1. 18 Feb, 2012 2 commits
    • Linus Torvalds's avatar
      i387: re-introduce FPU state preloading at context switch time · 34ddc81a
      Linus Torvalds authored
      After all the FPU state cleanups and finally finding the problem that
      caused all our FPU save/restore problems, this re-introduces the
      preloading of FPU state that was removed in commit b3b0870e ("i387:
      do not preload FPU state at task switch time").
      
      However, instead of simply reverting the removal, this reimplements
      preloading with several fixes, most notably
      
       - properly abstracted as a true FPU state switch, rather than as
         open-coded save and restore with various hacks.
      
         In particular, implementing it as a proper FPU state switch allows us
         to optimize the CR0.TS flag accesses: there is no reason to set the
         TS bit only to then almost immediately clear it again.  CR0 accesses
         are quite slow and expensive, don't flip the bit back and forth for
         no good reason.
      
       - Make sure that the same model works for both x86-32 and x86-64, so
         that there are no gratuitous differences between the two due to the
         way they save and restore segment state differently due to
         architectural differences that really don't matter to the FPU state.
      
       - Avoid exposing the "preload" state to the context switch routines,
         and in particular allow the concept of lazy state restore: if nothing
         else has used the FPU in the meantime, and the process is still on
         the same CPU, we can avoid restoring state from memory entirely, just
         re-expose the state that is still in the FPU unit.
      
         That optimized lazy restore isn't actually implemented here, but the
         infrastructure is set up for it.  Of course, older CPU's that use
         'fnsave' to save the state cannot take advantage of this, since the
         state saving also trashes the state.
      
      In other words, there is now an actual _design_ to the FPU state saving,
      rather than just random historical baggage.  Hopefully it's easier to
      follow as a result.
      Signed-off-by: default avatarLinus Torvalds <torvalds@linux-foundation.org>
      34ddc81a
    • Linus Torvalds's avatar
      i387: move TS_USEDFPU flag from thread_info to task_struct · f94edacf
      Linus Torvalds authored
      This moves the bit that indicates whether a thread has ownership of the
      FPU from the TS_USEDFPU bit in thread_info->status to a word of its own
      (called 'has_fpu') in task_struct->thread.has_fpu.
      
      This fixes two independent bugs at the same time:
      
       - changing 'thread_info->status' from the scheduler causes nasty
         problems for the other users of that variable, since it is defined to
         be thread-synchronous (that's what the "TS_" part of the naming was
         supposed to indicate).
      
         So perfectly valid code could (and did) do
      
      	ti->status |= TS_RESTORE_SIGMASK;
      
         and the compiler was free to do that as separate load, or and store
         instructions.  Which can cause problems with preemption, since a task
         switch could happen in between, and change the TS_USEDFPU bit. The
         change to TS_USEDFPU would be overwritten by the final store.
      
         In practice, this seldom happened, though, because the 'status' field
         was seldom used more than once, so gcc would generally tend to
         generate code that used a read-modify-write instruction and thus
         happened to avoid this problem - RMW instructions are naturally low
         fat and preemption-safe.
      
       - On x86-32, the current_thread_info() pointer would, during interrupts
         and softirqs, point to a *copy* of the real thread_info, because
         x86-32 uses %esp to calculate the thread_info address, and thus the
         separate irq (and softirq) stacks would cause these kinds of odd
         thread_info copy aliases.
      
         This is normally not a problem, since interrupts aren't supposed to
         look at thread information anyway (what thread is running at
         interrupt time really isn't very well-defined), but it confused the
         heck out of irq_fpu_usable() and the code that tried to squirrel
         away the FPU state.
      
         (It also caused untold confusion for us poor kernel developers).
      
      It also turns out that using 'task_struct' is actually much more natural
      for most of the call sites that care about the FPU state, since they
      tend to work with the task struct for other reasons anyway (ie
      scheduling).  And the FPU data that we are going to save/restore is
      found there too.
      
      Thanks to Arjan Van De Ven <arjan@linux.intel.com> for pointing us to
      the %esp issue.
      
      Cc: Arjan van de Ven <arjan@linux.intel.com>
      Reported-and-tested-by: default avatarRaphael Prevost <raphael@buro.asia>
      Acked-and-tested-by: default avatarSuresh Siddha <suresh.b.siddha@intel.com>
      Tested-by: default avatarPeter Anvin <hpa@zytor.com>
      Signed-off-by: default avatarLinus Torvalds <torvalds@linux-foundation.org>
      f94edacf
  2. 17 Feb, 2012 1 commit
    • Linus Torvalds's avatar
      i387: move AMD K7/K8 fpu fxsave/fxrstor workaround from save to restore · 4903062b
      Linus Torvalds authored
      The AMD K7/K8 CPUs don't save/restore FDP/FIP/FOP unless an exception is
      pending.  In order to not leak FIP state from one process to another, we
      need to do a floating point load after the fxsave of the old process,
      and before the fxrstor of the new FPU state.  That resets the state to
      the (uninteresting) kernel load, rather than some potentially sensitive
      user information.
      
      We used to do this directly after the FPU state save, but that is
      actually very inconvenient, since it
      
       (a) corrupts what is potentially perfectly good FPU state that we might
           want to lazy avoid restoring later and
      
       (b) on x86-64 it resulted in a very annoying ordering constraint, where
           "__unlazy_fpu()" in the task switch needs to be delayed until after
           the DS segment has been reloaded just to get the new DS value.
      
      Coupling it to the fxrstor instead of the fxsave automatically avoids
      both of these issues, and also ensures that we only do it when actually
      necessary (the FP state after a save may never actually get used).  It's
      simply a much more natural place for the leaked state cleanup.
      Signed-off-by: default avatarLinus Torvalds <torvalds@linux-foundation.org>
      4903062b
  3. 16 Feb, 2012 4 commits
    • Linus Torvalds's avatar
      i387: do not preload FPU state at task switch time · b3b0870e
      Linus Torvalds authored
      Yes, taking the trap to re-load the FPU/MMX state is expensive, but so
      is spending several days looking for a bug in the state save/restore
      code.  And the preload code has some rather subtle interactions with
      both paravirtualization support and segment state restore, so it's not
      nearly as simple as it should be.
      
      Also, now that we no longer necessarily depend on a single bit (ie
      TS_USEDFPU) for keeping track of the state of the FPU, we migth be able
      to do better.  If we are really switching between two processes that
      keep touching the FP state, save/restore is inevitable, but in the case
      of having one process that does most of the FPU usage, we may actually
      be able to do much better than the preloading.
      
      In particular, we may be able to keep track of which CPU the process ran
      on last, and also per CPU keep track of which process' FP state that CPU
      has.  For modern CPU's that don't destroy the FPU contents on save time,
      that would allow us to do a lazy restore by just re-enabling the
      existing FPU state - with no restore cost at all!
      Signed-off-by: default avatarLinus Torvalds <torvalds@linux-foundation.org>
      b3b0870e
    • Linus Torvalds's avatar
      i387: don't ever touch TS_USEDFPU directly, use helper functions · 6d59d7a9
      Linus Torvalds authored
      This creates three helper functions that do the TS_USEDFPU accesses, and
      makes everybody that used to do it by hand use those helpers instead.
      
      In addition, there's a couple of helper functions for the "change both
      CR0.TS and TS_USEDFPU at the same time" case, and the places that do
      that together have been changed to use those.  That means that we have
      fewer random places that open-code this situation.
      
      The intent is partly to clarify the code without actually changing any
      semantics yet (since we clearly still have some hard to reproduce bug in
      this area), but also to make it much easier to use another approach
      entirely to caching the CR0.TS bit for software accesses.
      
      Right now we use a bit in the thread-info 'status' variable (this patch
      does not change that), but we might want to make it a full field of its
      own or even make it a per-cpu variable.
      Signed-off-by: default avatarLinus Torvalds <torvalds@linux-foundation.org>
      6d59d7a9
    • Linus Torvalds's avatar
      i387: move TS_USEDFPU clearing out of __save_init_fpu and into callers · b6c66418
      Linus Torvalds authored
      Touching TS_USEDFPU without touching CR0.TS is confusing, so don't do
      it.  By moving it into the callers, we always do the TS_USEDFPU next to
      the CR0.TS accesses in the source code, and it's much easier to see how
      the two go hand in hand.
      Signed-off-by: default avatarLinus Torvalds <torvalds@linux-foundation.org>
      b6c66418
    • Linus Torvalds's avatar
      i387: fix x86-64 preemption-unsafe user stack save/restore · 15d8791c
      Linus Torvalds authored
      Commit 5b1cbac3 ("i387: make irq_fpu_usable() tests more robust")
      added a sanity check to the #NM handler to verify that we never cause
      the "Device Not Available" exception in kernel mode.
      
      However, that check actually pinpointed a (fundamental) race where we do
      cause that exception as part of the signal stack FPU state save/restore
      code.
      
      Because we use the floating point instructions themselves to save and
      restore state directly from user mode, we cannot do that atomically with
      testing the TS_USEDFPU bit: the user mode access itself may cause a page
      fault, which causes a task switch, which saves and restores the FP/MMX
      state from the kernel buffers.
      
      This kind of "recursive" FP state save is fine per se, but it means that
      when the signal stack save/restore gets restarted, it will now take the
      '#NM' exception we originally tried to avoid.  With preemption this can
      happen even without the page fault - but because of the user access, we
      cannot just disable preemption around the save/restore instruction.
      
      There are various ways to solve this, including using the
      "enable/disable_page_fault()" helpers to not allow page faults at all
      during the sequence, and fall back to copying things by hand without the
      use of the native FP state save/restore instructions.
      
      However, the simplest thing to do is to just allow the #NM from kernel
      space, but fix the race in setting and clearing CR0.TS that this all
      exposed: the TS bit changes and the TS_USEDFPU bit absolutely have to be
      atomic wrt scheduling, so while the actual state save/restore can be
      interrupted and restarted, the act of actually clearing/setting CR0.TS
      and the TS_USEDFPU bit together must not.
      
      Instead of just adding random "preempt_disable/enable()" calls to what
      is already excessively ugly code, this introduces some helper functions
      that mostly mirror the "kernel_fpu_begin/end()" functionality, just for
      the user state instead.
      
      Those helper functions should probably eventually replace the other
      ad-hoc CR0.TS and TS_USEDFPU tests too, but I'll need to think about it
      some more: the task switching functionality in particular needs to
      expose the difference between the 'prev' and 'next' threads, while the
      new helper functions intentionally were written to only work with
      'current'.
      Signed-off-by: default avatarLinus Torvalds <torvalds@linux-foundation.org>
      15d8791c
  4. 15 Feb, 2012 1 commit
    • Linus Torvalds's avatar
      i387: fix sense of sanity check · c38e2345
      Linus Torvalds authored
      The check for save_init_fpu() (introduced in commit 5b1cbac3: "i387:
      make irq_fpu_usable() tests more robust") was the wrong way around, but
      I hadn't noticed, because my "tests" were bogus: the FPU exceptions are
      disabled by default, so even doing a divide by zero never actually
      triggers this code at all unless you do extra work to enable them.
      
      So if anybody did enable them, they'd get one spurious warning.
      Signed-off-by: default avatarLinus Torvalds <torvalds@linux-foundation.org>
      c38e2345
  5. 14 Feb, 2012 32 commits