• Zachary Amsden's avatar
    [PATCH] x86/PAE: Fix pte_clear for the >4GB RAM case · 6e5882cf
    Zachary Amsden authored
    Proposed fix for ptep_get_and_clear_full PAE bug.  Pte_clear had the same bug,
    so use the same fix for both.  Turns out pmd_clear had it as well, but pgds
    are not affected.
    
    The problem is rather intricate.  Page table entries in PAE mode are 64-bits
    wide, but the only atomic 8-byte write operation available in 32-bit mode is
    cmpxchg8b, which is expensive (at least on P4), and thus avoided.  But it can
    happen that the processor may prefetch entries into the TLB in the middle of an
    operation which clears a page table entry.  So one must always clear the P-bit
    in the low word of the page table entry first when clearing it.
    
    Since the sequence *ptep = __pte(0) leaves the order of the write dependent on
    the compiler, it must be coded explicitly as a clear of the low word followed
    by a clear of the high word.  Further, there must be a write memory barrier
    here to enforce proper ordering by the compiler (and, in the future, by the
    processor as well).
    
    On > 4GB memory machines, the implementation of pte_clear for PAE was clearly
    deficient, as it could leave virtual mappings of physical memory above 4GB
    aliased to memory below 4GB in the TLB.  The implementation of
    ptep_get_and_clear_full has a similar bug, although not nearly as likely to
    occur, since the mappings being cleared are in the process of being destroyed,
    and should never be dereferenced again.
    
    But, as luck would have it, it is possible to trigger bugs even without ever
    dereferencing these bogus TLB mappings, even if the clear is followed fairly
    soon after with a TLB flush or invalidation.  The problem is that memory above
    4GB may now be aliased into the first 4GB of memory, and in fact, may hit a
    region of memory with non-memory semantics.  These regions include AGP and PCI
    space.  As such, these memory regions are not cached by the processor.  This
    introduces the bug.
    
    The processor can speculate memory operations, including memory writes, as long
    as they are committed with the proper ordering.  Speculating a memory write to
    a linear address that has a bogus TLB mapping is possible.  Normally, the
    speculation is harmless.  But for cached memory, it does leave the falsely
    speculated cacheline unmodified, but in a dirty state.  This cache line will be
    eventually written back.  If this cacheline happens to intersect a region of
    memory that is not protected by the cache coherency protocol, it can corrupt
    data in I/O memory, which is generally a very bad thing to do, and can cause
    total system failure or just plain undefined behavior.
    
    These bugs are extremely unlikely, but the severity is of such magnitude, and
    the fix so simple that I think fixing them immediately is justified.  Also,
    they are nearly impossible to debug.
    Signed-off-by: default avatarZachary Amsden <zach@vmware.com>
    Signed-off-by: default avatarLinus Torvalds <torvalds@osdl.org>
    6e5882cf
pgtable-3level.h 4.6 KB