We have been observing hangs, both of KVM guest vcpu tasks and more
generally, where a process that is woken doesn't properly wake up and
continue to run, but instead sticks in TASK_WAKING state. This
happens because the update of rq->wake_list in ttwu_queue_remote()
is not ordered with the update of ipi_message in
smp_muxed_ipi_message_pass(), and the reading of rq->wake_list in
scheduler_ipi() is not ordered with the reading of ipi_message in
smp_ipi_demux(). Thus it is possible for the IPI receiver not to see
the updated rq->wake_list and therefore conclude that there is nothing
for it to do.
In order to make sure that anything done before smp_send_reschedule()
is ordered before anything done in the resulting call to scheduler_ipi(),
this adds barriers in smp_muxed_message_pass() and smp_ipi_demux().
The barrier in smp_muxed_message_pass() is a full barrier to ensure that
there is a full ordering between the smp_send_reschedule() caller and
scheduler_ipi(). In smp_ipi_demux(), we use xchg() rather than
xchg_local() because xchg() includes release and acquire barriers.
Using xchg() rather than xchg_local() makes sense given that
ipi_message is not just accessed locally.
This moves the barrier between setting the message and calling the
cause_ipi() function into the individual cause_ipi implementations.
Most of them -- those that used outb, out_8 or similar -- already had
a full barrier because out_8 etc. include a sync before the MMIO
store. This adds an explicit barrier in the two remaining cases.
These changes made no measurable difference to the speed of IPIs as
measured using a simple ping-pong latency test across two CPUs on
different cores of a POWER7 machine.
The analysis of the reason why processes were not waking up properly
is due to Milton Miller.
Cc: stable@vger.kernel.org # v3.0+
Reported-by: Milton Miller <miltonm@bga.com>
Signed-off-by: Paul Mackerras <paulus@samba.org>
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
In preparation to remove the slow revmap path, eliminate the public
radix revmap lookup functions. This simplifies the code and makes the
slowpath removal patch a lot simpler.
Signed-off-by: Grant Likely <grant.likely@secretlab.ca>
Cc: Paul Mundt <lethal@linux-sh.org>
Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Rob Herring <rob.herring@calxeda.com>
I have an intermittent kdump fail where the hypervisor fails an H_EOI.
As a result our CPPR is never reset to 0xff and we no longer accept
interrupts.
This patch calls icp_hv_set_cppr to reset the CPPR if H_EOI fails,
fixing the kdump fail.
Signed-off-by: Anton Blanchard <anton@samba.org>
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
During kdump stress testing I sometimes see the kdump kernel panic
with:
Interrupt 0x306 (real) is invalid, disabling it.
Kernel panic - not syncing: bad return code EOI - rc = -4, value=ff000306
Instead of panicing print the error message, dump the stack the first
time it happens and continue on. Add some more information to the
debug messages as well.
Signed-off-by: Anton Blanchard <anton@samba.org>
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Consolidate the mux and demux of ipi messages into smp.c and call
a new smp_ops callback to actually trigger the ipi.
The powerpc architecture code is optimised for having 4 distinct
ipi triggers, which are mapped to 4 distinct messages (ipi many, ipi
single, scheduler ipi, and enter debugger). However, several interrupt
controllers only provide a single software triggered interrupt that
can be delivered to each cpu. To resolve this limitation, each smp_ops
implementation created a per-cpu variable that is manipulated with atomic
bitops. Since these lines will be contended they are optimialy marked as
shared_aligned and take a full cache line for each cpu. Distro kernels
may have 2 or 3 of these in their config, each taking per-cpu space
even though at most one will be in use.
This consolidation removes smp_message_recv and replaces the single call
actions cases with direct calls from the common message recognition loop.
The complicated debugger ipi case with its muxed crash handling code is
moved to debug_ipi_action which is now called from the demux code (instead
of the multi-message action calling smp_message_recv).
I put a call to reschedule_action to increase the likelyhood of correctly
merging the anticipated scheduler_ipi() hook coming from the scheduler
tree; that single required call can be inlined later.
The actual message decode is a copy of the old pseries xics code with its
memory barriers and cache line spacing, augmented with a per-cpu unsigned
long based on the book-e doorbell code. The optional data is set via a
callback from the implementation and is passed to the new cause-ipi hook
along with the logical cpu number. While currently only the doorbell
implemntation uses this data it should be almost zero cost to retrieve and
pass it -- it adds a single register load for the argument from the same
cache line to which we just completed a store and the register is dead
on return from the call. I extended the data element from unsigned int
to unsigned long in case some other code wanted to associate a pointer.
The doorbell check_self is replaced by a call to smp_muxed_ipi_resend,
conditioned on the CPU_DBELL feature. The ifdef guard could be relaxed
to CONFIG_SMP but I left it with BOOKE for now.
Also, the doorbell interrupt vector for book-e was not calling irq_enter
and irq_exit, which throws off cpu accounting and causes code to not
realize it is running in interrupt context. Add the missing calls.
Signed-off-by: Milton Miller <miltonm@bga.com>
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Now that smp_ops->smp_message_pass is always called with an (online) cpu
number for the target remove the checks for MSG_ALL and MSG_ALL_BUT_SELF.
Signed-off-by: Milton Miller <miltonm@bga.com>
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
This is a significant rework of the XICS driver, too significant to
conveniently break it up into a series of smaller patches to be honest.
The driver is moved to a more generic location to allow new platforms
to use it, and is broken up into separate ICP and ICS "backends". For
now we have the native and "hypervisor" ICP backends and one common
RTAS ICS backend.
The driver supports one ICP backend instanciation, and many ICS ones,
in order to accomodate future platforms with multiple possibly different
interrupt "sources" mechanisms.
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Paul Mackerras