Tomasz Chmielewski posted on Mon, 21 Oct 2013 13:36:44 +0900 as excerpted: >> > I'll do some more tests with lots of extents to see if it's >> > reproducible here as well. >> >> Interestingly, I've generally had qgroups enabled here as well, >> possibly on all of these systems. Could that be the culprit? > > Possibly. > > Is there a reliable way to create files with many extents? I've tried > playing with fallocate, but couldn't create files with too many extents > with it. While the absense of both qgroups and the problem here doesn't prove anything, it's a datapoint. However, I /am/ a bit suspicious of them, as it just seems people keep reporting weird problems and the common thread often seems to be qgroups. I'm not a coder let alone a kernel/btrfs coder so take this with a suitably sized grain of salt, but my intuition says qgroups add more stress to the system and likely some as yet unresolved race-points, and they raise the possibility of triggering otherwise unrelated problems that already exist, along with any issues they themselves might have. That would fit with the many-thousands-of-extents issues here as well, since especially on spinning rust that's going to create all sorts of stress on the drive and filesystem and all sorts of potential to trigger race conditions. If both of them increase stress and open up the critical sections during which races are possible, putting the two together is an effective torture-test likely indeed to bring out otherwise extremely small window race conditions that might otherwise very seldom trigger. Meanwhile, on the extents issue I believe creating a large enough file and then writing into random locations within it, saving, writing some more, should do it, given btrfs' copy-on-write defaults. Working with an almost full filesystem is likely to create worse fragmentation too, since at some point btrfs just gives up trying to create single-extent files and writes anywhere it can find space. Finally, btrfs compression is said to look like fragmentation to filefrag. I'm not sure whether it /acts/ like fragmentation or not, but any suitably large file that gets btrfs compression will look fragmented I believe because the compression blocks (128KB IIRC) are reported to filefrag as individual extents due to the way btrfs handles and tracks them. -- Duncan - List replies preferred. No HTML msgs. "Every nonfree program has a lord, a master -- and if you use the program, he is your master." Richard Stallman -- To unsubscribe from this list: send the line "unsubscribe linux-btrfs" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
