osd: Allow recovery to send additional reads

For now it doesn't include non-acting OSDs
Added test for this case

Signed-off-by: David Zafman <dzafman@redhat.com>

diff --git a/qa/standalone/erasure-code/test-erasure-eio.sh b/qa/standalone/erasure-code/test-erasure-eio.sh
index 49f61cd70eed103392015c3894e0bd0d5cb918b8..7d26d70a1c0a2a2801bc562bad3dd7e7ebebb11b 100755 (executable)
--- a/qa/standalone/erasure-code/test-erasure-eio.sh
+++ b/qa/standalone/erasure-code/test-erasure-eio.sh
@@ -380,6 +380,34 @@ function TEST_rados_get_with_subreadall_eio_shard_1() {
     delete_pool $poolname
 }
 
+# Test recovery the first k copies aren't all available
+function TEST_ec_recovery_errors() {
+    local dir=$1
+    local objname=myobject
+
+    setup_osds 7 || return 1
+
+    local poolname=pool-jerasure
+    create_erasure_coded_pool $poolname 3 2 || return 1
+
+    rados_put $dir $poolname $objname || return 1
+    inject_eio ec data $poolname $objname $dir 0 || return 1
+
+    local -a initial_osds=($(get_osds $poolname $objname))
+    local last_osd=${initial_osds[-1]}
+    # Kill OSD
+    kill_daemons $dir TERM osd.${last_osd} >&2 < /dev/null || return 1
+    ceph osd down ${last_osd} || return 1
+    ceph osd out ${last_osd} || return 1
+
+    # Cluster should recover this object
+    wait_for_clean || return 1
+
+    #rados_get_data_recovery eio $dir $shard_id || return 1
+
+    delete_pool $poolname
+}
+
 main test-erasure-eio "$@"
 
 # Local Variables:

diff --git a/src/osd/ECBackend.cc b/src/osd/ECBackend.cc
index 34b2f0f3e769c9200f561220c008679b4d84baf6..caadeb92ba1c64d39f8b7128ed6d45401c92e05c 100644 (file)
--- a/src/osd/ECBackend.cc
+++ b/src/osd/ECBackend.cc
@@ -1187,8 +1187,7 @@ void ECBackend::handle_sub_read_reply(
   unsigned is_complete = 0;
   // For redundant reads check for completion as each shard comes in,
   // or in a non-recovery read check for completion once all the shards read.
-  // TODO: It would be nice if recovery could send more reads too
-  if (rop.do_redundant_reads || (!rop.for_recovery && rop.in_progress.empty())) {
+  if (rop.do_redundant_reads || rop.in_progress.empty()) {
     for (map<hobject_t, read_result_t>::const_iterator iter =
         rop.complete.begin();
       iter != rop.complete.end();
@@ -1204,6 +1203,7 @@ void ECBackend::handle_sub_read_reply(
       set<int> want_to_read, dummy_minimum;
       get_want_to_read_shards(&want_to_read);
       int err;
+      // TODO: Should we include non-acting nodes here when for_recovery is set?
       if ((err = ec_impl->minimum_to_decode(want_to_read, have, &dummy_minimum)) < 0) {
        dout(20) << __func__ << " minimum_to_decode failed" << dendl;
         if (rop.in_progress.empty()) {