Hi! > So not to dispute that this could be a bug, but I couldn't cause a > deadlock. I wonder if you can see something wrong with my method? > > *** Set up *** > > - kernel 5.0.0-0.rc3.git0.1.fc30.x86_64 > - nbd-client 3.19-1.fc30 > - nbdkit 1.11.5 (git commit ef9d1978ce28) > > Baremetal machine was booted with mem=2G to artificially limit the > RAM. The machine has 64G of swap. > > # free -m > total used free shared buff/cache > available > Mem: 1806 329 1383 0 93 > 1357 > Swap: 65535 179 65356 > > *** Method *** > > I started nbdkit as a 4G RAM disk: > > ./nbdkit memory size=4G > > This is implemented as a sparse array with a 2 level page table, and > should allocate (using malloc) every time a new area of the disk is > written to. Exact implementation is here: > https://github.com/libguestfs/nbdkit/tree/master/common/sparse > > I started nbd-client using the -swap option which uses > mlockall(MCL_FUTURE) to lock the client into RAM. > > nbd-client -b 512 -swap localhost /dev/nbd0 > > I then created a filesystem on the RAM disk, mounted it, and copied a > 3G file into it. I tried this various ways, but the variation I was > eventually happy with was: > > mke2fs /dev/nbd0 > mount /dev/nbd0 /tmp/mnt > > dd if=/dev/zero of=/tmp/big bs=1M count=3000 > cp /tmp/big /tmp/mnt/big > > I couldn't get any kind of deadlock or failure in this test. > > (Note that if you repeat the test several times, in theory you could > delete the file and fstrim the filesystem, but when I was testing it > to be sure I unmounted everything and killed and restarted nbdkit > between each test.)
This looks like quite a good try. I'd try to use mmap() to dirty
memory very quickly.
But Shaun reported it happened somehow often for them, so he might
have a practical test case... better than my theories :-).
Best regards,
Pavel
--
(english) http://www.livejournal.com/~pavelmachek
(cesky, pictures)
http://atrey.karlin.mff.cuni.cz/~pavel/picture/horses/blog.html
signature.asc
Description: Digital signature
