On Wed, 2017-05-10 at 15:30 +0200, Miklos Szeredi wrote: > On Wed, May 10, 2017 at 3:20 PM, Jeff Layton <jlayton@xxxxxxxxxx> wrote: > > On Wed, 2017-05-10 at 09:05 +0100, David Howells wrote: > > > Miklos Szeredi <mszeredi@xxxxxxxxxx> wrote: > > > > > > > Possible rule of thumb: use it only at the place where the error > > > > originates and not where errors are just passed on. This would result > > > > in at most one report per syscall, normally. > > > > > > > > That might be hard to enforce in practice once you get into some > > complicated layering. What if we have device_mapper setting this along > > with filesystems too? We need clear rules here. > > If the error originates in the devicemapper, then why would the > filesystem set it? > > There's always a root cause of an error and that should be where the > detailed error is set. > > Am I missing something? > I was thinking that you'd need some well-defined way to tell whether the string should be replaced. If the thing just hangs out across syscalls, then you don't know when it got put there. Is it a leftover from a previous syscall or did a lower layer just put it there? But...maybe I'm making assumptions about how this would work and I should just wait until there are patches in flight. Getting the lifetime of these strings right will be crucial though. > > > > > > And the static string thing that David implemented is also a very good > > > > idea, IMO. > > > > > > There is an issue with it: it's fine as long as you keep a ref on the module > > > that generated it or clear all strings as part of module removal (which the > > > mount context in this patchset does). With the NFS mount context I did, I > > > have to keep a ref on the NFS protocol module as well as the NFS filesystem > > > module. > > > > > > I'm tempted to make it conditionally copy the string using kvasprintf_const() > > > - which would also permit format substitution. > > > > > > > On balance, I think this is a reasonable way to pass back detailed > > errors. Up until now, we've mostly relied on just printk'ing them. Now > > though, a lot of larger machines are running containerized setups. Good > > luck scraping dmesg for _your_ error in that situation. There may be > > tons of mounts failing all over the place. > > > > That said, I have some concerns here: > > > > What's the lifetime of these strings? Do they just hang around forever > > until the process goes away or they're replaced? If this becomes common, > > then you could easily end up with an extra string allocation per task in > > some cases. That could add up. > > That's why I liked the static string thing. It's just one assignment > and no worries about freeing. Not sure what to do about modules, > though. Can we somehow move the cost of checking the validity to the > place where the error is retrieved? > Seems a little dangerous, and could be limiting. Dynamically allocated strings seem like they could be more useful. > > > > One idea might be to always kfree it on syscall entry, and that might > > mitigate the problem assuming that not everything is erroring out. Then > > you could always do some trivial syscall to clear it manually. > > > > There's also the problem of how these should be formatted. Is English ok > > everywhere? Do we need a facility to allow translating these things? > > Messages in dmesg are in English too. If necessary userspace will do > the translation. I don't think the kernel would need to worry about > that. Fair enough. It _is_ still an improvement over dmesg, IMO. -- Jeff Layton <jlayton@xxxxxxxxxx>
