Async for reader side of BR.

Signed-off-by: Xiaoxi Chen <[email protected]>

Author: xiaoxichen

src/lib/homestore_backend/hs_homeobject.hpp

@@ -59,7 +59,8 @@ class HSHomeObject : public HomeObjectImpl {
                                                trace_id_t tid) const override;
     BlobManager::NullAsyncResult _del_blob(ShardInfo const&, blob_id_t, trace_id_t tid) override;
 
-    PGManager::NullAsyncResult _create_pg(PGInfo&& pg_info, std::set< peer_id_t > const& peers, trace_id_t tid) override;
+    PGManager::NullAsyncResult _create_pg(PGInfo&& pg_info, std::set< peer_id_t > const& peers,
+                                          trace_id_t tid) override;
     PGManager::NullAsyncResult _replace_member(pg_id_t id, peer_id_t const& old_member, PGMember const& new_member,
                                                uint32_t commit_quorum, trace_id_t tid) override;
 
@@ -435,9 +436,11 @@ class HSHomeObject : public HomeObjectImpl {
 
     struct PGBlobIterator {
         struct blob_read_result {
+            blob_id_t blob_id_;
             sisl::io_blob_safe blob_;
             ResyncBlobState state_;
-	    blob_read_result(sisl::io_blob_safe&& blob, ResyncBlobState state): blob_(std::move(blob)), state_(state) {}
+            blob_read_result(blob_id_t blob_id, sisl::io_blob_safe&& blob, ResyncBlobState state) :
+                    blob_id_(blob_id), blob_(std::move(blob)), state_(state) {}
         };
         PGBlobIterator(HSHomeObject& home_obj, homestore::group_id_t group_id, uint64_t upto_lsn = 0);
         PG* get_pg_metadata();
@@ -639,7 +642,7 @@ class HSHomeObject : public HomeObjectImpl {
     static ShardInfo deserialize_shard_info(const char* shard_info_str, size_t size);
     static std::string serialize_shard_info(const ShardInfo& info);
     void local_create_shard(ShardInfo shard_info, homestore::chunk_num_t v_chunk_id, homestore::chunk_num_t p_chunk_id,
-                            homestore::blk_count_t blk_count, trace_id_t tid=0);
+                            homestore::blk_count_t blk_count, trace_id_t tid = 0);
     void add_new_shard_to_map(ShardPtr&& shard);
     void update_shard_in_map(const ShardInfo& shard_info);
 
@@ -812,7 +815,7 @@ class HSHomeObject : public HomeObjectImpl {
                             const homestore::MultiBlkId& pbas, cintrusive< homestore::repl_req_ctx >& hs_ctx);
     void on_blob_del_commit(int64_t lsn, sisl::blob const& header, sisl::blob const& key,
                             cintrusive< homestore::repl_req_ctx >& hs_ctx);
-    bool local_add_blob_info(pg_id_t pg_id, BlobInfo const& blob_info, trace_id_t tid=0);
+    bool local_add_blob_info(pg_id_t pg_id, BlobInfo const& blob_info, trace_id_t tid = 0);
     homestore::ReplResult< homestore::blk_alloc_hints >
     blob_put_get_blk_alloc_hints(sisl::blob const& header, cintrusive< homestore::repl_req_ctx >& ctx);
     void compute_blob_payload_hash(BlobHeader::HashAlgorithm algorithm, const uint8_t* blob_bytes, size_t blob_size,
@@ -902,13 +905,13 @@ class HSHomeObject : public HomeObjectImpl {
 
     // only leader will call incr and decr pending request num
     void incr_pending_request_num() const {
-        uint64_t now=pending_request_num.fetch_add(1);
+        uint64_t now = pending_request_num.fetch_add(1);
         LOGT("inc pending req, was {}", now);
     }
-        void decr_pending_request_num() const {
+    void decr_pending_request_num() const {
         uint64_t now = pending_request_num.fetch_sub(1);
         LOGT("desc pending req, was {}", now);
-        DEBUG_ASSERT(now>0, "pending == 0 ");
+        DEBUG_ASSERT(now > 0, "pending == 0 ");
     }
 };
 

src/lib/homestore_backend/pg_blob_iterator.cpp

@@ -187,8 +187,8 @@ bool HSHomeObject::PGBlobIterator::create_shard_snapshot_data(sisl::io_blob_safe
     return true;
 }
 
-BlobManager::AsyncResult< HSHomeObject::PGBlobIterator::blob_read_result >
-HSHomeObject::PGBlobIterator::load_blob_data(const BlobInfo& blob_info) {
+typedef HSHomeObject::PGBlobIterator::blob_read_result blob_read_result;
+BlobManager::AsyncResult< blob_read_result > HSHomeObject::PGBlobIterator::load_blob_data(const BlobInfo& blob_info) {
     auto shard_id = blob_info.shard_id;
     auto blob_id = blob_info.blob_id;
     auto blkid = blob_info.pbas;
@@ -202,8 +202,8 @@ HSHomeObject::PGBlobIterator::load_blob_data(const BlobInfo& blob_info) {
     LOGD("Blob get request: shardID=0x{:x}, pg={}, shard=0x{:x}, blob_id={}, blkid={}", shard_id,
          (shard_id >> homeobject::shard_width), (shard_id & homeobject::shard_mask), blob_id, blkid.to_string());
     return repl_dev_->async_read(blkid, sgs, total_size)
-        .thenValue([this, blob_id, shard_id, read_buf = std::move(read_buf)](
-                       auto&& result) mutable -> BlobManager::AsyncResult< HSHomeObject::PGBlobIterator::blob_read_result > {
+        .thenValue([this, blob_id, shard_id, read_buf = std::move(read_buf)](auto&& result) mutable
+                       -> BlobManager::AsyncResult< HSHomeObject::PGBlobIterator::blob_read_result > {
             if (result) {
                 LOGE("Failed to get blob, shardID=0x{:x}, pg={}, shard=0x{:x}, blob_id={}, err={}", shard_id,
                      (shard_id >> homeobject::shard_width), (shard_id & homeobject::shard_mask), blob_id,
@@ -217,15 +217,15 @@ HSHomeObject::PGBlobIterator::load_blob_data(const BlobInfo& blob_info) {
                 LOGE("Invalid header found, shardID=0x{:x}, pg={}, shard=0x{:x}, blob_id={}, [header={}]", shard_id,
                      (shard_id >> homeobject::shard_width), (shard_id & homeobject::shard_mask), blob_id,
                      header->to_string());
-                return HSHomeObject::PGBlobIterator::blob_read_result(std::move(read_buf), ResyncBlobState::CORRUPTED);
+                return blob_read_result(blob_id, std::move(read_buf), ResyncBlobState::CORRUPTED);
             }
 
             if (header->shard_id != shard_id) {
                 // The metrics for corrupted blob is handled on the follower side.
                 LOGE("Invalid shard_id in header, shardID=0x{:x}, pg={}, shard=0x{:x}, blob_id={}, [header={}]",
                      shard_id, (shard_id >> homeobject::shard_width), (shard_id & homeobject::shard_mask), blob_id,
                      header->to_string());
-                return HSHomeObject::PGBlobIterator::blob_read_result(std::move(read_buf), ResyncBlobState::CORRUPTED);
+                return blob_read_result(blob_id, std::move(read_buf), ResyncBlobState::CORRUPTED);
             }
 
             std::string user_key = header->user_key_size
@@ -242,12 +242,12 @@ HSHomeObject::PGBlobIterator::load_blob_data(const BlobInfo& blob_info) {
                      "[{}] [computed={:np}]",
                      shard_id, (shard_id >> homeobject::shard_width), (shard_id & homeobject::shard_mask), blob_id,
                      header->to_string(), spdlog::to_hex(computed_hash, computed_hash + BlobHeader::blob_max_hash_len));
-                return HSHomeObject::PGBlobIterator::blob_read_result(std::move(read_buf), ResyncBlobState::CORRUPTED);
+                return blob_read_result(blob_id, std::move(read_buf), ResyncBlobState::CORRUPTED);
             }
 
             LOGD("Blob get success: shardID=0x{:x}, pg={}, shard=0x{:x}, blob_id={}", shard_id,
                  (shard_id >> homeobject::shard_width), (shard_id & homeobject::shard_mask), blob_id);
-            return HSHomeObject::PGBlobIterator::blob_read_result(std::move(read_buf), ResyncBlobState::NORMAL);
+            return blob_read_result(blob_id, std::move(read_buf), ResyncBlobState::NORMAL);
         });
 }
 
@@ -259,9 +259,10 @@ bool HSHomeObject::PGBlobIterator::create_blobs_snapshot_data(sisl::io_blob_safe
     uint64_t total_bytes = 0;
     auto idx = cur_start_blob_idx_;
 
+    std::vector< BlobManager::AsyncResult< blob_read_result > > futs;
+    folly::InlineExecutor executor;
     while (total_bytes < max_batch_size_ && idx < cur_blob_list_.size()) {
         auto info = cur_blob_list_[idx++];
-        ResyncBlobState state = ResyncBlobState::NORMAL;
         // handle deleted object
         if (info.pbas == tombstone_pbas) {
             LOGT("Blob is deleted: shardID=0x{:x}, pg={}, shard=0x{:x}, blob_id={}, blkid={}", info.shard_id,
@@ -271,11 +272,10 @@ bool HSHomeObject::PGBlobIterator::create_blobs_snapshot_data(sisl::io_blob_safe
             continue;
         }
 
-        auto blob_start = Clock::now();
-
 #ifdef _PRERELEASE
         if (iomgr_flip::instance()->test_flip("pg_blob_iterator_load_blob_data_error")) {
             LOGW("Simulating loading blob data error");
+            futs.emplace_back(folly::makeUnexpected(BlobError(BlobErrorCode::READ_FAILED)));
             return false;
         }
         auto delay = iomgr_flip::instance()->get_test_flip< long >("simulate_read_snapshot_load_blob_delay",
@@ -286,39 +286,36 @@ bool HSHomeObject::PGBlobIterator::create_blobs_snapshot_data(sisl::io_blob_safe
             std::this_thread::sleep_for(std::chrono::milliseconds(delay.get()));
         }
 #endif
-        sisl::io_blob_safe blob;
-
-        struct blob_read_result {
-            sisl::io_blob_safe blob;
-            int state;
-        };
-
-        uint8_t retries = HS_BACKEND_DYNAMIC_CONFIG(snapshot_blob_load_retry);
-        for (int i = 0; i < retries; i++) {
-            auto result = load_blob_data(info).get();
-            if (result.hasError() && result.error().code == BlobErrorCode::READ_FAILED) {
-                LOGW("Failed to retrieve blob for shardID=0x{:x}, pg={}, shard=0x{:x} blob={} pbas={}, err={}, "
-                     "attempt={}",
-                     info.shard_id, (info.shard_id >> homeobject::shard_width),
-                     (info.shard_id & homeobject::shard_mask), info.blob_id, info.pbas.to_string(), result.error(), i);
-            } else {
-                blob = std::move(result.value().blob_);
-                state = result.value().state_;
-                break;
-            }
-        }
-        if (blob.size() == 0) {
-            LOGE("Failed to retrieve blob for shardID=0x{:x}, pg={}, shard=0x{:x} blob={} pbas={}", info.shard_id,
-                 (info.shard_id >> homeobject::shard_width), (info.shard_id & homeobject::shard_mask), info.blob_id,
-                 info.pbas.to_string());
-            COUNTER_INCREMENT(*metrics_, snp_dnr_error_count, 1);
-            return false;
-        }
-
-        HISTOGRAM_OBSERVE(*metrics_, snp_dnr_blob_process_latency, get_elapsed_time_us(blob_start));
-        std::vector< uint8_t > data(blob.cbytes(), blob.cbytes() + blob.size());
-        blob_entries.push_back(CreateResyncBlobDataDirect(builder_, info.blob_id, (uint8_t)state, &data));
-        total_bytes += blob.size();
+        auto blob_start = Clock::now();
+        auto executorKeepAlive = folly::getKeepAliveToken(executor);
+
+        // Fixme: Re-enable retries uint8_t retries = HS_BACKEND_DYNAMIC_CONFIG(snapshot_blob_load_retry);
+        futs.emplace_back(
+            std::move(load_blob_data(info))
+                .via(std::move(executorKeepAlive))
+                .thenValue(
+                    [this, info, blob_start](auto&& result) mutable -> BlobManager::AsyncResult< blob_read_result > {
+                        if (result.hasError() && result.error().code == BlobErrorCode::READ_FAILED) {
+                            LOGE("Failed to retrieve blob for shardID=0x{:x}, pg={}, shard=0x{:x} blob={} pbas={}",
+                                 info.shard_id, (info.shard_id >> homeobject::shard_width),
+                                 (info.shard_id & homeobject::shard_mask), info.blob_id, info.pbas.to_string());
+                            COUNTER_INCREMENT(*metrics_, snp_dnr_error_count, 1);
+                        } else {
+                            HISTOGRAM_OBSERVE(*metrics_, snp_dnr_blob_process_latency, get_elapsed_time_us(blob_start));
+                        }
+                        return result;
+                    }));
+
+        auto const expect_blob_size = info.pbas.blk_count() * repl_dev_->get_blk_size();
+        total_bytes += expect_blob_size;
+    }
+    // collect futs and add to flatbuffer.
+    for (auto it = futs.begin(); it != futs.end(); it++) {
+        auto res = std::move(*it).get();
+        // fail whole batch.
+        if (res.hasError()) { return false; }
+        std::vector< uint8_t > data(res->blob_.cbytes(), res->blob_.cbytes() + res->blob_.size());
+        blob_entries.push_back(CreateResyncBlobDataDirect(builder_, res->blob_id_, (uint8_t)res->state_, &data));
     }
     // should include the deleted blobs
     cur_batch_blob_count_ = idx - cur_start_blob_idx_;