On Thu, Jun 13, 2019 at 09:25:54AM -0700, Dan Williams wrote: > On Thu, Jun 13, 2019 at 8:14 AM Jason Gunthorpe <jgg@xxxxxxxx> wrote: > > > > On Wed, Jun 12, 2019 at 06:14:46PM -0700, Dan Williams wrote: > > > > Effectively, we would need a way for an admin to close a specific file > > > > descriptor (or set of fds) which point to that file. AFAIK there is no way to > > > > do that at all, is there? > > > > > > Even if there were that gets back to my other question, does RDMA > > > teardown happen at close(fd), or at final fput() of the 'struct > > > file'? > > > > AFAIK there is no kernel side driver hook for close(fd). > > > > rdma uses a normal chardev so it's lifetime is linked to the file_ops > > release, which is called on last fput. So all the mmaps, all the dups, > > everything must go before it releases its resources. > > Oh, I must have missed where this conversation started talking about > the driver-device fd. In the first paragraph above where Ira is musing about 'close a specific file', he is talking about the driver-device fd. Ie unilaterally closing /dev/uverbs as a punishment for an application that used leases wrong: ie that released its lease with the RDMA is still ongoing. Jason
