Cache and TLB Flushing
                     Under Linux

            David S. Miller <davem@redhat.com>

This document describes the cache/tlb flushing interfaces called
by the Linux VM subsystem.  It enumerates over each interface,
describes it's intended purpose, and what side effect is expected
after the interface is invoked.

The side effects described below are stated for a uniprocessor
implementation, and what is to happen on that single processor.  The
SMP cases are a simple extension, in that you just extend the
definition such that the side effect for a particular interface occurs
on all processors in the system.  Don't let this scare you into
thinking SMP cache/tlb flushing must be so inefficient, this is in
fact an area where many optimizations are possible.  For example,
if it can be proven that a user address space has never executed
on a cpu (see vma->cpu_vm_mask), one need not perform a flush
for this address space on that cpu.

First, the TLB flushing interfaces, since they are the simplest.  The
"TLB" is abstracted under Linux as something the cpu uses to cache
virtual-->physical address translations obtained from the software
page tables.  Meaning that if the software page tables change, it is
possible for stale translations to exist in this "TLB" cache.
Therefore when software page table changes occur, the kernel will
invoke one of the following flush methods _after_ the page table
changes occur:

1) void flush_tlb_all(void)

        The most severe flush of all.  After this interface runs,
        any previous page table modification whatsoever will be
        visible to the cpu.

        This is usually invoked when the kernel page tables are
        changed, since such translations are "global" in nature.

2) void flush_tlb_mm(struct mm_struct *mm)

        This interface flushes an entire user address space from
        the TLB.  After running, this interface must make sure that
        any previous page table modifications for the address space
        'mm' will be visible to the cpu.  That is, after running,
        there will be no entries in the TLB for 'mm'.

        This interface is used to handle whole address space
        page table operations such as what happens during
        fork, and exec.

3) void flush_tlb_range(struct mm_struct *mm,
                        unsigned long start, unsigned long end)

        Here we are flushing a specific range of (user) virtual
        address translations from the TLB.  After running, this
        interface must make sure that any previous page table
        modifications for the address space 'mm' in the range 'start'
        to 'end' will be visible to the cpu.  That is, after running,
        there will be no entries in the TLB for 'mm' for virtual
        addresses in the range 'start' to 'end'.

        Primarily, this is used for munmap() type operations.

        The interface is provided in hopes that the port can find
        a suitably efficient method for removing multiple page
        sized translations from the TLB, instead of having the kernel
        call flush_tlb_page (see below) for each entry which may be
        modified.

4) void flush_tlb_page(struct vm_area_struct *vma, unsigned long page)

        This time we need to remove the PAGE_SIZE sized translation
        from the TLB.  The 'vma' is the backing structure used by
        Linux to keep track of mmap'd regions for a process, the
        address space is available via vma->vm_mm.  Also, one may
        test (vma->vm_flags & VM_EXEC) to see if this region is
        executable (and thus could be in the 'instruction TLB' in
        split-tlb type setups).

        After running, this interface must make sure that any previous
        page table modification for address space 'vma->vm_mm' for
        user virtual address 'page' will be visible to the cpu.  That
        is, after running, there will be no entries in the TLB for
        'vma->vm_mm' for virtual address 'page'.

        This is used primarily during fault processing.

5) void flush_tlb_pgtables(struct mm_struct *mm,
                           unsigned long start, unsigned long end)

   The software page tables for address space 'mm' for virtual
   addresses in the range 'start' to 'end' are being torn down.

   Some platforms cache the lowest level of the software page tables
   in a linear virtually mapped array, to make TLB miss processing
   more efficient.  On such platforms, since the TLB is caching the
   software page table structure, it needs to be flushed when parts
   of the software page table tree are unlinked/freed.

   Sparc64 is one example of a platform which does this.

   Usually, when munmap()'ing an area of user virtual address
   space, the kernel leaves the page table parts around and just
   marks the individual pte's as invalid.  However, if very large
   portions of the address space are unmapped, the kernel frees up
   those portions of the software page tables to prevent potential
   excessive kernel memory usage caused by erratic mmap/mmunmap
   sequences.  It is at these times that flush_tlb_pgtables will
   be invoked.

6) void update_mmu_cache(struct vm_area_struct *vma,
                         unsigned long address, pte_t pte)

        At the end of every page fault, this routine is invoked to
        tell the architecture specific code that a translation
        described by "pte" now exists at virtual address "address"
        for address space "vma->vm_mm", in the software page tables.

        A port may use this information in any way it so chooses.
        For example, it could use this event to pre-load TLB
        translations for software managed TLB configurations.
        The sparc64 port currently does this.

Next, we have the cache flushing interfaces.  In general, when Linux
is changing an existing virtual-->physical mapping to a new value,
the sequence will be in one of the following forms:

        1) flush_cache_mm(mm);
           change_all_page_tables_of(mm);
           flush_tlb_mm(mm);

        2) flush_cache_range(mm, start, end);
           change_range_of_page_tables(mm, start, end);
           flush_tlb_range(mm, start, end);

        3) flush_cache_page(vma, page);
           set_pte(pte_pointer, new_pte_val);
           flush_tlb_page(vma, page);

The cache level flush will always be first, because this allows
us to properly handle systems whose caches are strict and require
a virtual-->physical translation to exist for a virtual address
when that virtual address is flushed from the cache.  The HyperSparc
cpu is one such cpu with this attribute.

The cache flushing routines below need only deal with cache flushing
to the extent that it is necessary for a particular cpu.  Mostly,
these routines must be implemented for cpus which have virtually
indexed caches which must be flushed when virtual-->physical
translations are changed or removed.  So, for example, the physically
indexed physically tagged caches of IA32 processors have no need to
implement these interfaces since the caches are fully synchronized
and have no dependency on translation information.

Here are the routines, one by one:

1) void flush_cache_all(void)

        The most severe flush of all.  After this interface runs,
        the entire cpu cache is flushed.

        This is usually invoked when the kernel page tables are
        changed, since such translations are "global" in nature.

2) void flush_cache_mm(struct mm_struct *mm)

        This interface flushes an entire user address space from
        the caches.  That is, after running, there will be no cache
        lines associated with 'mm'.

        This interface is used to handle whole address space
        page table operations such as what happens during
        fork, exit, and exec.

3) void flush_cache_range(struct mm_struct *mm,
                          unsigned long start, unsigned long end)

        Here we are flushing a specific range of (user) virtual
        addresses from the cache.  After running, there will be no
        entries in the cache for 'mm' for virtual addresses in the
        range 'start' to 'end'.

        Primarily, this is used for munmap() type operations.

        The interface is provided in hopes that the port can find
        a suitably efficient method for removing multiple page
        sized regions from the cache, instead of having the kernel
        call flush_cache_page (see below) for each entry which may be
        modified.

4) void flush_cache_page(struct vm_area_struct *vma, unsigned long page)

        This time we need to remove a PAGE_SIZE sized range
        from the cache.  The 'vma' is the backing structure used by
        Linux to keep track of mmap'd regions for a process, the
        address space is available via vma->vm_mm.  Also, one may
        test (vma->vm_flags & VM_EXEC) to see if this region is
        executable (and thus could be in the 'instruction cache' in
        "Harvard" type cache layouts).

        After running, there will be no entries in the cache for
        'vma->vm_mm' for virtual address 'page'.

        This is used primarily during fault processing.

There exists another whole class of cpu cache issues which currently
require a whole different set of interfaces to handle properly.
The biggest problem is that of virtual aliasing in the data cache
of a processor.

Is your port susceptible to virtual aliasing in it's D-cache?
Well, if your D-cache is virtually indexed, is larger in size than
PAGE_SIZE, and does not prevent multiple cache lines for the same
physical address from existing at once, you have this problem.

If your D-cache has this problem, first define asm/shmparam.h SHMLBA
properly, it should essentially be the size of your virtually
addressed D-cache (or if the size is variable, the largest possible
size).  This setting will force the SYSv IPC layer to only allow user
processes to mmap shared memory at address which are a multiple of
this value.

NOTE: This does not fix shared mmaps, check out the sparc64 port for
one way to solve this (in particular SPARC_FLAG_MMAPSHARED).

Next, you have two methods to solve the D-cache aliasing issue for all
other cases.  Please keep in mind that fact that, for a given page
mapped into some user address space, there is always at least one more
mapping, that of the kernel in it's linear mapping starting at
PAGE_OFFSET.  So immediately, once the first user maps a given
physical page into its address space, by implication the D-cache
aliasing problem has the potential to exist since the kernel already
maps this page at its virtual address.

First, I describe the old method to deal with this problem.  I am
describing it for documentation purposes, but it is deprecated and the
latter method I describe next should be used by all new ports and all
existing ports should move over to the new mechanism as well.

  flush_page_to_ram(struct page *page)

        The physical page 'page' is about to be place into the
        user address space of a process.  If it is possible for
        stores done recently by the kernel into this physical
        page, to not be visible to an arbitrary mapping in userspace,
        you must flush this page from the D-cache.

        If the D-cache is writeback in nature, the dirty data (if
        any) for this physical page must be written back to main
        memory before the cache lines are invalidated.

Admittedly, the author did not think very much when designing this
interface.  It does not give the architecture enough information about
what exactly is going on, and there is no context to base a judgment
on about whether an alias is possible at all.  The new interfaces to
deal with D-cache aliasing are meant to address this by telling the
architecture specific code exactly which is going on at the proper points
in time.

Here is the new interface:

  void copy_user_page(void *to, void *from, unsigned long address)
  void clear_user_page(void *to, unsigned long address)

        These two routines store data in user anonymous or COW
        pages.  It allows a port to efficiently avoid D-cache alias
        issues between userspace and the kernel.

        For example, a port may temporarily map 'from' and 'to' to
        kernel virtual addresses during the copy.  The virtual address
        for these two pages is chosen in such a way that the kernel
        load/store instructions happen to virtual addresses which are
        of the same "color" as the user mapping of the page.  Sparc64
        for example, uses this technique.

        The "address" parameter tells the virtual address where the
        user will ultimately have this page mapped.

        If D-cache aliasing is not an issue, these two routines may
        simply call memcpy/memset directly and do nothing more.

  void flush_dcache_page(struct page *page)

        Any time the kernel writes to a page cache page, _OR_
        the kernel is about to read from a page cache page and
        user space shared/writable mappings of this page potentially
        exist, this routine is called.

        NOTE: This routine need only be called for page cache pages
              which can potentially ever be mapped into the address
              space of a user process.  So for example, VFS layer code
              handling vfs symlinks in the page cache need not call
              this interface at all.

        The phrase "kernel writes to a page cache page" means,
        specifically, that the kernel executes store instructions
        that dirty data in that page at the page->virtual mapping
        of that page.  It is important to flush here to handle
        D-cache aliasing, to make sure these kernel stores are
        visible to user space mappings of that page.

        The corollary case is just as important, if there are users
        which have shared+writable mappings of this file, we must make
        sure that kernel reads of these pages will see the most recent
        stores done by the user.

        If D-cache aliasing is not an issue, this routine may
        simply be defined as a nop on that architecture.

        There is a bit set aside in page->flags (PG_arch_1) as
        "architecture private".  The kernel guarantees that,
        for pagecache pages, it will clear this bit when such
        a page first enters the pagecache.

        This allows these interfaces to be implemented much more
        efficiently.  It allows one to "defer" (perhaps indefinitely)
        the actual flush if there are currently no user processes
        mapping this page.  See sparc64's flush_dcache_page and
        update_mmu_cache implementations for an example of how to go
        about doing this.

        The idea is, first at flush_dcache_page() time, if
        page->mapping->i_mmap{,_shared} are empty lists, just mark the
        architecture private page flag bit.  Later, in
        update_mmu_cache(), a check is made of this flag bit, and if
        set the flush is done and the flag bit is cleared.

        IMPORTANT NOTE: It is often important, if you defer the flush,
                        that the actual flush occurs on the same CPU
                        as did the cpu stores into the page to make it
                        dirty.  Again, see sparc64 for examples of how
                        to deal with this.

  void flush_icache_range(unsigned long start, unsigned long end)
        When the kernel stores into addresses that it will execute
        out of (eg when loading modules), this function is called.

        If the icache does not snoop stores then this routine will need
        to flush it.

  void flush_icache_page(struct vm_area_struct *vma, struct page *page)
        All the functionality of flush_icache_page can be implemented in
        flush_dcache_page and update_mmu_cache. In 2.5 the hope is to
        remove this interface completely.