• Deepthi Dharwar's avatar
    cpuidle/powerpc: Fix snooze state problem in the cpuidle design on pseries. · 83dac594
    Deepthi Dharwar authored
    Earlier without cpuidle framework on pseries, the native arch
    idle routine comprised of both snooze and nap
    states.  smt_snooze_delay variable was used to delay
    the idle process entry to deeper idle state like  nap.
    With the coming of cpuidle, this arch specific idle was replaced
    by two different idle routines, one for supporting snooze and other
    for nap. This enabled addition of more
    low level idle states on pseries in the future.
    
    On adopting the generic cpuidle framework for POWER systems,
    the decision of which idle state to choose from,  given a predicted
    idle time is taken by the menu governor based on
    target_residency and  exit_latency of the idle states.
    target_residency is the minimum time to be resident in that idle state.
    Exit_latency is time taken to exit out of idle state.
    Deeper the idle state, both the target residency and exit latency
    would be higher.
    
    In the current design, smt_snooze_delay is used as target_residency
    for the  snooze state which is incorrect, as it is not the
    minimum but the maximum duration to be in snooze state.
    This would  result in the governor in taking bad decision,
    as presently target_residency of nap < target_residency of snooze
    inspite of nap being deeper idle state.
    
    This patch aims to fix this problem by replacing the smt_snooze_delay loop
    in snooze state, with the need_resched()  as the governor is aware of
    entry and exit of various idle transitions based on which
    next idle time prediction.
    
    The governor is intelligent enough to determine the idle state the needs to
    be transitioned to and maintains a whole of heuristics including
    io load, previous idle states predictions etc for the same, based on
    which idle state entry decision is taken.
    
    With this fix, of setting target_residency of snooze to 0
    					     nap to smt_snooze_delay
    if the predicted idle time is less
    than smt_snooze_delay (target_residency of nap)
    value governor would pick snooze state, else nap. This adhers to the
    previous native idle design.
    Signed-off-by: default avatarDeepthi Dharwar <deepthi@linux.vnet.ibm.com>
    Signed-off-by: default avatarBenjamin Herrenschmidt <benh@kernel.crashing.org>
    83dac594
processor_idle.c 7.96 KB