On Sun, Mar 18, 2012 at 3:25 PM, Mandeep Singh Baines <msb@xxxxxxxxxxxx> wrote: > On Sun, Mar 18, 2012 at 1:23 PM, Alan Stern <stern@xxxxxxxxxxxxxxxxxxx> wrote: >> On Sat, 17 Mar 2012, Ted Ts'o wrote: >> >>> I can't help thinking that the fact that we're constantly playing >>> whack-a-mole trying to fix various random crashes when devices >>> disappear that perhaps we should consider if there's a better way to >>> do things. >> >> Indeed, as Jens's patch mentions, proper reference counting for the BDI >> stuff hasn't been implemented yet. Obviously it will require somebody >> who really does know the code (i.e., not me). >> >> For example, when Paul's patch assigns &default_backing_dev_info, is >> the assignment synchronized by any sort of lock? I can't tell -- but >> if it isn't then the possibility of a race will still exist. >> > > I think its safe without a lock (assuming the assignment is atomic) but it > wouldn't hurt to add an i_lock. That would also give you a barrier which > is needed to propagate the assignment to other CPUs. > > This is not a perfect fix but its pretty safe and is nice in that it works > independent of filesystem or bus-type. > > Regards, > Mandeep > >>> The fact that at the file system layer I have **no** idea that a >>> device has disappeared, and just blindly going on trying to write to a >>> device which is gone just seems a little crazy to me... why shouldn't >>> block layer inform the upper layers about something as fundamental as, >>> "the device is gone and is never coming back"? >> >> Playing devil's advocate... What would you do differently if you did >> know the device was gone? All I/O operations will fail regardless, and >> presumably with an error code like -ENODEV. Pretty much all you could >> do would be to fail them a little earlier. >> >>> > I suspect Paul's patch is the right thing to do. It might even make >>> > the ext4 fix unnecessary, although I don't understand the details well >>> > enough to verify it. Maybe Paul can check -- the commit I'm referring >>> > to is 7c2e70879fc0949b4220ee61b7c4553f6976a94d (ext4: add ext4-specific >>> > kludge to avoid an oops after the disk disappears). >>> >>> I have no idea either, because it's not obvious to me what data >>> structures can be relied upon, and what can't, and when things are >>> supposed to get freed on sudden device disconnects. The fact that >>> none of us are sure is part of what makes me think that the current >>> scheme is, perhaps, non-optimal... >> >> That's why someone like Jens or Al needs to take a close look at this >> (hint, hint). >> >> Alan Stern >> I have rerun my tests without my change on the 3.2.7 kernel and I was not able to get it to crash. I even put some code in to do the early detection so I didn't have to wait for another thread to stumble across the corruption. The way I test is with several flash drivers with ext2, ext3, ext4, FAT, and HPFS file systems and just repeatedly plug and unplug them. When a flash drive is plugged in with a file system, it is automatically mounted. Paul Taysom -- To unsubscribe from this list: send the line "unsubscribe linux-usb" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
