On Intel platforms machine check exceptions are always broadcast to
all CPUs.  This patch makes the machine check handler synchronize all
these machine checks, elect a Monarch to handle the event and collect
the worst event from all CPUs and then process it first.

This has some advantages:

- When there is a truly data corrupting error the system panics as
  quickly as possible. This improves containment of corrupted
  data and makes sure the corrupted data never hits stable storage.

- The panics are synchronized and do not reenter the panic code
  on multiple CPUs (which currently does not handle this well).

- All the errors are reported. Currently it often happens that
  another CPU happens to do the panic first, but reports useless
  information (empty machine check) because the real error
  happened on another CPU which came in later.
  This is a big advantage on Nehalem where the 8 threads per CPU
  lead to often the wrong CPU winning the race and dumping
  useless information on a machine check.  The problem also occurs
  in a less severe form on older CPUs.

- The system can detect when no CPUs detected a machine check
  and shut down the system.  This can happen when one CPU is so
  badly hung that that it cannot process a machine check anymore
  or when some external agent wants to stop the system by
  asserting the machine check pin.  This follows Intel hardware
  recommendations.

- This matches the recommended error model by the CPU designers.

- The events can be output in true severity order

- When a panic happens on another CPU it makes sure to be actually
  be able to process the stop IPI by enabling interrupts.

The code is extremly careful to handle timeouts while waiting
for other CPUs. It can't rely on the normal timing mechanisms
(jiffies, ktime_get) because of its asynchronous/lockless nature,
so it uses own timeouts using ndelay() and a "SPINUNIT"

The timeout is configurable. By default it waits for upto one
second for the other CPUs.  This can be also disabled.

From some informal testing AMD systems do not see to broadcast
machine checks, so right now it's always disabled by default on
non Intel CPUs or also on very old Intel systems.

Includes fixes from Ying Huang
Fixed a "ecception" in a comment (H.Seto)
Moved global_nwo reset later based on suggestion from H.Seto
v2: Avoid duplicate messages

[ Impact: feature, fixes long standing problems. ]
Signed-off-by: NAndi Kleen <ak@linux.intel.com>
Signed-off-by: NHidetoshi Seto <seto.hidetoshi@jp.fujitsu.com>
Signed-off-by: NH. Peter Anvin <hpa@zytor.com>

Document that check_interval set to 0 means no polling.
Noticed by Hidetoshi Seto

Also add a reference from boot options to the sysfs tunables
Acked-by: NHidetoshi Seto <seto.hidetoshi@jp.fujitsu.com>
Signed-off-by: NAndi Kleen <ak@linux.intel.com>
Signed-off-by: NHidetoshi Seto <seto.hidetoshi@jp.fujitsu.com>
Signed-off-by: NH. Peter Anvin <hpa@zytor.com>

The current organization of the x86 documentation makes it appear as
if the "i386" documentation doesn't apply to x86-64, which is does.
Thus, move that documentation into Documentation/x86, and move the
x86-64-specific stuff into Documentation/x86/x86_64 with the eventual
goal to move stuff that isn't actually 64-bit specific back into
Documentation/x86.
Signed-off-by: NH. Peter Anvin <hpa@zytor.com>

Background:
 The MCE handler has several paths that it can take, depending on various
 conditions of the MCE status and the value of the 'tolerant' knob.  The
 exact semantics are not well defined and the code is a bit twisty.

Description:
 This patch makes the MCE handler's behavior more clear by documenting the
 behavior for various 'tolerant' levels.  It also fixes or enhances
 several small things in the handler.  Specifically:
     * If RIPV is set it is not safe to restart, so set the 'no way out'
       flag rather than the 'kill it' flag.
     * Don't panic() on correctable MCEs.
     * If the _OVER bit is set *and* the _UC bit is set (meaning possibly
       dropped uncorrected errors), set the 'no way out' flag.
     * Use EIPV for testing whether an app can be killed (SIGBUS) rather
       than RIPV.  According to docs, EIPV indicates that the error is
       related to the IP, while RIPV simply means the IP is valid to
       restart from.
     * Don't clear the MCi_STATUS registers until after the panic() path.
       This leaves the status bits set after the panic() so clever BIOSes
       can find them (and dumb BIOSes can do nothing).

 This patch also calls nonseekable_open() in mce_open (as suggested by akpm).

Result:
 Tolerant levels behave almost identically to how they always have, but
 not it's well defined.  There's a slightly higher chance of panic()ing
 when multiple errors happen (a good thing, IMHO).  If you take an MBE and
 panic(), the error status bits are not cleared.

Alternatives:
 None.

Testing:
 I used software to inject correctable and uncorrectable errors.  With
 tolerant = 3, the system usually survives.  With tolerant = 2, the system
 usually panic()s (PCC) but not always.  With tolerant = 1, the system
 always panic()s.  When the system panic()s, the BIOS is able to detect
 that the cause of death was an MC4.  I was not able to reproduce the
 case of a non-PCC error in userspace, with EIPV, with (tolerant < 3).
 That will be rare at best.
Signed-off-by: NTim Hockin <thockin@google.com>
Signed-off-by: NAndrew Morton <akpm@linux-foundation.org>
Signed-off-by: NAndi Kleen <ak@suse.de>
Signed-off-by: NLinus Torvalds <torvalds@linux-foundation.org>

Background:
 We've found that MCEs (specifically DRAM SBEs) tend to come in bunches,
 especially when we are trying really hard to stress the system out.  The
 current MCE poller uses a static interval which does not care whether it
 has or has not found MCEs recently.

Description:
 This patch makes the MCE poller adjust the polling interval dynamically.
 If we find an MCE, poll 2x faster (down to 10 ms).  When we stop finding
 MCEs, poll 2x slower (up to check_interval seconds).  The check_interval
 tunable becomes the max polling interval.  The "Machine check events
 logged" printk() is rate limited to the check_interval, which should be
 identical behavior to the old functionality.

Result:
 If you start to take a lot of correctable errors (not exceptions), you
 log them faster and more accurately (less chance of overflowing the MCA
 registers).  If you don't take a lot of errors, you will see no change.

Alternatives:
 I considered simply reducing the polling interval to 10 ms immediately
 and keeping it there as long as we continue to find errors.  This felt a
 bit heavy handed, but does perform significantly better for the default
 check_interval of 5 minutes (we're using a few seconds when testing for
 DRAM errors).  I could be convinced to go with this, if anyone felt it
 was not too aggressive.

Testing:
 I used an error-injecting DIMM to create lots of correctable DRAM errors
 and verified that the polling interval accelerates.  The printk() only
 happens once per check_interval seconds.

Patch:
 This patch is against 2.6.21-rc7.
Signed-Off-By: NTim Hockin <thockin@google.com>
Signed-off-by: NAndi Kleen <ak@suse.de>

When a machine check event is detected (including a AMD RevF threshold
overflow event) allow to run a "trigger" program. This allows user space
to react to such events sooner.

The trigger is configured using a new trigger entry in the
machinecheck sysfs interface. It is currently shared between
all CPUs.

I also fixed the AMD threshold handler to run the machine
check polling code immediately to actually log any events
that might have caused the threshold interrupt.

Also added some documentation for the mce sysfs interface.
Signed-off-by: NAndi Kleen <ak@suse.de>