Description of problem:

dd IO failed with "dd: failed to open ‘/mnt/fuse/fuse/client1/dir1/1M_1/magna021_000000104856’: Cannot send after transport endpoint shutdown"

dd script used during testing:
function n {
    printf "%s_%012d" "$(hostname)" "$1"
}
dir=/mnt/fuse/fuse/client1/dir1
time mkdir -p $dir/1M_1 && for ((i = 0; i < 2**20/10; i++)); do echo "$i" > /root/1M_1_CURRENT; dd status=none if=/dev/urandom of="$dir/1M_1/$(n "$i")" bs=4k count=1; done

Fuse assert:
2018-05-10 10:20:50.709658 7f96cadb90c0  0 ~Inode: leftover objects on inode 0x0x10000082b16
2018-05-10 10:20:50.726976 7f96cadb90c0 -1 /builddir/build/BUILD/ceph-12.2.1/src/client/Inode.cc: In function 'Inode::~Inode()' thread 7f96cadb90c0 time 2018-05-10 10:20:50.709671
/builddir/build/BUILD/ceph-12.2.1/src/client/Inode.cc: 27: FAILED assert(oset.objects.empty())

 ceph version 12.2.1-46.el7cp (b6f6f1b141c306a43f669b974971b9ec44914cb0) luminous (stable)
 1: (ceph::__ceph_assert_fail(char const*, char const*, int, char const*)+0x110) [0x55ea9e2857d0]
 2: (Inode::~Inode()+0x5e5) [0x55ea9e250865]
 3: (Client::put_inode(Inode*, int)+0x350) [0x55ea9e1cdda0]
 4: (Client::unlink(Dentry*, bool, bool)+0x11a) [0x55ea9e1d2cca]
 5: (Client::trim_dentry(Dentry*)+0x93) [0x55ea9e1d3433]
 6: (Client::trim_cache(bool)+0x328) [0x55ea9e1d38c8]
 7: (Client::tear_down_cache()+0x2eb) [0x55ea9e1e987b]
 8: (Client::~Client()+0x53) [0x55ea9e1e9ab3]
 9: (StandaloneClient::~StandaloneClient()+0x9) [0x55ea9e1ea1c9]
 10: (main()+0x534) [0x55ea9e19c3a4]
 11: (__libc_start_main()+0xf5) [0x7f96c7af23d5]
 12: (()+0x212bf3) [0x55ea9e1a5bf3]
 NOTE: a copy of the executable, or `objdump -rdS <executable>` is needed to interpret this.

--- begin dump of recent events ---


Version-Release number of selected component (if applicable):
Client ceph version: 12.2.1-46.el7cp
Cluster daemons are in 12.2.4-10.el7cp --> cluster upgraded from 12.2.1-46 to 12.2.4-10.el7cp version.

How reproducible:
1/1

Steps to Reproduce:
1. Configure cluster with MDS using 12.2.1-46.el7cp
2. Created around 800k inodes in the FS
3. Did cluster upgrade to 12.2.4-10.el7cp using MDS upgrade procedures manually
4. After deactivating the active MDS it was in stopping state, observed client hung and fuse assert.
Actual results:


Expected results:
Client IO should not fail

Additional info:
NA