On 12/7/20 9:59 PM, Collin Walling wrote: > On 12/7/20 3:16 PM, Christian Borntraeger wrote: >> >> >> On 07.12.20 21:13, Collin Walling wrote: >>> On 12/7/20 3:09 PM, Christian Borntraeger wrote: >>>> >>>> >>>> On 07.12.20 21:06, Collin Walling wrote: >>>>> On 12/7/20 2:32 PM, Christian Borntraeger wrote: >>>>>> On 07.12.20 16:41, Collin Walling wrote: >>>>>>> The DIAGNOSE 0x0318 instruction, unique to s390x, is a privileged call >>>>>>> that must be intercepted via SIE, handled in userspace, and the >>>>>>> information set by the instruction is communicated back to KVM. >>>>>>> >>>>>>> To test the instruction interception, an ad-hoc handler is defined which >>>>>>> simply has a VM execute the instruction and then userspace will extract >>>>>>> the necessary info. The handler is defined such that the instruction >>>>>>> invocation occurs only once. It is up to the caller to determine how the >>>>>>> info returned by this handler should be used. >>>>>>> >>>>>>> The diag318 info is communicated from userspace to KVM via a sync_regs >>>>>>> call. This is tested During a sync_regs test, where the diag318 info is >>>>>>> requested via the handler, then the info is stored in the appropriate >>>>>>> register in KVM via a sync registers call. >>>>>>> >>>>>>> If KVM does not support diag318, then the tests will print a message >>>>>>> stating that diag318 was skipped, and the asserts will simply test >>>>>>> against a value of 0. >>>>>>> >>>>>>> Signed-off-by: Collin Walling <walling@xxxxxxxxxxxxx> >>>>>> >>>>>> Interestingly enough, this testcase actually trigger a bug: >>>>>> While we gracefully handle this (no crash) >>>>>> debugfs: Directory 'kvm-200206' with parent 's390dbf' already present! >>>>>> is certainly not ideal.... >>>>>> >>>>> >>>>> Odd... I wonder what triggered this behavior? >>>>> >>>>> I run my tests with a simple command: >>>>> >>>>> make summary=0 TARGETS=kvm kselftest >>>>> >>>>> This must have something to do with spinning up another VM to get the >>>>> diag318 data. I think if I have the sync_regs test call the diag handler >>>>> first, and then have the sync regs create a VM, that might solve that >>>>> issue... >>>> >>>> Yes, the s390dbf code will try to create a file named kvm-%pid. With >>>> 2 VMs the 2nd one fails. Luckily the kvm will be created anyway and >>>> also the shutdown seems to be fine, still.... >>>> >>>>> >>>>> May I ask how you encountered this bug so I may replicate in on my end? >>>> >>>> I just did >>>> make TARGETS=kvm selftests >>>> >>>> and then the error is on dmesg. >>>> >>> >>> Thanks. v5 with fix incoming. >> >> I think the test is actually fine and we should rather fix the kvm module to >> gracefully handle a userspace that starts up 2 or more guests. >> > > Looks like the root cause is within the inode.c file used for the debug > filesystem. Essentially, the 2nd VM starts / ends just fine as you > mentioned, but doesn't get a dbfs. > > Since this touches the dbfs related area, and I'm unsure how common this > problem is out in the wild, should we ping the kernel list and see if it > catches anyone's attention? > > A first thought would be to append a per-userspace incrementing value to > the end of the kvm-%pid part to account for any collisions, but that's > up to the folks that know more than I do. > Seems like I have a deja-vu, I remember having to fix that for the common code :-) For the KVM debugfs I solved that by appending the file descriptor after the pid. Maybe we can steal the string from the dentry. Have a look into vm_create_vm_debugfs(struct kvm *kvm, int fd).