rgw: utilities to support raw bucket index operations

and other related changes.

Signed-off-by: Yehuda Sadeh <yehuda@redhat.com>

diff --git a/src/rgw/rgw_rados.cc b/src/rgw/rgw_rados.cc
index 3e1229ed0eca103f0463abef1b33936f8b7108ec..e5cebb648351037111b19c445cfb672cf115dde1 100644 (file)
--- a/src/rgw/rgw_rados.cc
+++ b/src/rgw/rgw_rados.cc
@@ -5321,11 +5321,7 @@ int RGWRados::create_bucket(RGWUserInfo& owner, rgw_bucket& bucket,
       return ret;
 
     if (!pmaster_bucket) {
-      uint64_t iid = instance_id();
-      uint64_t bid = next_bucket_id();
-      char buf[get_zone_params().get_id().size() + 48];
-      snprintf(buf, sizeof(buf), "%s.%llu.%llu", get_zone_params().get_id().c_str(), (long long)iid, (long long)bid);
-      bucket.marker = buf;
+      create_bucket_id(&bucket.marker);
       bucket.bucket_id = bucket.marker;
     } else {
       bucket.marker = pmaster_bucket->marker;
@@ -6000,6 +5996,21 @@ int RGWRados::BucketShard::init(rgw_bucket& _bucket, rgw_obj& obj)
   return 0;
 }
 
+int RGWRados::BucketShard::init(rgw_bucket& _bucket, int sid)
+{
+  bucket = _bucket;
+  shard_id = sid;
+
+  int ret = store->open_bucket_index_shard(bucket, index_ctx, shard_id, &bucket_obj);
+  if (ret < 0) {
+    ldout(store->ctx(), 0) << "ERROR: open_bucket_index_shard() returned ret=" << ret << dendl;
+    return ret;
+  }
+  ldout(store->ctx(), 20) << " bucket index object: " << bucket_obj << dendl;
+
+  return 0;
+}
+
 
 /* Execute @handler on last item in bucket listing for bucket specified
  * in @bucket_info. @obj_prefix and @obj_delim narrow down the listing
@@ -7894,6 +7905,27 @@ int RGWRados::open_bucket_index_shard(rgw_bucket& bucket, librados::IoCtx& index
   return 0;
 }
 
+int RGWRados::open_bucket_index_shard(rgw_bucket& bucket, librados::IoCtx& index_ctx,
+                                      int shard_id, string *bucket_obj)
+{
+  string bucket_oid_base;
+  int ret = open_bucket_index_base(bucket, index_ctx, bucket_oid_base);
+  if (ret < 0)
+    return ret;
+
+  RGWObjectCtx obj_ctx(this);
+
+  // Get the bucket info
+  RGWBucketInfo binfo;
+  ret = get_bucket_instance_info(obj_ctx, bucket, binfo, NULL, NULL);
+  if (ret < 0)
+    return ret;
+
+  get_bucket_index_object(bucket_oid_base, binfo.num_shards,
+                          shard_id, bucket_obj);
+  return 0;
+}
+
 static void accumulate_raw_stats(const rgw_bucket_dir_header& header,
                                  map<RGWObjCategory, RGWStorageStats>& stats)
 {
@@ -11545,6 +11577,20 @@ int RGWRados::bi_get(rgw_bucket& bucket, rgw_obj& obj, BIIndexType index_type, r
   return 0;
 }
 
+void RGWRados::bi_put(ObjectWriteOperation& op, BucketShard& bs, rgw_cls_bi_entry& entry)
+{
+  cls_rgw_bi_put(op, bs.bucket_obj, entry);
+}
+
+int RGWRados::bi_put(BucketShard& bs, rgw_cls_bi_entry& entry)
+{
+  int ret = cls_rgw_bi_put(bs.index_ctx, bs.bucket_obj, entry);
+  if (ret < 0)
+    return ret;
+
+  return 0;
+}
+
 int RGWRados::bi_put(rgw_bucket& bucket, rgw_obj& obj, rgw_cls_bi_entry& entry)
 {
   BucketShard bs(this);
@@ -11554,11 +11600,7 @@ int RGWRados::bi_put(rgw_bucket& bucket, rgw_obj& obj, rgw_cls_bi_entry& entry)
     return ret;
   }
 
-  ret = cls_rgw_bi_put(bs.index_ctx, bs.bucket_obj, entry);
-  if (ret < 0)
-    return ret;
-
-  return 0;
+  return bi_put(bs, entry);
 }
 
 int RGWRados::bi_list(rgw_bucket& bucket, const string& obj_name, const string& marker, uint32_t max, list<rgw_cls_bi_entry> *entries, bool *is_truncated)
@@ -11578,6 +11620,29 @@ int RGWRados::bi_list(rgw_bucket& bucket, const string& obj_name, const string&
   return 0;
 }
 
+int RGWRados::bi_list(BucketShard& bs, const string& marker, uint32_t max, list<rgw_cls_bi_entry> *entries, bool *is_truncated)
+{
+  string filter_prefix;
+
+  int ret = cls_rgw_bi_list(bs.index_ctx, bs.bucket_obj, filter_prefix, marker, max, entries, is_truncated);
+  if (ret < 0)
+    return ret;
+
+  return 0;
+}
+
+int RGWRados::bi_list(rgw_bucket& bucket, int shard_id, const string& marker, uint32_t max, list<rgw_cls_bi_entry> *entries, bool *is_truncated)
+{
+  BucketShard bs(this);
+  int ret = bs.init(bucket, shard_id);
+  if (ret < 0) {
+    ldout(cct, 5) << "bs.init() returned ret=" << ret << dendl;
+    return ret;
+  }
+
+  return bi_list(bs, marker, max, entries, is_truncated);
+}
+
 int RGWRados::gc_operate(string& oid, librados::ObjectWriteOperation *op)
 {
   return gc_pool_ctx.operate(oid, op);
@@ -12333,6 +12398,44 @@ void RGWRados::get_bucket_instance_ids(RGWBucketInfo& bucket_info, int shard_id,
   }
 }
 
+int RGWRados::get_target_shard_id(const RGWBucketInfo& bucket_info, const string& obj_key,
+                                  int *shard_id)
+{
+  int r = 0;
+  switch (bucket_info.bucket_index_shard_hash_type) {
+    case RGWBucketInfo::MOD:
+      if (!bucket_info.num_shards) {
+        if (shard_id) {
+          *shard_id = -1;
+        }
+      } else {
+        uint32_t sid = ceph_str_hash_linux(obj_key.c_str(), obj_key.size());
+        uint32_t sid2 = sid ^ ((sid & 0xFF) << 24);
+        sid = sid2 % MAX_BUCKET_INDEX_SHARDS_PRIME % bucket_info.num_shards;
+        if (shard_id) {
+          *shard_id = (int)sid;
+        }
+      }
+      break;
+    default:
+      r = -ENOTSUP;
+  }
+  return r;
+}
+
+void RGWRados::get_bucket_index_object(const string& bucket_oid_base, uint32_t num_shards,
+                                      int shard_id, string *bucket_obj)
+{
+  if (!num_shards) {
+    // By default with no sharding, we use the bucket oid as itself
+    (*bucket_obj) = bucket_oid_base;
+  } else {
+    char buf[bucket_oid_base.size() + 32];
+    snprintf(buf, sizeof(buf), "%s.%d", bucket_oid_base.c_str(), shard_id);
+    (*bucket_obj) = buf;
+  }
+}
+
 int RGWRados::get_bucket_index_object(const string& bucket_oid_base, const string& obj_key,
     uint32_t num_shards, RGWBucketInfo::BIShardsHashType hash_type, string *bucket_obj, int *shard_id)
 {

diff --git a/src/rgw/rgw_rados.h b/src/rgw/rgw_rados.h
index b2092a160ac09f2da1997f5ed4175ed597092b29..d228b370c928c6766850043d6f91665ead5beff1 100644 (file)
--- a/src/rgw/rgw_rados.h
+++ b/src/rgw/rgw_rados.h
@@ -1819,6 +1819,8 @@ class RGWRados
       string& bucket_oid_base);
   int open_bucket_index_shard(rgw_bucket& bucket, librados::IoCtx& index_ctx,
       const string& obj_key, string *bucket_obj, int *shard_id);
+  int open_bucket_index_shard(rgw_bucket& bucket, librados::IoCtx& index_ctx,
+                              int shard_id, string *bucket_obj);
   int open_bucket_index(rgw_bucket& bucket, librados::IoCtx& index_ctx,
       map<int, string>& bucket_objs, int shard_id = -1, map<int, string> *bucket_instance_ids = NULL);
   template<typename T>
@@ -2160,6 +2162,7 @@ public:
                                  RGWZonePlacementInfo *rule_info);
   int set_bucket_location_by_rule(const string& location_rule, const string& tenant_name, const string& bucket_name, rgw_bucket& bucket,
                                   RGWZonePlacementInfo *rule_info);
+  void create_bucket_id(string *bucket_id);
   virtual int create_bucket(RGWUserInfo& owner, rgw_bucket& bucket,
                             const string& zonegroup_id,
                             const string& placement_rule,
@@ -2242,6 +2245,7 @@ public:
 
     explicit BucketShard(RGWRados *_store) : store(_store), shard_id(-1) {}
     int init(rgw_bucket& _bucket, rgw_obj& obj);
+    int init(rgw_bucket& _bucket, int sid);
   };
 
   class Object {
@@ -2923,7 +2927,11 @@ public:
 
   int bi_get_instance(rgw_obj& obj, rgw_bucket_dir_entry *dirent);
   int bi_get(rgw_bucket& bucket, rgw_obj& obj, BIIndexType index_type, rgw_cls_bi_entry *entry);
+  void bi_put(librados::ObjectWriteOperation& op, BucketShard& bs, rgw_cls_bi_entry& entry);
+  int bi_put(BucketShard& bs, rgw_cls_bi_entry& entry);
   int bi_put(rgw_bucket& bucket, rgw_obj& obj, rgw_cls_bi_entry& entry);
+  int bi_list(rgw_bucket& bucket, int shard_id, const string& marker, uint32_t max, list<rgw_cls_bi_entry> *entries, bool *is_truncated);
+  int bi_list(BucketShard& bs, const string& marker, uint32_t max, list<rgw_cls_bi_entry> *entries, bool *is_truncated);
   int bi_list(rgw_bucket& bucket, const string& obj_name, const string& marker, uint32_t max,
               list<rgw_cls_bi_entry> *entries, bool *is_truncated);
 
@@ -2936,6 +2944,7 @@ public:
   void shard_name(const string& prefix, unsigned max_shards, const string& key, string& name, int *shard_id);
   void shard_name(const string& prefix, unsigned max_shards, const string& section, const string& key, string& name);
   void shard_name(const string& prefix, unsigned shard_id, string& name);
+  int get_target_shard_id(const RGWBucketInfo& bucket_info, const string& obj_key, int *shard_id);
   void time_log_prepare_entry(cls_log_entry& entry, const ceph::real_time& ut, const string& section, const string& key, bufferlist& bl);
   int time_log_add_init(librados::IoCtx& io_ctx);
   int time_log_add(const string& oid, list<cls_log_entry>& entries,
@@ -3108,6 +3117,9 @@ public:
   int get_bucket_index_object(const string& bucket_oid_base, const string& obj_key,
       uint32_t num_shards, RGWBucketInfo::BIShardsHashType hash_type, string *bucket_obj, int *shard);
 
+  void get_bucket_index_object(const string& bucket_oid_base, uint32_t num_shards,
+                               int shard_id, string *bucket_obj);
+
   /**
    * Check the actual on-disk state of the object specified
    * by list_state, and fill in the time and size of object.