On Mon, Aug 19, 2019 at 05:49:48AM +0200, hch@xxxxxx wrote: > On Mon, Aug 19, 2019 at 10:08:31AM +1000, Dave Chinner wrote: > > > So I can also reproduce the same issue with the ramdisk driver, but not > > > with any other 4k sector size device (nvmet, scsi target, scsi_debug, > > > loop). > > > > How did you reproduce it? I tried with indentical pmem config to > > what Vishal was using and couldn't hit it... > > Just mkfs and mount the pmem device. Ran that in a loop overnight, no failures. > The fake pmem was create by the following kernel command line sniplet: > "memmap=2000M!1024M,2000M!4096M" and my kernel config is attached. I Very similar to what I'm using, just 2x8GB pmem instead of the smaller 1/4GB pair you are using. I even tried manually making it smaller w/ -d size= and changing log sizes, etc. I tried squeezing down memory to only 256MB unused (via mlock) but that didn't seem to change anything, either... > suspect you'll need CONFIG_SLUB to create the woes, but let me confirm > that.. Yup, that's what I have configured in my kernel, too. It's still a mystery to me why I'm not able to reproduce it easily. If it was just pmem, then I should have found it back when I tripped over the chaining bug, because the two VMs I were running testing on at the time were an iscsi test rig and an pmem test rig. The iscsi one tripped over the chaining bug, the pmem one has had no problems.... Cheers, Dave. -- Dave Chinner david@xxxxxxxxxxxxx
