| Message ID | 20170208204240.4002411-1-songliubraving@fb.com (mailing list archive) |
|---|---|
| State | Superseded, archived |
| Headers | show |
On Wed, Feb 08, 2017 at 12:42:40PM -0800, Song Liu wrote: > When a device is deleted through sysfs handle "delete", the code > locks shost->scan_mutex. If multiple devices are deleted at the > same time, these deletes will be handled in series. > > On the other hand, sd_shutdown() sometimes issues long latency > commands: sync cache and start_stop. It is not necessary for these > commands to run in series. > > To reduce latency of parallel "delete" requests, this patch unlock > shost->scan_mutex before long latency commands and relock the mutex > after the command. > > Fixed bug from previous version. > > Reported-by: kernel test robot <fengguang.wu@intel.com> > Signed-off-by: Song Liu <songliubraving@fb.com> > --- > drivers/scsi/sd.c | 15 +++++++++++++++ > 1 file changed, 15 insertions(+) > > diff --git a/drivers/scsi/sd.c b/drivers/scsi/sd.c > index 9e0783b..22add77 100644 > --- a/drivers/scsi/sd.c > +++ b/drivers/scsi/sd.c > @@ -3304,6 +3304,9 @@ static int sd_start_stop_device(struct scsi_disk *sdkp, int start) > static void sd_shutdown(struct device *dev) > { > struct scsi_disk *sdkp = dev_get_drvdata(dev); > + struct scsi_device *sdev; > + struct Scsi_Host *shost; > + bool scan_mutex_locked; > > if (!sdkp) > return; /* this can happen */ > @@ -3311,14 +3314,26 @@ static void sd_shutdown(struct device *dev) > if (pm_runtime_suspended(dev)) > return; > > + sdev = sdkp->device; > + shost = sdev->host; > + scan_mutex_locked = mutex_is_locked(&shost->scan_mutex); The use of mutex_is_locked outside of asserts is always bogus. The fix you want is most like in the caller of the method.
> On Feb 8, 2017, at 1:02 PM, Christoph Hellwig <hch@infradead.org> wrote: > > On Wed, Feb 08, 2017 at 12:42:40PM -0800, Song Liu wrote: >> When a device is deleted through sysfs handle "delete", the code >> locks shost->scan_mutex. If multiple devices are deleted at the >> same time, these deletes will be handled in series. >> >> On the other hand, sd_shutdown() sometimes issues long latency >> commands: sync cache and start_stop. It is not necessary for these >> commands to run in series. >> >> To reduce latency of parallel "delete" requests, this patch unlock >> shost->scan_mutex before long latency commands and relock the mutex >> after the command. >> >> Fixed bug from previous version. >> >> Reported-by: kernel test robot <fengguang.wu@intel.com> >> Signed-off-by: Song Liu <songliubraving@fb.com> >> --- >> drivers/scsi/sd.c | 15 +++++++++++++++ >> 1 file changed, 15 insertions(+) >> >> diff --git a/drivers/scsi/sd.c b/drivers/scsi/sd.c >> index 9e0783b..22add77 100644 >> --- a/drivers/scsi/sd.c >> +++ b/drivers/scsi/sd.c >> @@ -3304,6 +3304,9 @@ static int sd_start_stop_device(struct scsi_disk *sdkp, int start) >> static void sd_shutdown(struct device *dev) >> { >> struct scsi_disk *sdkp = dev_get_drvdata(dev); >> + struct scsi_device *sdev; >> + struct Scsi_Host *shost; >> + bool scan_mutex_locked; >> >> if (!sdkp) >> return; /* this can happen */ >> @@ -3311,14 +3314,26 @@ static void sd_shutdown(struct device *dev) >> if (pm_runtime_suspended(dev)) >> return; >> >> + sdev = sdkp->device; >> + shost = sdev->host; >> + scan_mutex_locked = mutex_is_locked(&shost->scan_mutex); > > The use of mutex_is_locked outside of asserts is always bogus. > > The fix you want is most like in the caller of the method. Hi Christoph, Thanks for the feedback. The other caller (that does not hold scan_mutex) of sd_shutdown() is device_shutdown() in drivers/base/core.c. I could not think of a good way to lock scan_mutex there. Would some variation of mutex_trylock() be a better option here? Thanks, Song
diff --git a/drivers/scsi/sd.c b/drivers/scsi/sd.c index 9e0783b..22add77 100644 --- a/drivers/scsi/sd.c +++ b/drivers/scsi/sd.c @@ -3304,6 +3304,9 @@ static int sd_start_stop_device(struct scsi_disk *sdkp, int start) static void sd_shutdown(struct device *dev) { struct scsi_disk *sdkp = dev_get_drvdata(dev); + struct scsi_device *sdev; + struct Scsi_Host *shost; + bool scan_mutex_locked; if (!sdkp) return; /* this can happen */ @@ -3311,14 +3314,26 @@ static void sd_shutdown(struct device *dev) if (pm_runtime_suspended(dev)) return; + sdev = sdkp->device; + shost = sdev->host; + scan_mutex_locked = mutex_is_locked(&shost->scan_mutex); + if (sdkp->WCE && sdkp->media_present) { + if (scan_mutex_locked) + mutex_unlock(&shost->scan_mutex); sd_printk(KERN_NOTICE, sdkp, "Synchronizing SCSI cache\n"); sd_sync_cache(sdkp); + if (scan_mutex_locked) + mutex_lock(&shost->scan_mutex); } if (system_state != SYSTEM_RESTART && sdkp->device->manage_start_stop) { + if (scan_mutex_locked) + mutex_unlock(&shost->scan_mutex); sd_printk(KERN_NOTICE, sdkp, "Stopping disk\n"); sd_start_stop_device(sdkp, 0); + if (scan_mutex_locked) + mutex_lock(&shost->scan_mutex); } }
When a device is deleted through sysfs handle "delete", the code locks shost->scan_mutex. If multiple devices are deleted at the same time, these deletes will be handled in series. On the other hand, sd_shutdown() sometimes issues long latency commands: sync cache and start_stop. It is not necessary for these commands to run in series. To reduce latency of parallel "delete" requests, this patch unlock shost->scan_mutex before long latency commands and relock the mutex after the command. Fixed bug from previous version. Reported-by: kernel test robot <fengguang.wu@intel.com> Signed-off-by: Song Liu <songliubraving@fb.com> --- drivers/scsi/sd.c | 15 +++++++++++++++ 1 file changed, 15 insertions(+)