On Wed, 2019-01-02 at 14:25 +-0800, stanley.chu+AEA-mediatek.com wrote: +AD4 From: Stanley Chu +ADw-stanley.chu+AEA-mediatek.com+AD4 +AD4 +AD4 The commit 356fd2663cff (+ACI-scsi: Set request queue runtime PM status +AD4 back to active on resume+ACI) fixed up the inconsistent RPM status between +AD4 request queue and device. However changing request queue RPM status +AD4 shall be done only on successful resume, otherwise status may be still +AD4 inconsistent as below, +AD4 +AD4 Request queue: RPM+AF8-ACTIVE +AD4 Device: RPM+AF8-SUSPENDED +AD4 +AD4 This ends up soft lockup because requests can be submitted to +AD4 underlying devices but those devices and their required resource +AD4 are not resumed. +AD4 +AD4 Signed-off-by: Stanley Chu +ADw-stanley.chu+AEA-mediatek.com+AD4 Please add +ACI-Fixes:+ACI and +ACI-Cc: stable+ACI tags and also Cc the author of commit 356fd2663cff. +AD4 --- +AD4 drivers/scsi/scsi+AF8-pm.c +AHw 24 +-+-+-+-+-+-+-+-+-+-+-+-+-+----------- +AD4 1 file changed, 14 insertions(+-), 10 deletions(-) +AD4 +AD4 diff --git a/drivers/scsi/scsi+AF8-pm.c b/drivers/scsi/scsi+AF8-pm.c +AD4 index a2b4179..eff3e59 100644 +AD4 --- a/drivers/scsi/scsi+AF8-pm.c +AD4 +-+-+- b/drivers/scsi/scsi+AF8-pm.c +AD4 +AEAAQA -82,6 +-82,20 +AEAAQA static int scsi+AF8-dev+AF8-type+AF8-resume(struct device +ACo-dev, +AD4 pm+AF8-runtime+AF8-disable(dev)+ADs +AD4 pm+AF8-runtime+AF8-set+AF8-active(dev)+ADs +AD4 pm+AF8-runtime+AF8-enable(dev)+ADs +AD4 +- +AD4 +- /+ACo +AD4 +- +ACo Forcibly set runtime PM status of request queue to +ACI-active+ACI +AD4 +- +ACo to make sure we can again get requests from the queue +AD4 +- +ACo (see also blk+AF8-pm+AF8-peek+AF8-request()). +AD4 +- +ACo +AD4 +- +ACo The resume hook will correct runtime PM status of the disk. +AD4 +- +ACo-/ +AD4 +- if (+ACE-err +ACYAJg scsi+AF8-is+AF8-sdev+AF8-device(dev)) +AHs +AD4 +- struct scsi+AF8-device +ACo-sdev +AD0 to+AF8-scsi+AF8-device(dev)+ADs +AD4 +- +AD4 +- if (sdev-+AD4-request+AF8-queue-+AD4-dev) +AD4 +- blk+AF8-set+AF8-runtime+AF8-active(sdev-+AD4-request+AF8-queue)+ADs +AD4 +- +AH0 What makes you think that the sdev-+AD4-request+AF8-queue-+AD4-dev test is necessary? The scsi+AF8-dev+AF8-type+AF8-resume() function is only called after blk+AF8-pm+AF8-runtime+AF8-init() has finished so I don't think that test is necessary. Additionally, since the above code occurs inside a block controlled by an +ACI-if (err +AD0APQ 0)+ACI statement, I think the +ACE-err test is redundant and should be left out. Thanks, Bart.
