There is a narrow race possible:
A: lookup foo
A: put on foo
A: foo --nref == 0
B: lookup foo
B: put foo
B: foo --nref == 0
B: try_remove() succeeds, removes
A: try_remove() tries to remove foo again, probably crashes
We could fix this by flagging the object in some way to indicate it was
removed (maybe clearing parent?), but then we need to be careful about
dereferencing foo to get parent from put().
Fix this by moving to a simpler model: make lookup fail if nref == 0.
This eliminates the races around put() entirely because once nref reaches
0 it never goes up again.
Fixes: http://tracker.ceph.com/issues/24211
Signed-off-by: Sage Weil <sage@redhat.com>
(cherry picked from commit
8c8944b2c45ca9dc5b8fd4db1590e1d24206c0b3)
<< " removing self from set " << get_parent()
<< dendl;
if (get_parent()) {
- if (get_parent()->try_remove(this)) {
- delete this;
- } else {
- ldout(coll->store->cct, 20)
- << __func__ << " " << this << " lost race to remove myself from set"
- << dendl;
- }
- } else {
- delete this;
+ get_parent()->remove_last(this);
}
+ delete this;
}
}
SharedBlobRef lookup(uint64_t sbid) {
std::lock_guard<std::mutex> l(lock);
auto p = sb_map.find(sbid);
- if (p == sb_map.end()) {
+ if (p == sb_map.end() ||
+ p->second->nref == 0) {
return nullptr;
}
return p->second;
sb->coll = coll;
}
- bool try_remove(SharedBlob *sb) {
+ void remove_last(SharedBlob *sb) {
std::lock_guard<std::mutex> l(lock);
- if (sb->nref == 0) {
- assert(sb->get_parent() == this);
- sb_map.erase(sb->get_sbid());
- return true;
- }
- return false;
+ assert(sb->nref == 0);
+ assert(sb->get_parent() == this);
+ sb_map.erase(sb->get_sbid());
}
void remove(SharedBlob *sb) {