On Fri, 9 Dec 2016, xxhdx1985126 wrote:
> Thanks for the quick reply:-)
>
> By the way, according to my understanding of the source code, the ondisk
> callback is trigger both when journal write complete or syscall(SYS_syncfs,
> fd) complete. Is this right?
The syncfs path only happens if you're using btrfs ('parallel' mode) since
with xfs we don't even write to the fs until the journaled event has
committed.
s
>
>
>
> At 2016-12-07 22:29:55, "Sage Weil" <sage@xxxxxxxxxxxx> wrote:
> >On Wed, 7 Dec 2016, xxhdx1985126 wrote:
> >> Hi, everyone.
> >>
> >>
> >> I'm trying to read the source code of ceph, and I found that some code se
> ems strange.
> >> In ReplicatedBackend::sub_op_modify_reply, there is the following code:
> >> if (r->ack_type & CEPH_OSD_FLAG_ONDISK) {
> >> assert(ip_op.waiting_for_commit.count(from));
> >> ip_op.waiting_for_commit.erase(from);
> >> if (ip_op.op) {
> >> ostringstream ss;
> >> ss << "sub_op_commit_rec_from_osd." << from.osd;
> >> ip_op.op->mark_event(ss.str());
> >> }
> >> } else {
> >> assert(ip_op.waiting_for_applied.count(from));
> >> if (ip_op.op) {
> >> ostringstream ss;
> >> ss << "sub_op_applied_rec_from_osd." << from.osd;
> >> ip_op.op->mark_event(ss.str());
> >> }
> >> }
> >> ip_op.waiting_for_applied.erase(from);
> >> It seems that the statement "ip_op.waiting_for_applied.erase(from)" shoul
> d be in the "else" clause, otherwise items in waiting_for_applied could be e
> rrorly erased when ack_type is CEPH_OSD_FLAG_ONDISK.
> >>
> >> Why put it outside the "else" clause?
> >
> >The commit/ondisk message implies applied.
> >
> >sage
>
>
>
>
>
>
