On Fri, Sep 07, 2018 at 09:05:43PM +0300, Leon Romanovsky wrote: > On Fri, Sep 07, 2018 at 08:40:10PM +0300, Leon Romanovsky wrote: > > On Thu, Sep 06, 2018 at 09:55:19PM -0600, Jason Gunthorpe wrote: > > > On Wed, Sep 05, 2018 at 05:21:37PM -0600, Jason Gunthorpe wrote: > > > > > > > Instead hold on to the actual mm directly inside the umme via mmgrab() > > > > and mmdrop(), just like mmu_notifiers already does internally. > > > > > > I coded up a series to do this, and more: > > > > > > https://github.com/jgunthorpe/linux/commits/tgid_removal > > > > > > I'll try to test it later, but it is the general idea.. ucontext->tgid > > > is an abomination and needs to be deleted. > > > > > > Have to do some testing on it.. > > > > I tried the series with my repro for use-after-free bug in ODP plus reverted > > commit "50704e039ab1 RDMA/umem: Restore lockdep check while downgrading lock" > > just to be sure and got the following splat. I have similar lockdep > > warning without reverting too. > > Something changed in my test system, I'm receiving such warning without > your patches, with reverted "lockdep" patch and reverted this ODP patch. I would expect reg_mr/dereg_mr to instantly cause this splat very reliably.. Assuming the tgid madness is working. Jason
