exception_table_entry *last, + unsigned long value) +{ + const struct exception_table_entry *mid; + + mid = check_exception_ranges(value); + if (mid) + return mid; + + while (first <= last) { + long diff; + + mid = (last - first) / 2 + first; + diff = mid->insn - value; + if (diff == 0) + return mid; + else if (diff < 0) + first = mid+1; + else + last = mid-1; + } + + return NULL; +} + +int fixup_exception(struct pt_regs *regs) +{ + const struct exception_table_entry *fixup; + + fixup = search_exception_tables(regs->pc); + if (fixup) { + regs->pc = fixup->fixup; + return 1; + } + + return 0; +} + diff --git a/arch/sh64/mm/fault.c b/arch/sh64/mm/fault.c new file mode 100644 index 00000000000..a24932881db --- /dev/null +++ b/arch/sh64/mm/fault.c @@ -0,0 +1,601 @@ +/* + * This file is subject to the terms and conditions of the GNU General Public + * License. See the file "COPYING" in the main directory of this archive + * for more details. + * + * arch/sh64/mm/fault.c + * + * Copyright (C) 2000, 2001 Paolo Alberelli + * Copyright (C) 2003 Richard Curnow (/proc/tlb, bug fixes) + * Copyright (C) 2003 Paul Mundt + * + */ + +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include + +#include +#include +#include +#include +#include +#include +#include /* required by inline asm statements */ + +#if defined(CONFIG_SH64_PROC_TLB) +#include +#include +/* Count numbers of tlb refills in each region */ +static unsigned long long calls_to_update_mmu_cache = 0ULL; +static unsigned long long calls_to_flush_tlb_page = 0ULL; +static unsigned long long calls_to_flush_tlb_range = 0ULL; +static unsigned long long calls_to_flush_tlb_mm = 0ULL; +static unsigned long long calls_to_flush_tlb_all = 0ULL; +unsigned long long calls_to_do_slow_page_fault = 0ULL; +unsigned long long calls_to_do_fast_page_fault = 0ULL; + +/* Count size of ranges for flush_tlb_range */ +static unsigned long long flush_tlb_range_1 = 0ULL; +static unsigned long long flush_tlb_range_2 = 0ULL; +static unsigned long long flush_tlb_range_3_4 = 0ULL; +static unsigned long long flush_tlb_range_5_7 = 0ULL; +static unsigned long long flush_tlb_range_8_11 = 0ULL; +static unsigned long long flush_tlb_range_12_15 = 0ULL; +static unsigned long long flush_tlb_range_16_up = 0ULL; + +static unsigned long long page_not_present = 0ULL; + +#endif + +extern void die(const char *,struct pt_regs *,long); + +#define PFLAG(val,flag) (( (val) & (flag) ) ? #flag : "" ) +#define PPROT(flag) PFLAG(pgprot_val(prot),flag) + +static inline void print_prots(pgprot_t prot) +{ + printk("prot is 0x%08lx\n",pgprot_val(prot)); + + printk("%s %s %s %s %s\n",PPROT(_PAGE_SHARED),PPROT(_PAGE_READ), + PPROT(_PAGE_EXECUTE),PPROT(_PAGE_WRITE),PPROT(_PAGE_USER)); +} + +static inline void print_vma(struct vm_area_struct *vma) +{ + printk("vma start 0x%08lx\n", vma->vm_start); + printk("vma end 0x%08lx\n", vma->vm_end); + + print_prots(vma->vm_page_prot); + printk("vm_flags 0x%08lx\n", vma->vm_flags); +} + +static inline void print_task(struct task_struct *tsk) +{ + printk("Task pid %d\n", tsk->pid); +} + +static pte_t *lookup_pte(struct mm_struct *mm, unsigned long address) +{ + pgd_t *dir; + pmd_t *pmd; + pte_t *pte; + pte_t entry; + + dir = pgd_offset(mm, address); + if (pgd_none(*dir)) { + return NULL; + } + + pmd = pmd_offset(dir, address); + if (pmd_none(*pmd)) { + return NULL; + } + + pte = pte_offset_kernel(pmd, address); + entry = *pte; + + if (pte_none(entry)) { + return NULL; + } + if (!pte_present(entry)) { + return NULL; + } + + return pte; +} + +/* + * This routine handles page faults. It determines the address, + * and the problem, and then passes it off to one of the appropriate + * routines. + */ +asmlinkage void do_page_fault(struct pt_regs *regs, unsigned long writeaccess, + unsigned long textaccess, unsigned long address) +{ + struct task_struct *tsk; + struct mm_struct *mm; + struct vm_area_struct * vma; + const struct exception_table_entry *fixup; + pte_t *pte; + +#if defined(CONFIG_SH64_PROC_TLB) + ++calls_to_do_slow_page_fault; +#endif + + /* SIM + * Note this is now called with interrupts still disabled + * This is to cope with being called for a missing IO port + * address with interupts disabled. This should be fixed as + * soon as we have a better 'fast path' miss handler. + * + * Plus take care how you try and debug this stuff. + * For example, writing debug data to a port which you + * have just faulted on is not going to work. + */ + + tsk = current; + mm = tsk->mm; + + /* Not an IO address, so reenable interrupts */ + local_irq_enable(); + + /* + * If we're in an interrupt or have no user + * context, we must not take the fault.. + */ + if (in_interrupt() || !mm) + goto no_context; + + /* TLB misses upon some cache flushes get done under cli() */ + down_read(&mm->mmap_sem); + + vma = find_vma(mm, address); + + if (!vma) { +#ifdef DEBUG_FAULT + print_task(tsk); + printk("%s:%d fault, address is 0x%08x PC %016Lx textaccess %d writeaccess %d\n", + __FUNCTION__,__LINE__, + address,regs->pc,textaccess,writeaccess); + show_regs(regs); +#endif + goto bad_area; + } + if (vma->vm_start <= address) { + goto good_area; + } + + if (!(vma->vm_flags & VM_GROWSDOWN)) { +#ifdef DEBUG_FAULT + print_task(tsk); + printk("%s:%d fault, address is 0x%08x PC %016Lx textaccess %d writeaccess %d\n", + __FUNCTION__,__LINE__, + address,regs->pc,textaccess,writeaccess); + show_regs(regs); + + print_vma(vma); +#endif + goto bad_area; + } + if (expand_stack(vma, address)) { +#ifdef DEBUG_FAULT + print_task(tsk); + printk("%s:%d fault, address is 0x%08x PC %016Lx textaccess %d writeaccess %d\n", + __FUNCTION__,__LINE__, + address,regs->pc,textaccess,writeaccess); + show_regs(regs); +#endif + goto bad_area; + } +/* + * Ok, we have a good vm_area for this memory access, so + * we can handle it.. + */ +good_area: + if (textaccess) { + if (!(vma->vm_flags & VM_EXEC)) + goto bad_area; + } else { + if (writeaccess) { + if (!(vma->vm_flags & VM_WRITE)) + goto bad_area; + } else { + if (!(vma->vm_flags & VM_READ)) + goto bad_area; + } + } + + /* + * If for any reason at all we couldn't handle the fault, + * make sure we exit gracefully rather than endlessly redo + * the fault. + */ +survive: + switch (handle_mm_fault(mm, vma, address, writeaccess)) { + case 1: + tsk->min_flt++; + break; + case 2: + tsk->maj_flt++; + break; + case 0: + goto do_sigbus; + default: + goto out_of_memory; + } + /* If we get here, the page fault has been handled. Do the TLB refill + now from the newly-setup PTE, to avoid having to fault again right + away on the same instruction. */ + pte = lookup_pte (mm, address); + if (!pte) { + /* From empirical evidence, we can get here, due to + !pte_present(pte). (e.g. if a swap-in occurs, and the page + is swapped back out again before the process that wanted it + gets rescheduled?) */ + goto no_pte; + } + + __do_tlb_refill(address, textaccess, pte); + +no_pte: + + up_read(&mm->mmap_sem); + return; + +/* + * Something tried to access memory that isn't in our memory map.. + * Fix it, but check if it's kernel or user first.. + */ +bad_area: +#ifdef DEBUG_FAULT + printk("fault:bad area\n"); +#endif + up_read(&mm->mmap_sem); + + if (user_mode(regs)) { + static int count=0; + siginfo_t info; + if (count < 4) { + /* This is really to help debug faults when starting + * usermode, so only need a few */ + count++; + printk("user mode bad_area address=%08lx pid=%d (%s) pc=%08lx\n", + address, current->pid, current->comm, + (unsigned long) regs->pc); +#if 0 + show_regs(regs); +#endif + } + if (tsk->pid == 1) { + panic("INIT had user mode bad_area\n"); + } + tsk->thread.address = address; + tsk->thread.error_code = writeaccess; + info.si_signo = SIGSEGV; + info.si_errno = 0; + info.si_addr = (void *) address; + force_sig_info(SIGSEGV, &info, tsk); + return; + } + +no_context: +#ifdef DEBUG_FAULT + printk("fault:No context\n"); +#endif + /* Are we prepared to handle this kernel fault? */ + fixup = search_exception_tables(regs->pc); + if (fixup) { + regs->pc = fixup->fixup; + return; + } + +/* + * Oops. The kernel tried to access some bad page. We'll have to + * terminate things with extreme prejudice. + * + */ + if (address < PAGE_SIZE) + printk(KERN_ALERT "Unable to handle kernel NULL pointer dereference"); + else + printk(KERN_ALERT "Unable to handle kernel paging request"); + printk(" at virtual address %08lx\n", address); + printk(KERN_ALERT "pc = %08Lx%08Lx\n", regs->pc >> 32, regs->pc & 0xffffffff); + die("Oops", regs, writeaccess); + do_exit(SIGKILL); + +/* + * We ran out of memory, or some other thing happened to us that made + * us unable to handle the page fault gracefully. + */ +out_of_memory: + if (current->pid == 1) { + panic("INIT out of memory\n"); + yield(); + goto survive; + } + printk("fault:Out of memory\n"); + up_read(&mm->mmap_sem); + if (current->pid == 1) { + yield(); + down_read(&mm->mmap_sem); + goto survive; + } + printk("VM: killing process %s\n", tsk->comm); + if (user_mode(regs)) + do_exit(SIGKILL); + goto no_context; + +do_sigbus: + printk("fault:Do sigbus\n"); + up_read(&mm->mmap_sem); + + /* + * Send a sigbus, regardless of whether we were in kernel + * or user mode. + */ + tsk->thread.address = address; + tsk->thread.error_code = writeaccess; + tsk->thread.trap_no = 14; + force_sig(SIGBUS, tsk); + + /* Kernel mode? Handle exceptions or die */ + if (!user_mode(regs)) + goto no_context; +} + + +void flush_tlb_all(void); + +void update_mmu_cache(struct vm_area_struct * vma, + unsigned long address, pte_t pte) +{ +#if defined(CONFIG_SH64_PROC_TLB) + ++calls_to_update_mmu_cache; +#endif + + /* + * This appears to get called once for every pte entry that gets + * established => I don't think it's efficient to try refilling the + * TLBs with the pages - some may not get accessed even. Also, for + * executable pages, it is impossible to determine reliably here which + * TLB they should be mapped into (or both even). + * + * So, just do nothing here and handle faults on demand. In the + * TLBMISS handling case, the refill is now done anyway after the pte + * has been fixed up, so that deals with most useful cases. + */ +} + +static void __flush_tlb_page(struct vm_area_struct *vma, unsigned long page) +{ + unsigned long long match, pteh=0, lpage; + unsigned long tlb; + struct mm_struct *mm; + + mm = vma->vm_mm; + + if (mm->context == NO_CONTEXT) + return; + + /* + * Sign-extend based on neff. + */ + lpage = (page & NEFF_SIGN) ? (page | NEFF_MASK) : page; + match = ((mm->context & MMU_CONTEXT_ASID_MASK) << PTEH_ASID_SHIFT) | PTEH_VALID; + match |= lpage; + + /* Do ITLB : don't bother for pages in non-exectutable VMAs */ + if (vma->vm_flags & VM_EXEC) { + for_each_itlb_entry(tlb) { + asm volatile ("getcfg %1, 0, %0" + : "=r" (pteh) + : "r" (tlb) ); + + if (pteh == match) { + __flush_tlb_slot(tlb); + break; + } + + } + } + + /* Do DTLB : any page could potentially be in here. */ + for_each_dtlb_entry(tlb) { + asm volatile ("getcfg %1, 0, %0" + : "=r" (pteh) + : "r" (tlb) ); + + if (pteh == match) { + __flush_tlb_slot(tlb); + break; + } + + } +} + +void flush_tlb_page(struct vm_area_struct *vma, unsigned long page) +{ + unsigned long flags; + +#if defined(CONFIG_SH64_PROC_TLB) + ++calls_to_flush_tlb_page; +#endif + + if (vma->vm_mm) { + page &= PAGE_MASK; + local_irq_save(flags); + __flush_tlb_page(vma, page); + local_irq_restore(flags); + } +} + +void flush_tlb_range(struct vm_area_struct *vma, unsigned long start, + unsigned long end) +{ + unsigned long flags; + unsigned long long match, pteh=0, pteh_epn, pteh_low; + unsigned long tlb; + struct mm_struct *mm; + + mm = vma->vm_mm; + +#if defined(CONFIG_SH64_PROC_TLB) + ++calls_to_flush_tlb_range; + + { + unsigned long size = (end - 1) - start; + size >>= 12; /* divide by PAGE_SIZE */ + size++; /* end=start+4096 => 1 page */ + switch (size) { + case 1 : flush_tlb_range_1++; break; + case 2 : flush_tlb_range_2++; break; + case 3 ... 4 : flush_tlb_range_3_4++; break; + case 5 ... 7 : flush_tlb_range_5_7++; break; + case 8 ... 11 : flush_tlb_range_8_11++; break; + case 12 ... 15 : flush_tlb_range_12_15++; break; + default : flush_tlb_range_16_up++; break; + } + } +#endif + + if (mm->context == NO_CONTEXT) + return; + + local_irq_save(flags); + + start &= PAGE_MASK; + end &= PAGE_MASK; + + match = ((mm->context & MMU_CONTEXT_ASID_MASK) << PTEH_ASID_SHIFT) | PTEH_VALID; + + /* Flush ITLB */ + for_each_itlb_entry(tlb) { + asm volatile ("getcfg %1, 0, %0" + : "=r" (pteh) + : "r" (tlb) ); + + pteh_epn = pteh & PAGE_MASK; + pteh_low = pteh & ~PAGE_MASK; + + if (pteh_low == match && pteh_epn >= start && pteh_epn <= end) + __flush_tlb_slot(tlb); + } + + /* Flush DTLB */ + for_each_dtlb_entry(tlb) { + asm volatile ("getcfg %1, 0, %0" + : "=r" (pteh) + : "r" (tlb) ); + + pteh_epn = pteh & PAGE_MASK; + pteh_low = pteh & ~PAGE_MASK; + + if (pteh_low == match && pteh_epn >= start && pteh_epn <= end) + __flush_tlb_slot(tlb); + } + + local_irq_restore(flags); +} + +void flush_tlb_mm(struct mm_struct *mm) +{ + unsigned long flags; + +#if defined(CONFIG_SH64_PROC_TLB) + ++calls_to_flush_tlb_mm; +#endif + + if (mm->context == NO_CONTEXT) + return; + + local_irq_save(flags); + + mm->context=NO_CONTEXT; + if(mm==current->mm) + activate_context(mm); + + local_irq_restore(flags); + +} + +void flush_tlb_all(void) +{ + /* Invalidate all, including shared pages, excluding fixed TLBs */ + + unsigned long flags, tlb; + +#if defined(CONFIG_SH64_PROC_TLB) + ++calls_to_flush_tlb_all; +#endif + + local_irq_save(flags); + + /* Flush each ITLB entry */ + for_each_itlb_entry(tlb) { + __flush_tlb_slot(tlb); + } + + /* Flush each DTLB entry */ + for_each_dtlb_entry(tlb) { + __flush_tlb_slot(tlb); + } + + local_irq_restore(flags); +} + +void flush_tlb_kernel_range(unsigned long start, unsigned long end) +{ + /* FIXME: Optimize this later.. */ + flush_tlb_all(); +} + +#if defined(CONFIG_SH64_PROC_TLB) +/* Procfs interface to read the performance information */ + +static int +tlb_proc_info(char *buf, char **start, off_t fpos, int length, int *eof, void *data) +{ + int len=0; + len += sprintf(buf+len, "do_fast_page_fault called %12lld times\n", calls_to_do_fast_page_fault); + len += sprintf(buf+len, "do_slow_page_fault called %12lld times\n", calls_to_do_slow_page_fault); + len += sprintf(buf+len, "update_mmu_cache called %12lld times\n", calls_to_update_mmu_cache); + len += sprintf(buf+len, "flush_tlb_page called %12lld times\n", calls_to_flush_tlb_page); + len += sprintf(buf+len, "flush_tlb_range called %12lld times\n", calls_to_flush_tlb_range); + len += sprintf(buf+len, "flush_tlb_mm called %12lld times\n", calls_to_flush_tlb_mm); + len += sprintf(buf+len, "flush_tlb_all called %12lld times\n", calls_to_flush_tlb_all); + len += sprintf(buf+len, "flush_tlb_range_sizes\n" + " 1 : %12lld\n" + " 2 : %12lld\n" + " 3 - 4 : %12lld\n" + " 5 - 7 : %12lld\n" + " 8 - 11 : %12lld\n" + "12 - 15 : %12lld\n" + "16+ : %12lld\n", + flush_tlb_range_1, flush_tlb_range_2, flush_tlb_range_3_4, + flush_tlb_range_5_7, flush_tlb_range_8_11, flush_tlb_range_12_15, + flush_tlb_range_16_up); + len += sprintf(buf+len, "page not present %12lld times\n", page_not_present); + *eof = 1; + return len; +} + +static int __init register_proc_tlb(void) +{ + create_proc_read_entry("tlb", 0, NULL, tlb_proc_info, NULL); + return 0; +} + +__initcall(register_proc_tlb); + +#endif diff --git a/arch/sh64/mm/hugetlbpage.c b/arch/sh64/mm/hugetlbpage.c new file mode 100644 index 00000000000..bcad2aefa4e --- /dev/null +++ b/arch/sh64/mm/hugetlbpage.c @@ -0,0 +1,264 @@ +/* + * arch/sh64/mm/hugetlbpage.c + * + * SuperH HugeTLB page support. + * + * Cloned from sparc64 by Paul Mundt. + * + * Copyright (C) 2002, 2003 David S. Miller (davem@redhat.com) + */ + +#include +#include +#include +#include +#include +#include +#include +#include +#include + +#include +#include +#include +#include +#include + +static pte_t *huge_pte_alloc(struct mm_struct *mm, unsigned long addr) +{ + pgd_t *pgd; + pmd_t *pmd; + pte_t *pte = NULL; + + pgd = pgd_offset(mm, addr); + if (pgd) { + pmd = pmd_alloc(mm, pgd, addr); + if (pmd) + pte = pte_alloc_map(mm, pmd, addr); + } + return pte; +} + +static pte_t *huge_pte_offset(struct mm_struct *mm, unsigned long addr) +{ + pgd_t *pgd; + pmd_t *pmd; + pte_t *pte = NULL; + + pgd = pgd_offset(mm, addr); + if (pgd) { + pmd = pmd_offset(pgd, addr); + if (pmd) + pte = pte_offset_map(pmd, addr); + } + return pte; +} + +#define mk_pte_huge(entry) do { pte_val(entry) |= _PAGE_SZHUGE; } while (0) + +static void set_huge_pte(struct mm_struct *mm, struct vm_area_struct *vma, + struct page *page, pte_t * page_table, int write_access) +{ + unsigned long i; + pte_t entry; + + add_mm_counter(mm, rss, HPAGE_SIZE / PAGE_SIZE); + + if (write_access) + entry = pte_mkwrite(pte_mkdirty(mk_pte(page, + vma->vm_page_prot))); + else + entry = pte_wrprotect(mk_pte(page, vma->vm_page_prot)); + entry = pte_mkyoung(entry); + mk_pte_huge(entry); + + for (i = 0; i < (1 << HUGETLB_PAGE_ORDER); i++) { + set_pte(page_table, entry); + page_table++; + + pte_val(entry) += PAGE_SIZE; + } +} + +/* + * This function checks for proper alignment of input addr and len parameters. + */ +int is_aligned_hugepage_range(unsigned long addr, unsigned long len) +{ + if (len & ~HPAGE_MASK) + return -EINVAL; + if (addr & ~HPAGE_MASK) + return -EINVAL; + return 0; +} + +int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src, + struct vm_area_struct *vma) +{ + pte_t *src_pte, *dst_pte, entry; + struct page *ptepage; + unsigned long addr = vma->vm_start; + unsigned long end = vma->vm_end; + int i; + + while (addr < end) { + dst_pte = huge_pte_alloc(dst, addr); + if (!dst_pte) + goto nomem; + src_pte = huge_pte_offset(src, addr); + BUG_ON(!src_pte || pte_none(*src_pte)); + entry = *src_pte; + ptepage = pte_page(entry); + get_page(ptepage); + for (i = 0; i < (1 << HUGETLB_PAGE_ORDER); i++) { + set_pte(dst_pte, entry); + pte_val(entry) += PAGE_SIZE; + dst_pte++; + } + add_mm_counter(dst, rss, HPAGE_SIZE / PAGE_SIZE); + addr += HPAGE_SIZE; + } + return 0; + +nomem: + return -ENOMEM; +} + +int follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma, + struct page **pages, struct vm_area_struct **vmas, + unsigned long *position, int *length, int i) +{ + unsigned long vaddr = *position; + int remainder = *length; + + WARN_ON(!is_vm_hugetlb_page(vma)); + + while (vaddr < vma->vm_end && remainder) { + if (pages) { + pte_t *pte; + struct page *page; + + pte = huge_pte_offset(mm, vaddr); + + /* hugetlb should be locked, and hence, prefaulted */ + BUG_ON(!pte || pte_none(*pte)); + + page = pte_page(*pte); + + WARN_ON(!PageCompound(page)); + + get_page(page); + pages[i] = page; + } + + if (vmas) + vmas[i] = vma; + + vaddr += PAGE_SIZE; + --remainder; + ++i; + } + + *length = remainder; + *position = vaddr; + + return i; +} + +struct page *follow_huge_addr(struct mm_struct *mm, + unsigned long address, int write) +{ + return ERR_PTR(-EINVAL); +} + +int pmd_huge(pmd_t pmd) +{ + return 0; +} + +struct page *follow_huge_pmd(struct mm_struct *mm, unsigned long address, + pmd_t *pmd, int write) +{ + return NULL; +} + +void unmap_hugepage_range(struct vm_area_struct *vma, + unsigned long start, unsigned long end) +{ + struct mm_struct *mm = vma->vm_mm; + unsigned long address; + pte_t *pte; + struct page *page; + int i; + + BUG_ON(start & (HPAGE_SIZE - 1)); + BUG_ON(end & (HPAGE_SIZE - 1)); + + for (address = start; address < end; address += HPAGE_SIZE) { + pte = huge_pte_offset(mm, address); + BUG_ON(!pte); + if (pte_none(*pte)) + continue; + page = pte_page(*pte); + put_page(page); + for (i = 0; i < (1 << HUGETLB_PAGE_ORDER); i++) { + pte_clear(mm, address+(i*PAGE_SIZE), pte); + pte++; + } + } + add_mm_counter(mm, rss, -((end - start) >> PAGE_SHIFT)); + flush_tlb_range(vma, start, end); +} + +int hugetlb_prefault(struct address_space *mapping, struct vm_area_struct *vma) +{ + struct mm_struct *mm = current->mm; + unsigned long addr; + int ret = 0; + + BUG_ON(vma->vm_start & ~HPAGE_MASK); + BUG_ON(vma->vm_end & ~HPAGE_MASK); + + spin_lock(&mm->page_table_lock); + for (addr = vma->vm_start; addr < vma->vm_end; addr += HPAGE_SIZE) { + unsigned long idx; + pte_t *pte = huge_pte_alloc(mm, addr); + struct page *page; + + if (!pte) { + ret = -ENOMEM; + goto out; + } + if (!pte_none(*pte)) + continue; + + idx = ((addr - vma->vm_start) >> HPAGE_SHIFT) + + (vma->vm_pgoff >> (HPAGE_SHIFT - PAGE_SHIFT)); + page = find_get_page(mapping, idx); + if (!page) { + /* charge the fs quota first */ + if (hugetlb_get_quota(mapping)) { + ret = -ENOMEM; + goto out; + } + page = alloc_huge_page(); + if (!page) { + hugetlb_put_quota(mapping); + ret = -ENOMEM; + goto out; + } + ret = add_to_page_cache(page, mapping, idx, GFP_ATOMIC); + if (! ret) { + unlock_page(page); + } else { + hugetlb_put_quota(mapping); + free_huge_page(page); + goto out; + } + } + set_huge_pte(mm, vma, page, pte, vma->vm_flags & VM_WRITE); + } +out: + spin_unlock(&mm->page_table_lock); + return ret; +} diff --git a/arch/sh64/mm/init.c b/arch/sh64/mm/init.c new file mode 100644 index 00000000000..a65e8bb2c3c --- /dev/null +++ b/arch/sh64/mm/init.c @@ -0,0 +1,196 @@ +/* + * This file is subject to the terms and conditions of the GNU General Public + * License. See the file "COPYING" in the main directory of this archive + * for more details. + * + * arch/sh64/mm/init.c + * + * Copyright (C) 2000, 2001 Paolo Alberelli + * Copyright (C) 2003, 2004 Paul Mundt + * + */ + +#include +#include +#include +#include +#include + +#include +#include +#include +#include +#include + +#ifdef CONFIG_BLK_DEV_INITRD +#include +#endif + +DEFINE_PER_CPU(struct mmu_gather, mmu_gathers); + +/* + * Cache of MMU context last used. + */ +unsigned long mmu_context_cache; +pgd_t * mmu_pdtp_cache; +int after_bootmem = 0; + +/* + * BAD_PAGE is the page that is used for page faults when linux + * is out-of-memory. Older versions of linux just did a + * do_exit(), but using this instead means there is less risk + * for a process dying in kernel mode, possibly leaving an inode + * unused etc.. + * + * BAD_PAGETABLE is the accompanying page-table: it is initialized + * to point to BAD_PAGE entries. + * + * ZERO_PAGE is a special page that is used for zero-initialized + * data and COW. + */ + +extern unsigned char empty_zero_page[PAGE_SIZE]; +extern unsigned char empty_bad_page[PAGE_SIZE]; +extern pte_t empty_bad_pte_table[PTRS_PER_PTE]; +extern pgd_t swapper_pg_dir[PTRS_PER_PGD]; + +extern char _text, _etext, _edata, __bss_start, _end; +extern char __init_begin, __init_end; + +/* It'd be good if these lines were in the standard header file. */ +#define START_PFN (NODE_DATA(0)->bdata->node_boot_start >> PAGE_SHIFT) +#define MAX_LOW_PFN (NODE_DATA(0)->bdata->node_low_pfn) + + +void show_mem(void) +{ + int i, total = 0, reserved = 0; + int shared = 0, cached = 0; + + printk("Mem-info:\n"); + show_free_areas(); + printk("Free swap: %6ldkB\n",nr_swap_pages<<(PAGE_SHIFT-10)); + i = max_mapnr; + while (i-- > 0) { + total++; + if (PageReserved(mem_map+i)) + reserved++; + else if (PageSwapCache(mem_map+i)) + cached++; + else if (page_count(mem_map+i)) + shared += page_count(mem_map+i) - 1; + } + printk("%d pages of RAM\n",total); + printk("%d reserved pages\n",reserved); + printk("%d pages shared\n",shared); + printk("%d pages swap cached\n",cached); + printk("%ld pages in page table cache\n",pgtable_cache_size); +} + +/* + * paging_init() sets up the page tables. + * + * head.S already did a lot to set up address translation for the kernel. + * Here we comes with: + * . MMU enabled + * . ASID set (SR) + * . some 512MB regions being mapped of which the most relevant here is: + * . CACHED segment (ASID 0 [irrelevant], shared AND NOT user) + * . possible variable length regions being mapped as: + * . UNCACHED segment (ASID 0 [irrelevant], shared AND NOT user) + * . All of the memory regions are placed, independently from the platform + * on high addresses, above 0x80000000. + * . swapper_pg_dir is already cleared out by the .space directive + * in any case swapper does not require a real page directory since + * it's all kernel contained. + * + * Those pesky NULL-reference errors in the kernel are then + * dealt with by not mapping address 0x00000000 at all. + * + */ +void __init paging_init(void) +{ + unsigned long zones_size[MAX_NR_ZONES] = {0, 0, 0}; + + pgd_init((unsigned long)swapper_pg_dir); + pgd_init((unsigned long)swapper_pg_dir + + sizeof(pgd_t) * USER_PTRS_PER_PGD); + + mmu_context_cache = MMU_CONTEXT_FIRST_VERSION; + + /* + * All memory is good as ZONE_NORMAL (fall-through) and ZONE_DMA. + */ + zones_size[ZONE_DMA] = MAX_LOW_PFN - START_PFN; + NODE_DATA(0)->node_mem_map = NULL; + free_area_init_node(0, NODE_DATA(0), zones_size, __MEMORY_START >> PAGE_SHIFT, 0); +} + +void __init mem_init(void) +{ + int codesize, reservedpages, datasize, initsize; + int tmp; + + max_mapnr = num_physpages = MAX_LOW_PFN - START_PFN; + high_memory = (void *)__va(MAX_LOW_PFN * PAGE_SIZE); + + /* + * Clear the zero-page. + * This is not required but we might want to re-use + * this very page to pass boot parameters, one day. + */ + memset(empty_zero_page, 0, PAGE_SIZE); + + /* this will put all low memory onto the freelists */ + totalram_pages += free_all_bootmem_node(NODE_DATA(0)); + reservedpages = 0; + for (tmp = 0; tmp < num_physpages; tmp++) + /* + * Only count reserved RAM pages + */ + if (PageReserved(mem_map+tmp)) + reservedpages++; + + after_bootmem = 1; + + codesize = (unsigned long) &_etext - (unsigned long) &_text; + datasize = (unsigned long) &_edata - (unsigned long) &_etext; + initsize = (unsigned long) &__init_end - (unsigned long) &__init_begin; + + printk("Memory: %luk/%luk available (%dk kernel code, %dk reserved, %dk data, %dk init)\n", + (unsigned long) nr_free_pages() << (PAGE_SHIFT-10), + max_mapnr << (PAGE_SHIFT-10), + codesize >> 10, + reservedpages << (PAGE_SHIFT-10), + datasize >> 10, + initsize >> 10); +} + +void free_initmem(void) +{ + unsigned long addr; + + addr = (unsigned long)(&__init_begin); + for (; addr < (unsigned long)(&__init_end); addr += PAGE_SIZE) { + ClearPageReserved(virt_to_page(addr)); + set_page_count(virt_to_page(addr), 1); + free_page(addr); + totalram_pages++; + } + printk ("Freeing unused kernel memory: %ldk freed\n", (&__init_end - &__init_begin) >> 10); +} + +#ifdef CONFIG_BLK_DEV_INITRD +void free_initrd_mem(unsigned long start, unsigned long end) +{ + unsigned long p; + for (p = start; p < end; p += PAGE_SIZE) { + ClearPageReserved(virt_to_page(p)); + set_page_count(virt_to_page(p), 1); + free_page(p); + totalram_pages++; + } + printk ("Freeing initrd memory: %ldk freed\n", (end - start) >> 10); +} +#endif + diff --git a/arch/sh64/mm/ioremap.c b/arch/sh64/mm/ioremap.c new file mode 100644 index 00000000000..f4003da556b --- /dev/null +++ b/arch/sh64/mm/ioremap.c @@ -0,0 +1,469 @@ +/* + * This file is subject to the terms and conditions of the GNU General Public + * License. See the file "COPYING" in the main directory of this archive + * for more details. + * + * arch/sh64/mm/ioremap.c + * + * Copyright (C) 2000, 2001 Paolo Alberelli + * Copyright (C) 2003, 2004 Paul Mundt + * + * Mostly derived from arch/sh/mm/ioremap.c which, in turn is mostly + * derived from arch/i386/mm/ioremap.c . + * + * (C) Copyright 1995 1996 Linus Torvalds + */ +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include + +static void shmedia_mapioaddr(unsigned long, unsigned long); +static unsigned long shmedia_ioremap(struct resource *, u32, int); + +static inline void remap_area_pte(pte_t * pte, unsigned long address, unsigned long size, + unsigned long phys_addr, unsigned long flags) +{ + unsigned long end; + unsigned long pfn; + pgprot_t pgprot = __pgprot(_PAGE_PRESENT | _PAGE_READ | + _PAGE_WRITE | _PAGE_DIRTY | + _PAGE_ACCESSED | _PAGE_SHARED | flags); + + address &= ~PMD_MASK; + end = address + size; + if (end > PMD_SIZE) + end = PMD_SIZE; + if (address >= end) + BUG(); + + pfn = phys_addr >> PAGE_SHIFT; + + pr_debug(" %s: pte %p address %lx size %lx phys_addr %lx\n", + __FUNCTION__,pte,address,size,phys_addr); + + do { + if (!pte_none(*pte)) { + printk("remap_area_pte: page already exists\n"); + BUG(); + } + + set_pte(pte, pfn_pte(pfn, pgprot)); + address += PAGE_SIZE; + pfn++; + pte++; + } while (address && (address < end)); +} + +static inline int remap_area_pmd(pmd_t * pmd, unsigned long address, unsigned long size, + unsigned long phys_addr, unsigned long flags) +{ + unsigned long end; + + address &= ~PGDIR_MASK; + end = address + size; + + if (end > PGDIR_SIZE) + end = PGDIR_SIZE; + + phys_addr -= address; + + if (address >= end) + BUG(); + + do { + pte_t * pte = pte_alloc_kernel(&init_mm, pmd, address); + if (!pte) + return -ENOMEM; + remap_area_pte(pte, address, end - address, address + phys_addr, flags); + address = (address + PMD_SIZE) & PMD_MASK; + pmd++; + } while (address && (address < end)); + return 0; +} + +static int remap_area_pages(unsigned long address, unsigned long phys_addr, + unsigned long size, unsigned long flags) +{ + int error; + pgd_t * dir; + unsigned long end = address + size; + + phys_addr -= address; + dir = pgd_offset_k(address); + flush_cache_all(); + if (address >= end) + BUG(); + spin_lock(&init_mm.page_table_lock); + do { + pmd_t *pmd = pmd_alloc(&init_mm, dir, address); + error = -ENOMEM; + if (!pmd) + break; + if (remap_area_pmd(pmd, address, end - address, + phys_addr + address, flags)) { + break; + } + error = 0; + address = (address + PGDIR_SIZE) & PGDIR_MASK; + dir++; + } while (address && (address < end)); + spin_unlock(&init_mm.page_table_lock); + flush_tlb_all(); + return 0; +} + +/* + * Generic mapping function (not visible outside): + */ + +/* + * Remap an arbitrary physical address space into the kernel virtual + * address space. Needed when the kernel wants to access high addresses + * directly. + * + * NOTE! We need to allow non-page-aligned mappings too: we will obviously + * have to convert them into an offset in a page-aligned mapping, but the + * caller shouldn't need to know that small detail. + */ +void * __ioremap(unsigned long phys_addr, unsigned long size, unsigned long flags) +{ + void * addr; + struct vm_struct * area; + unsigned long offset, last_addr; + + /* Don't allow wraparound or zero size */ + last_addr = phys_addr + size - 1; + if (!size || last_addr < phys_addr) + return NULL; + + /* + * Mappings have to be page-aligned + */ + offset = phys_addr & ~PAGE_MASK; + phys_addr &= PAGE_MASK; + size = PAGE_ALIGN(last_addr + 1) - phys_addr; + + /* + * Ok, go for it.. + */ + area = get_vm_area(size, VM_IOREMAP); + pr_debug("Get vm_area returns %p addr %p\n",area,area->addr); + if (!area) + return NULL; + area->phys_addr = phys_addr; + addr = area->addr; + if (remap_area_pages((unsigned long)addr, phys_addr, size, flags)) { + vunmap(addr); + return NULL; + } + return (void *) (offset + (char *)addr); +} + +void iounmap(void *addr) +{ + struct vm_struct *area; + + vfree((void *) (PAGE_MASK & (unsigned long) addr)); + area = remove_vm_area((void *) (PAGE_MASK & (unsigned long) addr)); + if (!area) { + printk(KERN_ERR "iounmap: bad address %p\n", addr); + return; + } + + kfree(area); +} + +static struct resource shmedia_iomap = { + .name = "shmedia_iomap", + .start = IOBASE_VADDR + PAGE_SIZE, + .end = IOBASE_END - 1, +}; + +static void shmedia_mapioaddr(unsigned long pa, unsigned long va); +static void shmedia_unmapioaddr(unsigned long vaddr); +static unsigned long shmedia_ioremap(struct resource *res, u32 pa, int sz); + +/* + * We have the same problem as the SPARC, so lets have the same comment: + * Our mini-allocator... + * Boy this is gross! We need it because we must map I/O for + * timers and interrupt controller before the kmalloc is available. + */ + +#define XNMLN 15 +#define XNRES 10 + +struct xresource { + struct resource xres; /* Must be first */ + int xflag; /* 1 == used */ + char xname[XNMLN+1]; +}; + +static struct xresource xresv[XNRES]; + +static struct xresource *xres_alloc(void) +{ + struct xresource *xrp; + int n; + + xrp = xresv; + for (n = 0; n < XNRES; n++) { + if (xrp->xflag == 0) { + xrp->xflag = 1; + return xrp; + } + xrp++; + } + return NULL; +} + +static void xres_free(struct xresource *xrp) +{ + xrp->xflag = 0; +} + +static struct resource *shmedia_find_resource(struct resource *root, + unsigned long vaddr) +{ + struct resource *res; + + for (res = root->child; res; res = res->sibling) + if (res->start <= vaddr && res->end >= vaddr) + return res; + + return NULL; +} + +static unsigned long shmedia_alloc_io(unsigned long phys, unsigned long size, + const char *name) +{ + static int printed_full = 0; + struct xresource *xres; + struct resource *res; + char *tack; + int tlen; + + if (name == NULL) name = "???"; + + if ((xres = xres_alloc()) != 0) { + tack = xres->xname; + res = &xres->xres; + } else { + if (!printed_full) { + printk("%s: done with statics, switching to kmalloc\n", + __FUNCTION__); + printed_full = 1; + } + tlen = strlen(name); + tack = kmalloc(sizeof (struct resource) + tlen + 1, GFP_KERNEL); + if (!tack) + return -ENOMEM; + memset(tack, 0, sizeof(struct resource)); + res = (struct resource *) tack; + tack += sizeof (struct resource); + } + + strncpy(tack, name, XNMLN); + tack[XNMLN] = 0; + res->name = tack; + + return shmedia_ioremap(res, phys, size); +} + +static unsigned long shmedia_ioremap(struct resource *res, u32 pa, int sz) +{ + unsigned long offset = ((unsigned long) pa) & (~PAGE_MASK); + unsigned long round_sz = (offset + sz + PAGE_SIZE-1) & PAGE_MASK; + unsigned long va; + unsigned int psz; + + if (allocate_resource(&shmedia_iomap, res, round_sz, + shmedia_iomap.start, shmedia_iomap.end, + PAGE_SIZE, NULL, NULL) != 0) { + panic("alloc_io_res(%s): cannot occupy\n", + (res->name != NULL)? res->name: "???"); + } + + va = res->start; + pa &= PAGE_MASK; + + psz = (res->end - res->start + (PAGE_SIZE - 1)) / PAGE_SIZE; + + /* log at boot time ... */ + printk("mapioaddr: %6s [%2d page%s] va 0x%08lx pa 0x%08x\n", + ((res->name != NULL) ? res->name : "???"), + psz, psz == 1 ? " " : "s", va, pa); + + for (psz = res->end - res->start + 1; psz != 0; psz -= PAGE_SIZE) { + shmedia_mapioaddr(pa, va); + va += PAGE_SIZE; + pa += PAGE_SIZE; + } + + res->start += offset; + res->end = res->start + sz - 1; /* not strictly necessary.. */ + + return res->start; +} + +static void shmedia_free_io(struct resource *res) +{ + unsigned long len = res->end - res->start + 1; + + BUG_ON((len & (PAGE_SIZE - 1)) != 0); + + while (len) { + len -= PAGE_SIZE; + shmedia_unmapioaddr(res->start + len); + } + + release_resource(res); +} + +static void *sh64_get_page(void) +{ + extern int after_bootmem; + void *page; + + if (after_bootmem) { + page = (void *)get_zeroed_page(GFP_ATOMIC); + } else { + page = alloc_bootmem_pages(PAGE_SIZE); + } + + if (!page || ((unsigned long)page & ~PAGE_MASK)) + panic("sh64_get_page: Out of memory already?\n"); + + return page; +} + +static void shmedia_mapioaddr(unsigned long pa, unsigned long va) +{ + pgd_t *pgdp; + pmd_t *pmdp; + pte_t *ptep, pte; + pgprot_t prot; + unsigned long flags = 1; /* 1 = CB0-1 device */ + + pr_debug("shmedia_mapiopage pa %08lx va %08lx\n", pa, va); + + pgdp = pgd_offset_k(va); + if (pgd_none(*pgdp) || !pgd_present(*pgdp)) { + pmdp = (pmd_t *)sh64_get_page(); + set_pgd(pgdp, __pgd((unsigned long)pmdp | _KERNPG_TABLE)); + } + + pmdp = pmd_offset(pgdp, va); + if (pmd_none(*pmdp) || !pmd_present(*pmdp) ) { + ptep = (pte_t *)sh64_get_page(); + set_pmd(pmdp, __pmd((unsigned long)ptep + _PAGE_TABLE)); + } + + prot = __pgprot(_PAGE_PRESENT | _PAGE_READ | _PAGE_WRITE | + _PAGE_DIRTY | _PAGE_ACCESSED | _PAGE_SHARED | flags); + + pte = pfn_pte(pa >> PAGE_SHIFT, prot); + ptep = pte_offset_kernel(pmdp, va); + + if (!pte_none(*ptep) && + pte_val(*ptep) != pte_val(pte)) + pte_ERROR(*ptep); + + set_pte(ptep, pte); + + flush_tlb_kernel_range(va, PAGE_SIZE); +} + +static void shmedia_unmapioaddr(unsigned long vaddr) +{ + pgd_t *pgdp; + pmd_t *pmdp; + pte_t *ptep; + + pgdp = pgd_offset_k(vaddr); + pmdp = pmd_offset(pgdp, vaddr); + + if (pmd_none(*pmdp) || pmd_bad(*pmdp)) + return; + + ptep = pte_offset_kernel(pmdp, vaddr); + + if (pte_none(*ptep) || !pte_present(*ptep)) + return; + + clear_page((void *)ptep); + pte_clear(&init_mm, vaddr, ptep); +} + +unsigned long onchip_remap(unsigned long phys, unsigned long size, const char *name) +{ + if (size < PAGE_SIZE) + size = PAGE_SIZE; + + return shmedia_alloc_io(phys, size, name); +} + +void onchip_unmap(unsigned long vaddr) +{ + struct resource *res; + unsigned int psz; + + res = shmedia_find_resource(&shmedia_iomap, vaddr); + if (!res) { + printk(KERN_ERR "%s: Failed to free 0x%08lx\n", + __FUNCTION__, vaddr); + return; + } + + psz = (res->end - res->start + (PAGE_SIZE - 1)) / PAGE_SIZE; + + printk(KERN_DEBUG "unmapioaddr: %6s [%2d page%s] freed\n", + res->name, psz, psz == 1 ? " " : "s"); + + shmedia_free_io(res); + + if ((char *)res >= (char *)xresv && + (char *)res < (char *)&xresv[XNRES]) { + xres_free((struct xresource *)res); + } else { + kfree(res); + } +} + +#ifdef CONFIG_PROC_FS +static int +ioremap_proc_info(char *buf, char **start, off_t fpos, int length, int *eof, + void *data) +{ + char *p = buf, *e = buf + length; + struct resource *r; + const char *nm; + + for (r = ((struct resource *)data)->child; r != NULL; r = r->sibling) { + if (p + 32 >= e) /* Better than nothing */ + break; + if ((nm = r->name) == 0) nm = "???"; + p += sprintf(p, "%08lx-%08lx: %s\n", r->start, r->end, nm); + } + + return p-buf; +} +#endif /* CONFIG_PROC_FS */ + +static int __init register_proc_onchip(void) +{ +#ifdef CONFIG_PROC_FS + create_proc_read_entry("io_map",0,0, ioremap_proc_info, &shmedia_iomap); +#endif + return 0; +} + +__initcall(register_proc_onchip); diff --git a/arch/sh64/mm/tlb.c b/arch/sh64/mm/tlb.c new file mode 100644 index 00000000000..d517e7d7034 --- /dev/null +++ b/arch/sh64/mm/tlb.c @@ -0,0 +1,166 @@ +/* + * arch/sh64/mm/tlb.c + * + * Copyright (C) 2003 Paul Mundt + * Copyright (C) 2003 Richard Curnow + * + * This file is subject to the terms and conditions of the GNU General Public + * License. See the file "COPYING" in the main directory of this archive + * for more details. + * + */ +#include +#include +#include +#include +#include + +/** + * sh64_tlb_init + * + * Perform initial setup for the DTLB and ITLB. + */ +int __init sh64_tlb_init(void) +{ + /* Assign some sane DTLB defaults */ + cpu_data->dtlb.entries = 64; + cpu_data->dtlb.step = 0x10; + + cpu_data->dtlb.first = DTLB_FIXED | cpu_data->dtlb.step; + cpu_data->dtlb.next = cpu_data->dtlb.first; + + cpu_data->dtlb.last = DTLB_FIXED | + ((cpu_data->dtlb.entries - 1) * + cpu_data->dtlb.step); + + /* And again for the ITLB */ + cpu_data->itlb.entries = 64; + cpu_data->itlb.step = 0x10; + + cpu_data->itlb.first = ITLB_FIXED | cpu_data->itlb.step; + cpu_data->itlb.next = cpu_data->itlb.first; + cpu_data->itlb.last = ITLB_FIXED | + ((cpu_data->itlb.entries - 1) * + cpu_data->itlb.step); + + return 0; +} + +/** + * sh64_next_free_dtlb_entry + * + * Find the next available DTLB entry + */ +unsigned long long sh64_next_free_dtlb_entry(void) +{ + return cpu_data->dtlb.next; +} + +/** + * sh64_get_wired_dtlb_entry + * + * Allocate a wired (locked-in) entry in the DTLB + */ +unsigned long long sh64_get_wired_dtlb_entry(void) +{ + unsigned long long entry = sh64_next_free_dtlb_entry(); + + cpu_data->dtlb.first += cpu_data->dtlb.step; + cpu_data->dtlb.next += cpu_data->dtlb.step; + + return entry; +} + +/** + * sh64_put_wired_dtlb_entry + * + * @entry: Address of TLB slot. + * + * Free a wired (locked-in) entry in the DTLB. + * + * Works like a stack, last one to allocate must be first one to free. + */ +int sh64_put_wired_dtlb_entry(unsigned long long entry) +{ + __flush_tlb_slot(entry); + + /* + * We don't do any particularly useful tracking of wired entries, + * so this approach works like a stack .. last one to be allocated + * has to be the first one to be freed. + * + * We could potentially load wired entries into a list and work on + * rebalancing the list periodically (which also entails moving the + * contents of a TLB entry) .. though I have a feeling that this is + * more trouble than it's worth. + */ + + /* + * Entry must be valid .. we don't want any ITLB addresses! + */ + if (entry <= DTLB_FIXED) + return -EINVAL; + + /* + * Next, check if we're within range to be freed. (ie, must be the + * entry beneath the first 'free' entry! + */ + if (entry < (cpu_data->dtlb.first - cpu_data->dtlb.step)) + return -EINVAL; + + /* If we are, then bring this entry back into the list */ + cpu_data->dtlb.first -= cpu_data->dtlb.step; + cpu_data->dtlb.next = entry; + + return 0; +} + +/** + * sh64_setup_tlb_slot + * + * @config_addr: Address of TLB slot. + * @eaddr: Virtual address. + * @asid: Address Space Identifier. + * @paddr: Physical address. + * + * Load up a virtual<->physical translation for @eaddr<->@paddr in the + * pre-allocated TLB slot @config_addr (see sh64_get_wired_dtlb_entry). + */ +inline void sh64_setup_tlb_slot(unsigned long long config_addr, + unsigned long eaddr, + unsigned long asid, + unsigned long paddr) +{ + unsigned long long pteh, ptel; + + /* Sign extension */ +#if (NEFF == 32) + pteh = (unsigned long long)(signed long long)(signed long) eaddr; +#else +#error "Can't sign extend more than 32 bits yet" +#endif + pteh &= PAGE_MASK; + pteh |= (asid << PTEH_ASID_SHIFT) | PTEH_VALID; +#if (NEFF == 32) + ptel = (unsigned long long)(signed long long)(signed long) paddr; +#else +#error "Can't sign extend more than 32 bits yet" +#endif + ptel &= PAGE_MASK; + ptel |= (_PAGE_CACHABLE | _PAGE_READ | _PAGE_WRITE); + + asm volatile("putcfg %0, 1, %1\n\t" + "putcfg %0, 0, %2\n" + : : "r" (config_addr), "r" (ptel), "r" (pteh)); +} + +/** + * sh64_teardown_tlb_slot + * + * @config_addr: Address of TLB slot. + * + * Teardown any existing mapping in the TLB slot @config_addr. + */ +inline void sh64_teardown_tlb_slot(unsigned long long config_addr) + __attribute__ ((alias("__flush_tlb_slot"))); + diff --git a/arch/sh64/mm/tlbmiss.c b/arch/sh64/mm/tlbmiss.c new file mode 100644 index 00000000000..c8615954aaa --- /dev/null +++ b/arch/sh64/mm/tlbmiss.c @@ -0,0 +1,280 @@ +/* + * This file is subject to the terms and conditions of the GNU General Public + * License. See the file "COPYING" in the main directory of this archive + * for more details. + * + * arch/sh64/mm/tlbmiss.c + * + * Original code from fault.c + * Copyright (C) 2000, 2001 Paolo Alberelli + * + * Fast PTE->TLB refill path + * Copyright (C) 2003 Richard.Curnow@superh.com + * + * IMPORTANT NOTES : + * The do_fast_page_fault function is called from a context in entry.S where very few registers + * have been saved. In particular, the code in this file must be compiled not to use ANY + * caller-save regiseters that are not part of the restricted save set. Also, it means that + * code in this file must not make calls to functions elsewhere in the kernel, or else the + * excepting context will see corruption in its caller-save registers. Plus, the entry.S save + * area is non-reentrant, so this code has to run with SR.BL==1, i.e. no interrupts taken inside + * it and panic on any exception. + * + */ + +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include + +#include +#include +#include +#include +#include +#include +#include /* required by inline asm statements */ + +/* Callable from fault.c, so not static */ +inline void __do_tlb_refill(unsigned long address, + unsigned long long is_text_not_data, pte_t *pte) +{ + unsigned long long ptel; + unsigned long long pteh=0; + struct tlb_info *tlbp; + unsigned long long next; + + /* Get PTEL first */ + ptel = pte_val(*pte); + + /* + * Set PTEH register + */ + pteh = address & MMU_VPN_MASK; + + /* Sign extend based on neff. */ +#if (NEFF == 32) + /* Faster sign extension */ + pteh = (unsigned long long)(signed long long)(signed long)pteh; +#else + /* General case */ + pteh = (pteh & NEFF_SIGN) ? (pteh | NEFF_MASK) : pteh; +#endif + + /* Set the ASID. */ + pteh |= get_asid() << PTEH_ASID_SHIFT; + pteh |= PTEH_VALID; + + /* Set PTEL register, set_pte has performed the sign extension */ + ptel &= _PAGE_FLAGS_HARDWARE_MASK; /* drop software flags */ + + tlbp = is_text_not_data ? &(cpu_data->itlb) : &(cpu_data->dtlb); + next = tlbp->next; + __flush_tlb_slot(next); + asm volatile ("putcfg %0,1,%2\n\n\t" + "putcfg %0,0,%1\n" + : : "r" (next), "r" (pteh), "r" (ptel) ); + + next += TLB_STEP; + if (next > tlbp->last) next = tlbp->first; + tlbp->next = next; + +} + +static int handle_vmalloc_fault(struct mm_struct *mm, unsigned long protection_flags, + unsigned long long textaccess, + unsigned long address) +{ + pgd_t *dir; + pmd_t *pmd; + static pte_t *pte; + pte_t entry; + + dir = pgd_offset_k(address); + pmd = pmd_offset(dir, address); + + if (pmd_none(*pmd)) { + return 0; + } + + if (pmd_bad(*pmd)) { + pmd_clear(pmd); + return 0; + } + + pte = pte_offset_kernel(pmd, address); + entry = *pte; + + if (pte_none(entry) || !pte_present(entry)) { + return 0; + } + + if ((pte_val(entry) & protection_flags) != protection_flags) { + return 0; + } + + __do_tlb_refill(address, textaccess, pte); + + return 1; +} + +static int handle_tlbmiss(struct mm_struct *mm, unsigned long long protection_flags, + unsigned long long textaccess, + unsigned long address) +{ + pgd_t *dir; + pmd_t *pmd; + pte_t *pte; + pte_t entry; + + /* NB. The PGD currently only contains a single entry - there is no + page table tree stored for the top half of the address space since + virtual pages in that region should never be mapped in user mode. + (In kernel mode, the only things in that region are the 512Mb super + page (locked in), and vmalloc (modules) + I/O device pages (handled + by handle_vmalloc_fault), so no PGD for the upper half is required + by kernel mode either). + + See how mm->pgd is allocated and initialised in pgd_alloc to see why + the next test is necessary. - RPC */ + if (address >= (unsigned long) TASK_SIZE) { + /* upper half - never has page table entries. */ + return 0; + } + dir = pgd_offset(mm, address); + if (pgd_none(*dir)) { + return 0; + } + if (!pgd_present(*dir)) { + return 0; + } + + pmd = pmd_offset(dir, address); + if (pmd_none(*pmd)) { + return 0; + } + if (!pmd_present(*pmd)) { + return 0; + } + pte = pte_offset_kernel(pmd, address); + entry = *pte; + if (pte_none(entry)) { + return 0; + } + if (!pte_present(entry)) { + return 0; + } + + /* If the page doesn't have sufficient protection bits set to service the + kind of fault being handled, there's not much point doing the TLB refill. + Punt the fault to the general handler. */ + if ((pte_val(entry) & protection_flags) != protection_flags) { + return 0; + } + + __do_tlb_refill(address, textaccess, pte); + + return 1; +} + +/* Put all this information into one structure so that everything is just arithmetic + relative to a single base address. This reduces the number of movi/shori pairs needed + just to load addresses of static data. */ +struct expevt_lookup { + unsigned short protection_flags[8]; + unsigned char is_text_access[8]; + unsigned char is_write_access[8]; +}; + +#define PRU (1<<9) +#define PRW (1<<8) +#define PRX (1<<7) +#define PRR (1<<6) + +#define DIRTY (_PAGE_DIRTY | _PAGE_ACCESSED) +#define YOUNG (_PAGE_ACCESSED) + +/* Sized as 8 rather than 4 to allow checking the PTE's PRU bit against whether + the fault happened in user mode or privileged mode. */ +static struct expevt_lookup expevt_lookup_table = { + .protection_flags = {PRX, PRX, 0, 0, PRR, PRR, PRW, PRW}, + .is_text_access = {1, 1, 0, 0, 0, 0, 0, 0} +}; + +/* + This routine handles page faults that can be serviced just by refilling a + TLB entry from an existing page table entry. (This case represents a very + large majority of page faults.) Return 1 if the fault was successfully + handled. Return 0 if the fault could not be handled. (This leads into the + general fault handling in fault.c which deals with mapping file-backed + pages, stack growth, segmentation faults, swapping etc etc) + */ +asmlinkage int do_fast_page_fault(unsigned long long ssr_md, unsigned long long expevt, + unsigned long address) +{ + struct task_struct *tsk; + struct mm_struct *mm; + unsigned long long textaccess; + unsigned long long protection_flags; + unsigned long long index; + unsigned long long expevt4; + + /* The next few lines implement a way of hashing EXPEVT into a small array index + which can be used to lookup parameters specific to the type of TLBMISS being + handled. Note: + ITLBMISS has EXPEVT==0xa40 + RTLBMISS has EXPEVT==0x040 + WTLBMISS has EXPEVT==0x060 + */ + + expevt4 = (expevt >> 4); + /* TODO : xor ssr_md into this expression too. Then we can check that PRU is set + when it needs to be. */ + index = expevt4 ^ (expevt4 >> 5); + index &= 7; + protection_flags = expevt_lookup_table.protection_flags[index]; + textaccess = expevt_lookup_table.is_text_access[index]; + +#ifdef CONFIG_SH64_PROC_TLB + ++calls_to_do_fast_page_fault; +#endif + + /* SIM + * Note this is now called with interrupts still disabled + * This is to cope with being called for a missing IO port + * address with interupts disabled. This should be fixed as + * soon as we have a better 'fast path' miss handler. + * + * Plus take care how you try and debug this stuff. + * For example, writing debug data to a port which you + * have just faulted on is not going to work. + */ + + tsk = current; + mm = tsk->mm; + + if ((address >= VMALLOC_START && address < VMALLOC_END) || + (address >= IOBASE_VADDR && address < IOBASE_END)) { + if (ssr_md) { + /* Process-contexts can never have this address range mapped */ + if (handle_vmalloc_fault(mm, protection_flags, textaccess, address)) { + return 1; + } + } + } else if (!in_interrupt() && mm) { + if (handle_tlbmiss(mm, protection_flags, textaccess, address)) { + return 1; + } + } + + return 0; +} + -- cgit v1.2.3