• Jeff Dike's avatar
    [PATCH] uml: skas0 - separate kernel address space on stock hosts · d67b569f
    Jeff Dike authored
    UML has had two modes of operation - an insecure, slow mode (tt mode) in
    which the kernel is mapped into every process address space which requires
    no host kernel modifications, and a secure, faster mode (skas mode) in
    which the UML kernel is in a separate host address space, which requires a
    patch to the host kernel.
    
    This patch implements something very close to skas mode for hosts which
    don't support skas - I'm calling this skas0.  It provides the security of
    the skas host patch, and some of the performance gains.
    
    The two main things that are provided by the skas patch, /proc/mm and
    PTRACE_FAULTINFO, are implemented in a way that require no host patch.
    
    For the remote address space changing stuff (mmap, munmap, and mprotect),
    we set aside two pages in the process above its stack, one of which
    contains a little bit of code which can call mmap et al.
    
    To update the address space, the system call information (system call
    number and arguments) are written to the stub page above the code.  The
    %esp is set to the beginning of the data, the %eip is set the the start of
    the stub, and it repeatedly pops the information into its registers and
    makes the system call until it sees a system call number of zero.  This is
    to amortize the cost of the context switch across multiple address space
    updates.
    
    When the updates are done, it SIGSTOPs itself, and the kernel process
    continues what it was doing.
    
    For a PTRACE_FAULTINFO replacement, we set up a SIGSEGV handler in the
    child, and let it handle segfaults rather than nullifying them.  The
    handler is in the same page as the mmap stub.  The second page is used as
    the stack.  The handler reads cr2 and err from the sigcontext, sticks them
    at the base of the stack in a faultinfo struct, and SIGSTOPs itself.  The
    kernel then reads the faultinfo and handles the fault.
    
    A complication on x86_64 is that this involves resetting the registers to
    the segfault values when the process is inside the kill system call.  This
    breaks on x86_64 because %rcx will contain %rip because you tell SYSRET
    where to return to by putting the value in %rcx.  So, this corrupts $rcx on
    return from the segfault.  To work around this, I added an
    arch_finish_segv, which on x86 does nothing, but which on x86_64 ptraces
    the child back through the sigreturn.  This causes %rcx to be restored by
    sigreturn and avoids the corruption.  Ultimately, I think I will replace
    this with the trick of having it send itself a blocked signal which will be
    unblocked by the sigreturn.  This will allow it to be stopped just after
    the sigreturn, and PTRACE_SYSCALLed without all the back-and-forth of
    PTRACE_SYSCALLing it through sigreturn.
    
    This runs on a stock host, so theoretically (and hopefully), tt mode isn't
    needed any more.  We need to make sure that this is better in every way
    than tt mode, though.  I'm concerned about the speed of address space
    updates and page fault handling, since they involve extra round-trips to
    the child.  We can amortize the round-trip cost for large address space
    updates by writing all of the operations to the data page and having the
    child execute them all at the same time.  This will help fork and exec, but
    not page faults, since they involve only one page.
    
    I can't think of any way to help page faults, except to add something like
    PTRACE_FAULTINFO to the host.  There is PTRACE_SIGINFO, but UML doesn't use
    siginfo for SIGSEGV (or anything else) because there isn't enough
    information in the siginfo struct to handle page faults (the faulting
    operation type is missing).  Adding that would make PTRACE_SIGINFO a usable
    equivalent to PTRACE_FAULTINFO.
    
    As for the code itself:
    
    - The system call stub is in arch/um/kernel/sys-$(SUBARCH)/stub.S.  It is
      put in its own section of the binary along with stub_segv_handler in
      arch/um/kernel/skas/process.c.  This is manipulated with run_syscall_stub
      in arch/um/kernel/skas/mem_user.c.  syscall_stub will execute any system
      call at all, but it's only used for mmap, munmap, and mprotect.
    
    - The x86_64 stub calls sigreturn by hand rather than allowing the normal
      sigreturn to happen, because the normal sigreturn is a SA_RESTORER in
      UML's address space provided by libc.  Needless to say, this is not
      available in the child's address space.  Also, it does a couple of odd
      pops before that which restore the stack to the state it was in at the
      time the signal handler was called.
    
    - There is a new field in the arch mmu_context, which is now a union.
      This is the pid to be manipulated rather than the /proc/mm file
      descriptor.  Code which deals with this now checks proc_mm to see whether
      it should use the usual skas code or the new code.
    
    - userspace_tramp is now used to create a new host process for every UML
      process, rather than one per UML processor.  It checks proc_mm and
      ptrace_faultinfo to decide whether to map in the pages above its stack.
    
    - start_userspace now makes CLONE_VM conditional on proc_mm since we need
      separate address spaces now.
    
    - switch_mm_skas now just sets userspace_pid[0] to the new pid rather
      than PTRACE_SWITCH_MM.  There is an addition to userspace which updates
      its idea of the pid being manipulated each time around the loop.  This is
      important on exec, when the pid will change underneath userspace().
    
    - The stub page has a pte, but it can't be mapped in using tlb_flush
      because it is part of tlb_flush.  This is why it's required for it to be
      mapped in by userspace_tramp.
    
    Other random things:
    
    - The stub section in uml.lds.S is page aligned.  This page is written
      out to the backing vm file in setup_physmem because it is mapped from
      there into user processes.
    
    - There's some confusion with TASK_SIZE now that there are a couple of
      extra pages that the process can't use.  TASK_SIZE is considered by the
      elf code to be the usable process memory, which is reasonable, so it is
      decreased by two pages.  This confuses the definition of
      USER_PGDS_IN_LAST_PML4, making it too small because of the rounding down
      of the uneven division.  So we round it to the nearest PGDIR_SIZE rather
      than the lower one.
    
    - I added a missing PT_SYSCALL_ARG6_OFFSET macro.
    
    - um_mmu.h was made into a userspace-usable file.
    
    - proc_mm and ptrace_faultinfo are globals which say whether the host
      supports these features.
    
    - There is a bad interaction between the mm.nr_ptes check at the end of
      exit_mmap, stack randomization, and skas0.  exit_mmap will stop freeing
      pages at the PGDIR_SIZE boundary after the last vma.  If the stack isn't
      on the last page table page, the last pte page won't be freed, as it
      should be since the stub ptes are there, and exit_mmap will BUG because
      there is an unfreed page.  To get around this, TASK_SIZE is set to the
      next lowest PGDIR_SIZE boundary and mm->nr_ptes is decremented after the
      calls to init_stub_pte.  This ensures that we know the process stack (and
      all other process mappings) will be below the top page table page, and
      thus we know that mm->nr_ptes will be one too many, and can be
      decremented.
    
    Things that need fixing:
    
    - We may need better assurrences that the stub code is PIC.
    
    - The stub pte is set up in init_new_context_skas.
    
    - alloc_pgdir is probably the right place.
    Signed-off-by: default avatarJeff Dike <jdike@addtoit.com>
    Signed-off-by: default avatarAndrew Morton <akpm@osdl.org>
    Signed-off-by: default avatarLinus Torvalds <torvalds@osdl.org>
    d67b569f
dyn.lds.S 6.18 KB