osd: add tier_evict for manifest tier

Truncate the object size to 0 if the mainfest object is composed of chunks.

Signed-off-by: Myoungwon Oh <myoungwon.oh@samsung.com>

src/include/rados.h

@@ -327,6 +327,7 @@ extern const char *ceph_osd_state_name(int s);
 	f(TIER_PROMOTE,	__CEPH_OSD_OP(WR, DATA, 41),	"tier-promote")	    \
 	f(UNSET_MANIFEST, __CEPH_OSD_OP(WR, DATA, 42),	"unset-manifest")   \
 	f(TIER_FLUSH, __CEPH_OSD_OP(CACHE, DATA, 43),	"tier-flush")	    \
+	f(TIER_EVICT, __CEPH_OSD_OP(CACHE, DATA, 44),	"tier-evict")	    \
 									    \
 	/** attrs **/							    \
 	/* read */							    \

src/include/rados/librados.hpp

@@ -753,6 +753,11 @@ inline namespace v14_2_0 {
      * updates.
      */
     void tier_flush();
+    /**
+     * evict a manifest tier object to backing tier; will block racing
+     * updates.
+     */
+    void tier_evict();
   };
 
   /* IoCtx : This is a context in which we can perform I/O.

src/librados/librados_cxx.cc

@@ -638,6 +638,13 @@ void librados::ObjectReadOperation::tier_flush()
   o->tier_flush();
 }
 
+void librados::ObjectReadOperation::tier_evict()
+{
+  ceph_assert(impl);
+  ::ObjectOperation *o = &impl->o;
+  o->tier_evict();
+}
+
 void librados::ObjectWriteOperation::set_redirect(const std::string& tgt_obj, 
 						  const IoCtx& tgt_ioctx,
 						  uint64_t tgt_version,

src/osd/PrimaryLogPG.cc

@@ -2375,7 +2375,8 @@ PrimaryLogPG::cache_result_t PrimaryLogPG::maybe_handle_manifest_detail(
 	op.op == CEPH_OSD_OP_SET_CHUNK || 
 	op.op == CEPH_OSD_OP_UNSET_MANIFEST ||
 	op.op == CEPH_OSD_OP_TIER_PROMOTE ||
-	op.op == CEPH_OSD_OP_TIER_FLUSH) {
+	op.op == CEPH_OSD_OP_TIER_FLUSH ||
+	op.op == CEPH_OSD_OP_TIER_EVICT) {
       return cache_result_t::NOOP;
     } 
   }
@@ -5655,6 +5656,7 @@ int PrimaryLogPG::do_osd_ops(OpContext *ctx, vector<OSDOp>& ops)
     case CEPH_OSD_OP_SET_CHUNK:
     case CEPH_OSD_OP_TIER_PROMOTE:
     case CEPH_OSD_OP_TIER_FLUSH:
+    case CEPH_OSD_OP_TIER_EVICT:
       break;
     default:
       if (op.op & CEPH_OSD_OP_MODE_WR)
@@ -7038,6 +7040,48 @@ int PrimaryLogPG::do_osd_ops(OpContext *ctx, vector<OSDOp>& ops)
 
       break;
 
+    case CEPH_OSD_OP_TIER_EVICT:
+      ++ctx->num_write;
+      result = 0;
+      {
+	if (pool.info.is_tier()) {
+	  result = -EINVAL;
+	  break;
+	}
+	if (!obs.exists) {
+	  result = -ENOENT;
+	  break;
+	}
+	if (get_osdmap()->require_osd_release < ceph_release_t::octopus) {
+	  result = -EOPNOTSUPP;
+	  break;
+	}
+	if (!obs.oi.has_manifest()) {
+	  result = -EINVAL;
+	  break;
+	}
+	
+	// The chunks already has a reference, so it is just enough to invoke truncate if necessary
+	uint64_t chunk_length = 0;
+	for (auto p : obs.oi.manifest.chunk_map) {
+	  chunk_length += p.second.length;
+	}
+	if (chunk_length == obs.oi.size) {
+	  // truncate
+	  for (auto p : obs.oi.manifest.chunk_map) {
+	    p.second.set_flag(chunk_info_t::FLAG_MISSING);
+	  }
+	  t->truncate(obs.oi.soid, 0);
+	  ctx->delta_stats.num_bytes -= obs.oi.size;
+	  ctx->delta_stats.num_wr++;
+	  oi.size = 0;
+	  ctx->cache_operation = true;
+	}
+	osd->logger->inc(l_osd_tier_evict);
+      }
+
+      break;
+
     case CEPH_OSD_OP_UNSET_MANIFEST:
       ++ctx->num_write;
       result = 0;

src/osdc/Objecter.h

@@ -1557,6 +1557,10 @@ struct ObjectOperation {
     add_op(CEPH_OSD_OP_TIER_FLUSH);
   }
 
+  void tier_evict() {
+    add_op(CEPH_OSD_OP_TIER_EVICT);
+  }
+
   void set_alloc_hint(uint64_t expected_object_size,
                       uint64_t expected_write_size,
 		      uint32_t flags) {

src/tools/rados/rados.cc

@@ -132,6 +132,7 @@ void usage(ostream& out)
 "   tier-promote <obj-name>	     promote the object to the base tier\n"
 "   unset-manifest <obj-name>	     unset redirect or chunked object\n"
 "   tier-flush <obj-name>	     flush the chunked object\n"
+"   tier-evict <obj-name>	     evict the chunked object\n"
 "\n"
 "IMPORT AND EXPORT\n"
 "   export [filename]\n"
@@ -3820,6 +3821,29 @@ static int rados_tool_common(const std::map < std::string, std::string > &opts,
 	   << cpp_strerror(ret) << std::endl;
       return 1;
     }
+  } else if (strcmp(nargs[0], "tier-evict") == 0) {
+    if (!pool_name || nargs.size() < 2) {
+      usage(cerr);
+      return 1;
+    }
+    string oid(nargs[1]);
+
+    ObjectReadOperation op;
+    op.tier_evict();
+    librados::AioCompletion *completion =
+      librados::Rados::aio_create_completion();
+    io_ctx.aio_operate(oid.c_str(), completion, &op,
+		       librados::OPERATION_IGNORE_CACHE |
+		       librados::OPERATION_IGNORE_OVERLAY,
+		       NULL);
+    completion->wait_for_complete();
+    ret = completion->get_return_value();
+    completion->release();
+    if (ret < 0) {
+      cerr << "error tier-evict " << pool_name << "/" << oid << " : " 
+	   << cpp_strerror(ret) << std::endl;
+      return 1;
+    }
   } else if (strcmp(nargs[0], "export") == 0) {
     // export [filename]
     if (!pool_name || nargs.size() > 2) {