RepoMirrors/ceph
1
0
Fork 0
mirror of https://github.com/ceph/ceph synced 2025-01-31 07:22:56 +00:00
Code Issues Packages Projects Releases Wiki Activity

Merge pull request #12779 from badone/wip-check_new_interval-comment

Browse Source 
osd/osd_types: Move comment to more relevant position

Reviewed-by: Samuel Just <sjust@redhat.com>
This commit is contained in:
Samuel Just 2017-01-04 14:03:19 -08:00 committed by GitHub
commit 5ead0588e8
2 changed files with 41 additions and 42 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

42
src/osd/PG.cc
View File

@ -7859,48 +7859,6 @@ PG::PriorSet::PriorSet(CephContext* cct,
const PG *debug_pg)
: cct(cct), ec_pool(ec_pool), pg_down(false), pcontdec(c)
{
/*
* We have to be careful to gracefully deal with situations like
* so. Say we have a power outage or something that takes out both
* OSDs, but the monitor doesn't mark them down in the same epoch.
* The history may look like
*
* 1: A B
* 2: B
* 3: let's say B dies for good, too (say, from the power spike)
* 4: A
*
* which makes it look like B may have applied updates to the PG
* that we need in order to proceed. This sucks...
*
* To minimize the risk of this happening, we CANNOT go active if
* _any_ OSDs in the prior set are down until we send an MOSDAlive
* to the monitor such that the OSDMap sets osd_up_thru to an epoch.
* Then, we have something like
*
* 1: A B
* 2: B up_thru[B]=0
* 3:
* 4: A
*
* -> we can ignore B, bc it couldn't have gone active (alive_thru
* still 0).
*
* or,
*
* 1: A B
* 2: B up_thru[B]=0
* 3: B up_thru[B]=2
* 4:
* 5: A
*
* -> we must wait for B, bc it was alive through 2, and could have
* written to the pg.
*
* If B is really dead, then an administrator will need to manually
* intervene by marking the OSD as "lost."
*/
// Include current acting and up nodes... not because they may
// contain old data (this interval hasn't gone active, obviously),
// but because we want their pg_info to inform choose_acting(), and

41
src/osd/osd_types.cc
View File

@ -3099,6 +3099,47 @@ bool pg_interval_t::check_new_interval(
map<epoch_t, pg_interval_t> *past_intervals,
std::ostream *out)
{
/*
* We have to be careful to gracefully deal with situations like
* so. Say we have a power outage or something that takes out both
* OSDs, but the monitor doesn't mark them down in the same epoch.
* The history may look like
*
* 1: A B
* 2: B
* 3: let's say B dies for good, too (say, from the power spike)
* 4: A
*
* which makes it look like B may have applied updates to the PG
* that we need in order to proceed. This sucks...
*
* To minimize the risk of this happening, we CANNOT go active if
* _any_ OSDs in the prior set are down until we send an MOSDAlive
* to the monitor such that the OSDMap sets osd_up_thru to an epoch.
* Then, we have something like
*
* 1: A B
* 2: B up_thru[B]=0
* 3:
* 4: A
*
* -> we can ignore B, bc it couldn't have gone active (up_thru still 0).
*
* or,
*
* 1: A B
* 2: B up_thru[B]=0
* 3: B up_thru[B]=2
* 4:
* 5: A
*
* -> we must wait for B, bc it was alive through 2, and could have
* written to the pg.
*
* If B is really dead, then an administrator will need to manually
* intervene by marking the OSD as "lost."
*/
// remember past interval
// NOTE: a change in the up set primary triggers an interval
// change, even though the interval members in the pg_interval_t