On Fri, Mar 15, 2019 at 01:26:04PM +1100, Dave Chinner wrote: > On Thu, Mar 14, 2019 at 12:34:51AM -0700, Dan Williams wrote: > > On Mon, Mar 11, 2019 at 9:38 PM Dave Chinner <david@xxxxxxxxxxxxx> wrote: > > > > > > On Mon, Mar 11, 2019 at 08:35:05PM -0700, Dan Williams wrote: > > > > On Mon, Mar 11, 2019 at 8:10 AM Matthew Wilcox <willy@xxxxxxxxxxxxx> wrote: > > > > > > > > > > On Thu, Mar 07, 2019 at 10:16:17PM -0800, Dan Williams wrote: > > > > > > Hi Willy, > > > > > > > > > > > > We're seeing a case where RocksDB hangs and becomes defunct when > > > > > > trying to kill the process. v4.19 succeeds and v4.20 fails. Robert was > > > > > > able to bisect this to commit b15cd800682f "dax: Convert page fault > > > > > > handlers to XArray". > > > > > > > > > > > > I see some direct usage of xa_index and wonder if there are some more > > > > > > pmd fixups to do? > > > > > > > > > > > > Other thoughts? > > > > > > > > > > I don't see why killing a process would have much to do with PMD > > > > > misalignment. The symptoms (hanging on a signal) smell much more like > > > > > leaving a locked entry in the tree. Is this easy to reproduce? Can you > > > > > get /proc/$pid/stack for a hung task? > > > > > > > > It's fairly easy to reproduce, I'll see if I can package up all the > > > > dependencies into something that fails in a VM. > > > > > > > > It's limited to xfs, no failure on ext4 to date. > > > > > > > > The hung process appears to be: > > > > > > > > kworker/53:1-xfs-sync/pmem0 > > > > > > That's completely internal to XFS. Every 30s the work is triggered > > > and it either does a log flush (if the fs is active) or it syncs the > > > superblock to clean the log and idle the filesystem. It has nothing > > > to do with user processes, and I don't see why killing a process has > > > any effect on what it does... > > > > > > > ...and then the rest of the database processes grind to a halt from there. > > > > > > > > Robert was kind enough to capture /proc/$pid/stack, but nothing interesting: > > > > > > > > [<0>] worker_thread+0xb2/0x380 > > > > [<0>] kthread+0x112/0x130 > > > > [<0>] ret_from_fork+0x1f/0x40 > > > > [<0>] 0xffffffffffffffff > > > > > > Much more useful would be: > > > > > > # echo w > /proc/sysrq-trigger > > > > > > And post the entire output of dmesg. > > > > Here it is: > > > > https://gist.github.com/djbw/ca7117023305f325aca6f8ef30e11556 > > Which tells us nothing. :( Nothing from a filesystem side, perhaps, but I find it quite interesting. We have a number of threads blocking in down_read() on mmap_sem. That means a task is holding the mmap_sem for write, or is blocked trying to take the mmap_sem for write. I think it's the latter; pid 4650 is blocked in munmap(). pid 4673 is blocking in get_unlocked_entry() and will be holding the mmap_sem for read while doing so. Since this is provoked by a fatal signal, it must have something to do with a killable or interruptible sleep. There's only one of those in the DAX code; fatal_signal_pending() in dax_iomap_actor(). Does rocksdb do I/O with write() or through a writable mmap()? I'd like to know before I chase too far down this fault tree analysis. My current suspicion is that we have a PMD fault being not-woken by a PTE modification, and the evidence seems to fit, but I don't quite see it yet. (I meant to ask Dan about this while we were in San Juan, but with all the other excitement, it slipped my mind).
