| Message ID | 20210519222640.5153-1-michael.christie@oracle.com |
|---|---|
| State | New |
| Headers | show |
| Series | scsi: target: iblock: Fix smp_processor_id BUG messages | expand |
On May 19, 2021 / 17:26, Mike Christie wrote: > This has us use raw_smp_processor_id in iblock's plug_device callout. > smp_processor_id is not needed here, because we are running from a per CPU > work item that is also queued to run on a worker thread that is normally > bound to a specific CPU. If the worker thread did end up switching CPUs > then it's handled the same way we handle when the work got moved to a > different CPU's worker thread, where we will just end up sending IO from > the new CPU. > > Fixes: 415ccd9811da ("scsi: target: iblock: Add backend plug/unplug > callouts") > Signed-off-by: Mike Christie <michael.christie@oracle.com> Thanks Mike. To confirm this patch, I tried to trigger the BUG message using block backstore on kernel configured with DEBUG_PREEMPT, but I was not able to trigger it. Having said that, I think this change prevents the potential BUG message with the same reasoning as the other fix in target_core_transport [1]. As far as I read through the code, the code should work even when CPU switches. [1] https://marc.info/?l=linux-scsi&m=162106219905019 -- Best Regards, Shin'ichiro Kawasaki
On Wed, 19 May 2021 17:26:40 -0500, Mike Christie wrote: > This has us use raw_smp_processor_id in iblock's plug_device callout. > smp_processor_id is not needed here, because we are running from a per CPU > work item that is also queued to run on a worker thread that is normally > bound to a specific CPU. If the worker thread did end up switching CPUs > then it's handled the same way we handle when the work got moved to a > different CPU's worker thread, where we will just end up sending IO from > the new CPU. Applied to 5.13/scsi-fixes, thanks! [1/1] scsi: target: iblock: Fix smp_processor_id BUG messages https://git.kernel.org/mkp/scsi/c/5aaeca258f55 -- Martin K. Petersen Oracle Linux Engineering
diff --git a/drivers/target/target_core_iblock.c b/drivers/target/target_core_iblock.c index d6fdd1c61f90..a526f9678c34 100644 --- a/drivers/target/target_core_iblock.c +++ b/drivers/target/target_core_iblock.c @@ -204,11 +204,11 @@ static struct se_dev_plug *iblock_plug_device(struct se_device *se_dev) struct iblock_dev_plug *ib_dev_plug; /* - * Each se_device has a per cpu work this can be run from. Wwe + * Each se_device has a per cpu work this can be run from. We * shouldn't have multiple threads on the same cpu calling this * at the same time. */ - ib_dev_plug = &ib_dev->ibd_plug[smp_processor_id()]; + ib_dev_plug = &ib_dev->ibd_plug[raw_smp_processor_id()]; if (test_and_set_bit(IBD_PLUGF_PLUGGED, &ib_dev_plug->flags)) return NULL;
This has us use raw_smp_processor_id in iblock's plug_device callout. smp_processor_id is not needed here, because we are running from a per CPU work item that is also queued to run on a worker thread that is normally bound to a specific CPU. If the worker thread did end up switching CPUs then it's handled the same way we handle when the work got moved to a different CPU's worker thread, where we will just end up sending IO from the new CPU. Fixes: 415ccd9811da ("scsi: target: iblock: Add backend plug/unplug callouts") Signed-off-by: Mike Christie <michael.christie@oracle.com> --- drivers/target/target_core_iblock.c | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-)