Patches 7cba160ad "powernv/cpuidle: Redesign idle states management"
and 77b54e9f2 "powernv/powerpc: Add winkle support for offline cpus"
use non-volatile condition registers (cr2, cr3 and cr4) early in the system
reset interrupt handler (system_reset_pSeries()) before it has been determined
if state loss has occurred. If state loss has not occurred, control returns via
the power7_wakeup_noloss() path which does not restore those condition
registers, leaving them corrupted.
Fix this by restoring the condition registers in the power7_wakeup_noloss()
case.
This is apparent when running a KVM guest on hardware that does not
support winkle or sleep and the guest makes use of secondary threads. In
practice this means Power7 machines, though some early unreleased Power8
machines may also be susceptible.
The secondary CPUs are taken off line before the guest is started and
they call pnv_smp_cpu_kill_self(). This checks support for sleep
states (in this case there is no support) and power7_nap() is called.
When the CPU is woken, power7_nap() returns and because the CPU is
still off line, the main while loop executes again. The sleep states
support test is executed again, but because the tested values cannot
have changed, the compiler has optimized the test away and instead we
rely on the result of the first test, which has been left in cr3
and/or cr4. With the result overwritten, the wrong branch is taken and
power7_winkle() is called on a CPU that does not support it, leading
to it stalling.
Fixes: 7cba160ad7 ("powernv/cpuidle: Redesign idle states management")
Fixes: 77b54e9f21 ("powernv/powerpc: Add winkle support for offline cpus")
[mpe: Massage change log a bit more]
Signed-off-by: Sam Bobroff <sam.bobroff@au1.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
The power7_nap(), power7_sleep() and power7_winkle() functions are
called from pnv_smp_cpu_kill_self(), which expects them to return the
SRR1 value set by the hardware on wakeup, or 0 if no nap/sleep/winkle
occurred. However, in the case where an interrupt needs to be
replayed, the logic in power7_powersave_common (the common code for
power7_nap et al.) doesn't set r3 to 0 in this case. Instead what we
get as the return value is the selector for the type of power-saving
mode requested (1, 2 or 3). In fact this should not affect the
operation of pnv_smp_cpu_kill_self(), but it is better to get this
correct, so this adds an instruction to set r3 to 0 in this case.
Signed-off-by: Paul Mackerras <paulus@samba.org>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Winkle is a deep idle state supported in power8 chips. A core enters
winkle when all the threads of the core enter winkle. In this state
power supply to the entire chiplet i.e core, private L2 and private L3
is turned off. As a result it gives higher powersavings compared to
sleep.
But entering winkle results in a total hypervisor state loss. Hence the
hypervisor context has to be preserved before entering winkle and
restored upon wake up.
Power-on Reset Engine (PORE) is a dedicated engine which is responsible
for powering on the chiplet during wake up. It can be programmed to
restore the register contests of a few specific registers. This patch
uses PORE to restore register state wherever possible and uses stack to
save and restore rest of the necessary registers.
With hypervisor state restore things fall under three categories-
per-core state, per-subcore state and per-thread state. To manage this,
extend the infrastructure introduced for sleep. Mainly we add a paca
variable subcore_sibling_mask. Using this and the core_idle_state we can
distingush first thread in core and subcore.
Signed-off-by: Shreyas B. Prabhu <shreyas@linux.vnet.ibm.com>
Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Michael Ellerman <mpe@ellerman.id.au>
Cc: linuxppc-dev@lists.ozlabs.org
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Deep idle states like sleep and winkle are per core idle states. A core
enters these states only when all the threads enter either the
particular idle state or a deeper one. There are tasks like fastsleep
hardware bug workaround and hypervisor core state save which have to be
done only by the last thread of the core entering deep idle state and
similarly tasks like timebase resync, hypervisor core register restore
that have to be done only by the first thread waking up from these
state.
The current idle state management does not have a way to distinguish the
first/last thread of the core waking/entering idle states. Tasks like
timebase resync are done for all the threads. This is not only is
suboptimal, but can cause functionality issues when subcores and kvm is
involved.
This patch adds the necessary infrastructure to track idle states of
threads in a per-core structure. It uses this info to perform tasks like
fastsleep workaround and timebase resync only once per core.
Signed-off-by: Shreyas B. Prabhu <shreyas@linux.vnet.ibm.com>
Originally-by: Preeti U. Murthy <preeti@linux.vnet.ibm.com>
Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Michael Ellerman <mpe@ellerman.id.au>
Cc: Rafael J. Wysocki <rjw@rjwysocki.net>
Cc: linux-pm@vger.kernel.org
Cc: linuxppc-dev@lists.ozlabs.org
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Currently, when going idle, we set the flag indicating that we are in
nap mode (paca->kvm_hstate.hwthread_state) and then execute the nap
(or sleep or rvwinkle) instruction, all with the MMU on. This is bad
for two reasons: (a) the architecture specifies that those instructions
must be executed with the MMU off, and in fact with only the SF, HV, ME
and possibly RI bits set, and (b) this introduces a race, because as
soon as we set the flag, another thread can switch the MMU to a guest
context. If the race is lost, this thread will typically start looping
on relocation-on ISIs at 0xc...4400.
This fixes it by setting the MSR as required by the architecture before
setting the flag or executing the nap/sleep/rvwinkle instruction.
Cc: stable@vger.kernel.org
[ shreyas@linux.vnet.ibm.com: Edited to handle LE ]
Signed-off-by: Paul Mackerras <paulus@samba.org>
Signed-off-by: Shreyas B. Prabhu <shreyas@linux.vnet.ibm.com>
Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Cc: Michael Ellerman <mpe@ellerman.id.au>
Cc: linuxppc-dev@lists.ozlabs.org
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
When a secondary hardware thread has finished running a KVM guest, we
currently put that thread into nap mode using a nap instruction in
the KVM code. This changes the code so that instead of doing a nap
instruction directly, we instead cause the call to power7_nap() that
put the thread into nap mode to return. The reason for doing this is
to avoid having the KVM code having to know what low-power mode to
put the thread into.
In the case of a secondary thread used to run a KVM guest, the thread
will be offline from the point of view of the host kernel, and the
relevant power7_nap() call is the one in pnv_smp_cpu_disable().
In this case we don't want to clear pending IPIs in the offline loop
in that function, since that might cause us to miss the wakeup for
the next time the thread needs to run a guest. To tell whether or
not to clear the interrupt, we use the SRR1 value returned from
power7_nap(), and check if it indicates an external interrupt. We
arrange that the return from power7_nap() when we have finished running
a guest returns 0, so pending interrupts don't get flushed in that
case.
Note that it is important a secondary thread that has finished
executing in the guest, or that didn't have a guest to run, should
not return to power7_nap's caller while the kvm_hstate.hwthread_req
flag in the PACA is non-zero, because the return from power7_nap
will reenable the MMU, and the MMU might still be in guest context.
In this situation we spin at low priority in real mode waiting for
hwthread_req to become zero.
Signed-off-by: Paul Mackerras <paulus@samba.org>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
On PowerNV platforms, when a CPU is offline, we put it into nap mode.
It's possible that the CPU wakes up from nap mode while it is still
offline due to a stray IPI. A misdirected device interrupt could also
potentially cause it to wake up. In that circumstance, we need to clear
the interrupt so that the CPU can go back to nap mode.
In the past the clearing of the interrupt was accomplished by briefly
enabling interrupts and allowing the normal interrupt handling code
(do_IRQ() etc.) to handle the interrupt. This has the problem that
this code calls irq_enter() and irq_exit(), which call functions such
as account_system_vtime() which use RCU internally. Use of RCU is not
permitted on offline CPUs and will trigger errors if RCU checking is
enabled.
To avoid calling into any generic code which might use RCU, we adopt
a different method of clearing interrupts on offline CPUs. Since we
are on the PowerNV platform, we know that the system interrupt
controller is a XICS being driven directly (i.e. not via hcalls) by
the kernel. Hence this adds a new icp_native_flush_interrupt()
function to the native-mode XICS driver and arranges to call that
when an offline CPU is woken from nap. This new function reads the
interrupt from the XICS. If it is an IPI, it clears the IPI; if it
is a device interrupt, it prints a warning and disables the source.
Then it does the end-of-interrupt processing for the interrupt.
The other thing that briefly enabling interrupts did was to check and
clear the irq_happened flag in this CPU's PACA. Therefore, after
flushing the interrupt from the XICS, we also clear all bits except
the PACA_IRQ_HARD_DIS (interrupts are hard disabled) bit from the
irq_happened flag. The PACA_IRQ_HARD_DIS flag is set by power7_nap()
and is left set to indicate that interrupts are hard disabled. This
means we then have to ignore that flag in power7_nap(), which is
reasonable since it doesn't indicate that any interrupt event needs
servicing.
Signed-off-by: Paul Mackerras <paulus@samba.org>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
(NOTE: This patch depends on upstream HMI handling patchset at
https://lists.ozlabs.org/pipermail/linuxppc-dev/2014-July/119731.html)
The current HMI handling on napping cpus does not take care of endianess
issue. On LE host kernel when we wake up from nap due to HMI interrupt we
would checkstop while jumping into opal call. There is a similar issue in
case of fast sleep wakeup where the code invokes opal_resync_tb opal call
without handling LE issue. This patch fixes that as well.
With this patch applied, HMIs handling on LE host kernel works fine.
Signed-off-by: Mahesh Salgaonkar <mahesh@linux.vnet.ibm.com>
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
HMIs are thread specific and can come while thread is in sleep/nap mode.
Hence with SMT=off mode we can receive HMIs on sleeping threads. For
interrupt received in nap mode, cpu wakes up at system reset vector, clears
the interrupt and go back to nap mode again. But HMIs are sticky and they
keep happening until we clear reason bits from HMER. Hence add a special
check for HMI in reset vector (through power7_wakeup_* functions) and
invoke opal call to handle HMI.
Signed-off-by: Mahesh Salgaonkar <mahesh@linux.vnet.ibm.com>
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Commit 8d6f7c5a: "powerpc/powernv: Make it possible to skip the IRQHAPPENED
check in power7_nap()" added code that prevents cpus from checking for
pending interrupts just before entering sleep state, which is wrong. These
interrupts are delivered during the soft irq disabled state of the cpu.
A cpu cannot enter any idle state with pending interrupts because they will
never be serviced until the next time the cpu is woken up by some other
interrupt. Its only then that the pending interrupts are replayed. This can result
in device timeouts or warnings about this cpu being stuck.
This patch fixes ths issue by ensuring that cpus check for pending interrupts
just before entering any idle state as long as they are not in the path of split
core operations.
Signed-off-by: Preeti U Murthy <preeti@linux.vnet.ibm.com>
Acked-by: Michael Ellerman <mpe@ellerman.id.au>
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
To support split core we need to be able to force all secondaries into
nap, so the core can detect they are idle and do an unsplit.
Currently power7_nap() will return without napping if there is an irq
pending. We want to ignore the pending irq and nap anyway, we will deal
with the interrupt later.
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Signed-off-by: Michael Neuling <mikey@neuling.org>
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
binutils is smart enough to know that a branch to a function
descriptor is actually a branch to the functions text address.
Alan tells me that binutils has been doing this for 9 years.
Signed-off-by: Anton Blanchard <anton@samba.org>
During "Fast-sleep" and deeper power savings state, decrementer and
timebase could be stopped making it out of sync with rest
of the cores in the system.
Add a firmware call to request platform to resync timebase
using low level platform methods.
Signed-off-by: Vaidyanathan Srinivasan <svaidy@linux.vnet.ibm.com>
Signed-off-by: Preeti U. Murthy <preeti@linux.vnet.ibm.com>
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Before adding Fast-Sleep into the cpuidle framework, some low level
support needs to be added to enable it. This includes saving and
restoring of certain registers at entry and exit time of this state
respectively just like we do in the NAP idle state.
Signed-off-by: Vaidyanathan Srinivasan <svaidy@linux.vnet.ibm.com>
[Changelog modified by Preeti U. Murthy <preeti@linux.vnet.ibm.com>]
Signed-off-by: Preeti U. Murthy <preeti@linux.vnet.ibm.com>
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
We can get machine checks from any context. We need to make sure that
we handle all of them correctly. If we are coming from hypervisor user-space,
we can continue in host kernel in virtual mode to deliver the MC event.
If we got woken up from power-saving mode then we may come in with one of
the following state:
a. No state loss
b. Supervisor state loss
c. Hypervisor state loss
For (a) and (b), we go back to nap again. State (c) is fatal, keep spinning.
For all other context which we not sure of queue up the MCE event and return
from the interrupt.
Signed-off-by: Mahesh Salgaonkar <mahesh@linux.vnet.ibm.com>
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
This help ups to select the relevant code in the kernel code
when we later move HV and PR bits as seperate modules. The patch
also makes the config options for PR KVM selectable
Signed-off-by: Aneesh Kumar K.V <aneesh.kumar@linux.vnet.ibm.com>
Signed-off-by: Alexander Graf <agraf@suse.de>
The CPU hotplug code for the powernv platform currently only puts
offline CPUs into nap mode if the powersave_nap variable is set.
However, HV-style KVM on this platform requires secondary CPU threads
to be offline and in nap mode. Since we know nap mode works just
fine on all POWER7 machines, and the only machines that support the
powernv platform are POWER7 machines, this changes the code to
always put offline CPUs into nap mode, regardless of powersave_nap.
Powersave_nap still controls whether or not CPUs go into nap mode
when idle, as before.
Signed-off-by: Paul Mackerras <paulus@samba.org>
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Currently on POWER7, if we are running the guest on a core and we don't
need all the hardware threads, we do nothing to ensure that the unused
threads aren't executing in the kernel (other than checking that they
are offline). We just assume they're napping and we don't do anything
to stop them trying to enter the kernel while the guest is running.
This means that a stray IPI can wake up the hardware thread and it will
then try to enter the kernel, but since the core is in guest context,
it will execute code from the guest in hypervisor mode once it turns the
MMU on, which tends to lead to crashes or hangs in the host.
This fixes the problem by adding two new one-byte flags in the
kvmppc_host_state structure in the PACA which are used to interlock
between the primary thread and the unused secondary threads when entering
the guest. With these flags, the primary thread can ensure that the
unused secondaries are not already in kernel mode (i.e. handling a stray
IPI) and then indicate that they should not try to enter the kernel
if they do get woken for any reason. Instead they will go into KVM code,
find that there is no vcpu to run, acknowledge and clear the IPI and go
back to nap mode.
Signed-off-by: Paul Mackerras <paulus@samba.org>
Signed-off-by: Alexander Graf <agraf@suse.de>
Signed-off-by: Avi Kivity <avi@redhat.com>
The current implementation of lazy interrupts handling has some
issues that this tries to address.
We don't do the various workarounds we need to do when re-enabling
interrupts in some cases such as when returning from an interrupt
and thus we may still lose or get delayed decrementer or doorbell
interrupts.
The current scheme also makes it much harder to handle the external
"edge" interrupts provided by some BookE processors when using the
EPR facility (External Proxy) and the Freescale Hypervisor.
Additionally, we tend to keep interrupts hard disabled in a number
of cases, such as decrementer interrupts, external interrupts, or
when a masked decrementer interrupt is pending. This is sub-optimal.
This is an attempt at fixing it all in one go by reworking the way
we do the lazy interrupt disabling from the ground up.
The base idea is to replace the "hard_enabled" field with a
"irq_happened" field in which we store a bit mask of what interrupt
occurred while soft-disabled.
When re-enabling, either via arch_local_irq_restore() or when returning
from an interrupt, we can now decide what to do by testing bits in that
field.
We then implement replaying of the missed interrupts either by
re-using the existing exception frame (in exception exit case) or via
the creation of a new one from an assembly trampoline (in the
arch_local_irq_enable case).
This removes the need to play with the decrementer to try to create
fake interrupts, among others.
In addition, this adds a few refinements:
- We no longer hard disable decrementer interrupts that occur
while soft-disabled. We now simply bump the decrementer back to max
(on BookS) or leave it stopped (on BookE) and continue with hard interrupts
enabled, which means that we'll potentially get better sample quality from
performance monitor interrupts.
- Timer, decrementer and doorbell interrupts now hard-enable
shortly after removing the source of the interrupt, which means
they no longer run entirely hard disabled. Again, this will improve
perf sample quality.
- On Book3E 64-bit, we now make the performance monitor interrupt
act as an NMI like Book3S (the necessary C code for that to work
appear to already be present in the FSL perf code, notably calling
nmi_enter instead of irq_enter). (This also fixes a bug where BookE
perfmon interrupts could clobber r14 ... oops)
- We could make "masked" decrementer interrupts act as NMIs when doing
timer-based perf sampling to improve the sample quality.
Signed-off-by-yet: Benjamin Herrenschmidt <benh@kernel.crashing.org>
---
v2:
- Add hard-enable to decrementer, timer and doorbells
- Fix CR clobber in masked irq handling on BookE
- Make embedded perf interrupt act as an NMI
- Add a PACA_HAPPENED_EE_EDGE for use by FSL if they want
to retrigger an interrupt without preventing hard-enable
v3:
- Fix or vs. ori bug on Book3E
- Fix enabling of interrupts for some exceptions on Book3E
v4:
- Fix resend of doorbells on return from interrupt on Book3E
v5:
- Rebased on top of my latest series, which involves some significant
rework of some aspects of the patch.
v6:
- 32-bit compile fix
- more compile fixes with various .config combos
- factor out the asm code to soft-disable interrupts
- remove the C wrapper around preempt_schedule_irq
v7:
- Fix a bug with hard irq state tracking on native power7
This fixes a problem where a CPU thread coming out of nap mode can
think it has valid values in the nonvolatile GPRs (r14 - r31) as saved
away in power7_idle, but in fact the values have been trashed because
the thread was used for KVM in the mean time. The result is that the
thread crashes because code that called power7_idle (e.g.,
pnv_smp_cpu_kill_self()) goes to use values in registers that have
been trashed.
The bit field in SRR1 that tells whether state was lost only reflects
the most recent nap, which may not have been the nap instruction in
power7_idle. So we need an extra PACA field to indicate that state
has been lost even if SRR1 indicates that the most recent nap didn't
lose state. We clear this field when saving the state in power7_idle,
we set it to a non-zero value when we use the thread for KVM, and we
test it in power7_wakeup_noloss.
Signed-off-by: Paul Mackerras <paulus@samba.org>
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
This lifts the restriction that book3s_hv guests can only run one
hardware thread per core, and allows them to use up to 4 threads
per core on POWER7. The host still has to run single-threaded.
This capability is advertised to qemu through a new KVM_CAP_PPC_SMT
capability. The return value of the ioctl querying this capability
is the number of vcpus per virtual CPU core (vcore), currently 4.
To use this, the host kernel should be booted with all threads
active, and then all the secondary threads should be offlined.
This will put the secondary threads into nap mode. KVM will then
wake them from nap mode and use them for running guest code (while
they are still offline). To wake the secondary threads, we send
them an IPI using a new xics_wake_cpu() function, implemented in
arch/powerpc/sysdev/xics/icp-native.c. In other words, at this stage
we assume that the platform has a XICS interrupt controller and
we are using icp-native.c to drive it. Since the woken thread will
need to acknowledge and clear the IPI, we also export the base
physical address of the XICS registers using kvmppc_set_xics_phys()
for use in the low-level KVM book3s code.
When a vcpu is created, it is assigned to a virtual CPU core.
The vcore number is obtained by dividing the vcpu number by the
number of threads per core in the host. This number is exported
to userspace via the KVM_CAP_PPC_SMT capability. If qemu wishes
to run the guest in single-threaded mode, it should make all vcpu
numbers be multiples of the number of threads per core.
We distinguish three states of a vcpu: runnable (i.e., ready to execute
the guest), blocked (that is, idle), and busy in host. We currently
implement a policy that the vcore can run only when all its threads
are runnable or blocked. This way, if a vcpu needs to execute elsewhere
in the kernel or in qemu, it can do so without being starved of CPU
by the other vcpus.
When a vcore starts to run, it executes in the context of one of the
vcpu threads. The other vcpu threads all go to sleep and stay asleep
until something happens requiring the vcpu thread to return to qemu,
or to wake up to run the vcore (this can happen when another vcpu
thread goes from busy in host state to blocked).
It can happen that a vcpu goes from blocked to runnable state (e.g.
because of an interrupt), and the vcore it belongs to is already
running. In that case it can start to run immediately as long as
the none of the vcpus in the vcore have started to exit the guest.
We send the next free thread in the vcore an IPI to get it to start
to execute the guest. It synchronizes with the other threads via
the vcore->entry_exit_count field to make sure that it doesn't go
into the guest if the other vcpus are exiting by the time that it
is ready to actually enter the guest.
Note that there is no fixed relationship between the hardware thread
number and the vcpu number. Hardware threads are assigned to vcpus
as they become runnable, so we will always use the lower-numbered
hardware threads in preference to higher-numbered threads if not all
the vcpus in the vcore are runnable, regardless of which vcpus are
runnable.
Signed-off-by: Paul Mackerras <paulus@samba.org>
Signed-off-by: Alexander Graf <agraf@suse.de>
Wakeup comes from the system reset handler with a potential loss of
the non-hypervisor CPU state. We save the non-volatile state on the
stack and a pointer to it in the PACA, which the system reset handler
uses to restore things
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Sam Bobroff