os/bluestore: Expand bluestore_onode_t, v=2 -> v=3

Add segment_size field for bluestore_onode_t.
It makes more reliable when onode keeps its segment size.
Upgraded v2 to v3.

Object creation on v3 (new BlueStore):
Object gets its segment_size field initialized from bluestore_onode_segment_size.
However, if pool opt compression_max_blob_size is set and it is larger, it will be used.

Upgrade:
Object upgraded has its segment_size = 0 by default, so new BlueStore
does not use segmentation. Spanning blobs can be created.

Downgrade:
Object downgraded (by being written by older BlueStore) is losing its segment_size setting.
Older BlueStore will have no problem understanding it.

Signed-off-by: Adam Kupczyk <akupczyk@ibm.com>

diff --git a/src/os/bluestore/BlueStore.cc b/src/os/bluestore/BlueStore.cc
index b47bb61ef7c799dc0bf6acc21f5ed0f550bd2e7a..b7fad7e75f3f5810390866b3cdd36d52d8115bd0 100644 (file)
--- a/src/os/bluestore/BlueStore.cc
+++ b/src/os/bluestore/BlueStore.cc
@@ -4860,6 +4860,13 @@ BlueStore::Onode* BlueStore::Onode::create_decode(
     } else {
       on->extent_map.init_shards(false, false);
     }
+  } else {
+    // init segment_size
+    uint32_t segment_size = c->store->segment_size.load();
+    if (c->comp_max_blob_size.has_value() && segment_size < c->comp_max_blob_size.value()) {
+      segment_size = c->comp_max_blob_size.value(); // compression larger than global segment_size, use it
+    }
+    on->onode.segment_size = segment_size;
   }
   return on;
 }
@@ -17418,7 +17425,7 @@ void BlueStore::_do_write_data(
     if (head_length) {
       _do_write_small(txc, c, o, head_offset, head_length, p, wctx);
     }
-    uint32_t segment_size = this->segment_size.load();
+    uint32_t segment_size = o->onode.segment_size;
     if (segment_size) {
       // split data to chunks
       uint64_t write_offset = middle_offset;
@@ -17708,8 +17715,8 @@ int BlueStore::_do_write_v2(
     // if we have compression, skip to write_v1
     return _do_write(txc, c, o, offset, length, bl, fadvise_flags);
   }
-  if (segment_size != 0 && wctx.target_blob_size > segment_size) {
-    wctx.target_blob_size = segment_size;
+  if (o->onode.segment_size != 0 && wctx.target_blob_size > o->onode.segment_size) {
+    wctx.target_blob_size = o->onode.segment_size;
   }
   if (bl.length() != length) {
     bl.splice(length, bl.length() - length);
@@ -19185,7 +19192,7 @@ void BlueStore::_record_onode(OnodeRef& o, KeyValueDB::Transaction &txn)
   // finalize extent_map shards
   o->extent_map.update(txn, false);
   if (o->extent_map.needs_reshard()) {
-    o->extent_map.reshard(db, txn, segment_size);
+    o->extent_map.reshard(db, txn, o->onode.segment_size);
     o->extent_map.update(txn, true);
     if (o->extent_map.needs_reshard()) {
       dout(20) << __func__ << " warning: still wants reshard, check options?"

diff --git a/src/os/bluestore/bluestore_types.h b/src/os/bluestore/bluestore_types.h
index a7dc6cc52f4af890ba92b0551eef096ad0da42a2..a05d2ae9c9702e432611928eaa51618d34bf128c 100644 (file)
--- a/src/os/bluestore/bluestore_types.h
+++ b/src/os/bluestore/bluestore_types.h
@@ -1135,6 +1135,7 @@ struct bluestore_onode_t {
   uint32_t expected_object_size = 0;
   uint32_t expected_write_size = 0;
   uint32_t alloc_hint_flags = 0;
+  uint32_t segment_size = 0; ///< mandatory segment lines to never cross; helps with sharding
 
   uint8_t flags = 0;
 
@@ -1214,7 +1215,7 @@ struct bluestore_onode_t {
   }
 
   DENC(bluestore_onode_t, v, p) {
-    DENC_START(2, 1, p);
+    DENC_START(3, 1, p);
     denc_varint(v.nid, p);
     denc_varint(v.size, p);
     denc(v.attrs, p);
@@ -1226,6 +1227,9 @@ struct bluestore_onode_t {
     if (struct_v >= 2) {
       denc(v.zone_offset_refs, p);
     }
+    if (struct_v >= 3) {
+      denc(v.segment_size, p);
+    }
     DENC_FINISH(p);
   }
   void dump(ceph::Formatter *f) const;

diff --git a/src/tools/ceph-dencoder/osd_types.h b/src/tools/ceph-dencoder/osd_types.h
index 749233e3b7d1203620f2b1d396025c8aa8e5603e..35bb5a45c9324517e8f7a61a9f00e2390accae2a 100644 (file)
--- a/src/tools/ceph-dencoder/osd_types.h
+++ b/src/tools/ceph-dencoder/osd_types.h
@@ -118,7 +118,7 @@ TYPE(bluestore_blob_use_tracker_t)
 // approach.
 // TYPE_FEATUREFUL(bluestore_blob_t)
 TYPE(bluestore_shared_blob_t)
-TYPE(bluestore_onode_t)
+TYPE_FEATUREFUL(bluestore_onode_t)
 TYPE(bluestore_onode_t::shard_info)
 using shard_info = bluestore_onode_t::shard_info;
 TYPE(shard_info)