Thanks for the blazing fast and detailed reply! Am I correct assuming that the jbd2 problem is only (mostly?) relevant in case of data=journal mode? > It also generally doesn't buy you much for most file system workloads, so it hasn't been high on my priority list to fix. Completely understandable. I wasn't really hunting for those juice performance bits, however it seemed like a nice optimization that's in the kernel for quite long, plus enabling it is just a bit flip, then why not? But then I haven't found anything conclusive on it. Thanks again for clearing things up! Pas On Sun, 15 Sep 2019 at 01:28, Theodore Y. Ts'o <tytso@xxxxxxx> wrote: > > On Sat, Sep 14, 2019 at 08:06:04PM +0200, Pas wrote: > > > > I hope this is the correct forum to ask these questions. (If not, then > > sorry for the noise! Though then could you recommend where to ask > > them?) > > There are some known issues with with the inline_data feature; in > particular, it violates some of the jbd2 rules about how to jorunal > data blocks. As such, bad things(tm) can happen on crash recovery. > For the most part it works OK for the original intended use case, was > for bigalloc file systems where there was a desire to handle small > directories more efficiently, which was how the developer who created > the feature used said feature. I didn't realize this particular > rather serious bug until after the feature went into the kernel, and > it's been on my todo list to fix; I just haven't had the time. > > It doesn't happen all that often; you need to start files that are > small enough such that they can fit in an inline_data file, and then > grow them via an appending write so that they need to be shifted to an > block allocated file, and then force an unclean shutdown at an > inopportune time. > > But yeah, there is a good reason why it's not a default-enabled > feature. It also generally doesn't buy you much for most file system > workloads, so it hasn't been high on my priority list to fix. > > Regards, > > - Ted
