From: Mingxin Liu <mingxin@xsky.com>
Date: Thu, 22 Jul 2021 04:21:53 +0000 (+0800)
Subject: osd/PGLog: set acceptable rollback_info_trimmed_to for pg of replicated pool
X-Git-Tag: v17.1.0~1236^2
X-Git-Url: http://git-server-git.apps.pok.os.sepia.ceph.com/?a=commitdiff_plain;h=refs%2Fpull%2F42445%2Fhead;p=ceph.git

osd/PGLog: set acceptable rollback_info_trimmed_to for pg of replicated pool

if this PG belongs to a replica pool, since replica pool does not
require rollback, see `pg_pool_t::require_rollback()`, we don't
update the omap entry for noting down the
`on_disk_rollback_info_trimmed_to`. see
`PGLog::_write_log_and_missing()` and `PGLog::_write_log_and_missing()`.
in this case, `read_log_and_missing()` should not be able to find
the value for "rollback_info_trimmed_to" in the omap entries.

because `roll_foward_to()` replays all entries starting at
`rollback_info_trimmed_to`, if it is empty, when the primary osd
is merging the pglog from its peers while peering, it would have
to walk through *all* pglog entries. this would incur unnecessary
overhead, and slows down the peering.

in this change, if this PG belongs to a replica pool, we just fast
forward it to `info.last_update` to avoid this overhead.

Signed-off-by: Mingxin Liu <mingxin@xsky.com>
---

diff --git a/src/osd/PGLog.cc b/src/osd/PGLog.cc
index be3f7414d69..c5e0a10ea96 100644
--- a/src/osd/PGLog.cc
+++ b/src/osd/PGLog.cc
@@ -1095,6 +1095,11 @@ namespace {
           return iter->next();
         });
       }).then([this] {
+        if (info.pgid.is_no_shard()) {
+          // replicated pool pg does not persist this key
+          assert(on_disk_rollback_info_trimmed_to == eversion_t());
+          on_disk_rollback_info_trimmed_to = info.last_update;
+        }
         log = PGLog::IndexedLog(
              info.last_update,
              info.log_tail,
diff --git a/src/osd/PGLog.h b/src/osd/PGLog.h
index 920b40dabc1..25c88217c51 100644
--- a/src/osd/PGLog.h
+++ b/src/osd/PGLog.h
@@ -1490,6 +1490,11 @@ public:
 	}
       }
     }
+    if (info.pgid.is_no_shard()) {
+      // replicated pool pg does not persist this key
+      assert(on_disk_rollback_info_trimmed_to == eversion_t());
+      on_disk_rollback_info_trimmed_to = info.last_update;
+    }
     log = IndexedLog(
       info.last_update,
       info.log_tail,