On Mon, 2017-06-26 at 19:38 +0200, Stephan Müller wrote: > Am Montag, 26. Juni 2017, 03:23:09 CEST schrieb Nicholas A. Bellinger: > > Hi Nicholas, > > > Hi Stephan, Lee & Jason, > > > > (Adding target-devel CC') > > > > Apologies for coming late to the discussion. Comments below. > > > > On Sun, 2017-06-18 at 10:04 +0200, Stephan Müller wrote: > > > Am Samstag, 17. Juni 2017, 05:45:57 CEST schrieb Lee Duncan: > > > > > > Hi Lee, > > > > > > > In your testing, how long might a process have to wait? Are we talking > > > > seconds? Longer? What about timeouts? > > > > > > In current kernels (starting with 4.8) this timeout should clear within a > > > few seconds after boot. > > > > > > In older kernels (pre 4.8), my KVM takes up to 90 seconds to reach that > > > seeding point. I have heard that on IBM System Z this trigger point > > > requires minutes to be reached. > > > > I share the same concern as Lee wrt to introducing latency into the > > existing iscsi-target login sequence. > > > > Namely in the use-cases where a single node is supporting ~1K unique > > iscsi-target IQNs, and fail-over + re-balancing of IQNs where 100s of > > login attempts are expected to occur in parallel. > > > > If environments exist that require non trivial amounts of time for RNG > > seeding to be ready for iscsi-target usage, then enforcing this > > requirement at iscsi login time can open up problems, especially when > > iscsi host environments may be sensitive to login timeouts, I/O > > timeouts, et al. > > > > That said, I'd prefer to simply wait for RNG to be seeded at modprobe > > iscsi_target_mod time, instead of trying to enforce randomness during > > login. > > > > This way, those of use who have distributed storage platforms can know > > to avoid putting a node into a state to accept iscsi-target IQN export > > migration, before modprobe iscsi_target_mod has successfully loaded and > > RNG seeding has been confirmed. > > > > WDYT..? > > We may have a chicken and egg problem when the wait is at the modprobe time. > Assume the modprobe happens during initramfs time to get access to the root > file system. In this case, you entire boot process will lock up for an > indefinite amount of time. The reason is that in order to obtain events > detected by the RNG, devices need to be initialized and working. Such devices > commonly start working after the the root partition is mounted as it contains > all drivers, all configuration information etc. > > Note, during the development of my /dev/random implementation, I added the > getrandom-like blocking behavior to /dev/urandom (which is the equivalent to > Jason's patch except that it applies to user space). The boot process locked > up since systemd wanted data from /dev/urandom while it processed the > initramfs. As it did not get any, the boot process did not commence that could > deliver new events to be picked up by the RNG. > > As I do not have such an iscsi system, I cannot test Jason's patch. But maybe > the mentioned chicken-and-egg problem I mentioned above is already visible > with the current patch as it will lead to a blocking of the mounting of the > root partition in case the root partition is on an iscsi target. AFAIK, there are no distro initramfs dependencies for iscsi_target_mod, and every environment I've ever seen loads iscsi_target_mod after switching to the real rootfs. For an iscsi initiator that might not been the case, especially if the rootfs is running atop a iscsi LUN. But at least for iscsi-target mode, any blocking during modprobe waiting for RNG seeding would happen outside of initramfs.
