[Filestore] Implement zero-copy read on VFS Fuse level

antonmyagkov · antonmyagkov · commit 019ee219625f · 2025-12-05T13:31:09.000+01:00
diff --git a/cloud/filestore/config/filesystem.proto b/cloud/filestore/config/filesystem.proto
@@ -97,4 +97,7 @@ message TFileSystemConfig
 
     // Enable FUSE_HANDLE_KILLPRIV_V2 on guest (FUSE client)
     optional bool GuestHandleKillPrivV2Enabled = 29;
+
+    // Enable zero-copy for read operations in the VFS FUSE layer
+    optional bool ZeroCopyReadEnabled = 30;
 }
diff --git a/cloud/filestore/config/server.proto b/cloud/filestore/config/server.proto
@@ -203,6 +203,9 @@ message TLocalServiceConfig
 
     // Enable FUSE_HANDLE_KILLPRIV_V2 on guest (FUSE client)
     optional bool GuestHandleKillPrivV2Enabled = 36;
+
+    // Enable zero-copy for read operations in the VFS FUSE layer
+    optional bool ZeroCopyReadEnabled = 37;
 }
 
 ////////////////////////////////////////////////////////////////////////////////
diff --git a/cloud/filestore/config/storage.proto b/cloud/filestore/config/storage.proto
@@ -656,6 +656,9 @@ message TStorageConfig
     // prototxt.
     optional bool UseBinaryFormatForTabletBootInfoBackup = 442;
 
+    // Enable zero-copy for read operations in the VFS FUSE layer
+    optional bool ZeroCopyReadEnabled = 443;
+
     // Allows to set node's priority for system (tenant) tablets (hive,
     // schemeshard, etc.). Lower value means lower priority.
     // NOTE: can be negative.
diff --git a/cloud/filestore/libs/service_local/config.cpp b/cloud/filestore/libs/service_local/config.cpp
@@ -49,6 +49,7 @@ namespace {
     xxx(DirectoryHandlesStorageEnabled, bool,       false                     )\
     xxx(DirectoryHandlesTableSize,   ui64,          100'000                   )\
     xxx(GuestHandleKillPrivV2Enabled,   bool,       false                     )\
+    xxx(ZeroCopyReadEnabled,            bool,       false                     )\
 // FILESTORE_SERVICE_CONFIG
 
 #define FILESTORE_SERVICE_NULL_FILE_IO_CONFIG(xxx)                             \
diff --git a/cloud/filestore/libs/service_local/config.h b/cloud/filestore/libs/service_local/config.h
@@ -131,6 +131,8 @@ class TLocalFileStoreConfig
     ui64 GetDirectoryHandlesTableSize() const;
 
     bool GetGuestHandleKillPrivV2Enabled() const;
+
+    bool GetZeroCopyReadEnabled() const;
 };
 
 }   // namespace NCloud::NFileStore
diff --git a/cloud/filestore/libs/storage/core/config.cpp b/cloud/filestore/libs/storage/core/config.cpp
@@ -301,6 +301,8 @@ using TAliases = NProto::TStorageConfig::TFilestoreAliases;
     xxx(DirectoryHandlesTableSize,         ui64,      100'000                 )\
     xxx(GuestHandleKillPrivV2Enabled,      bool,      false                   )\
     xxx(AllowAdditionalSystemTablets,      bool,      false                   )\
+                                                                               \
+    xxx(ZeroCopyReadEnabled,               bool,      false                   )\
 // FILESTORE_STORAGE_CONFIG
 
 #define FILESTORE_STORAGE_CONFIG_REF(xxx)                                      \
diff --git a/cloud/filestore/libs/storage/core/config.h b/cloud/filestore/libs/storage/core/config.h
@@ -353,6 +353,8 @@ class TStorageConfig
     bool GetGuestHandleKillPrivV2Enabled() const;
 
     [[nodiscard]] bool GetAllowAdditionalSystemTablets() const;
+
+    bool GetZeroCopyReadEnabled() const;
 };
 
 }   // namespace NCloud::NFileStore::NStorage
diff --git a/cloud/filestore/libs/storage/tablet/tablet_actor_createsession.cpp b/cloud/filestore/libs/storage/tablet/tablet_actor_createsession.cpp
@@ -89,6 +89,8 @@ void FillFeatures(
 
     features->SetGuestHandleKillPrivV2Enabled(
         config.GetGuestHandleKillPrivV2Enabled());
+
+    features->SetZeroCopyReadEnabled(config.GetZeroCopyReadEnabled());
 }
 
 ////////////////////////////////////////////////////////////////////////////////
diff --git a/cloud/filestore/libs/storage/tablet/tablet_ut_sessions.cpp b/cloud/filestore/libs/storage/tablet/tablet_ut_sessions.cpp
@@ -850,6 +850,7 @@ Y_UNIT_TEST_SUITE(TIndexTabletTest_Sessions)
         config.SetAllowHandlelessIO(true);
         config.SetZeroCopyWriteEnabled(true);
         config.SetGuestHandleKillPrivV2Enabled(true);
+        config.SetZeroCopyReadEnabled(true);
 
         features.SetTwoStageReadEnabled(true);
         features.SetEntryTimeout(TDuration::Seconds(10).MilliSeconds());
@@ -868,6 +869,7 @@ Y_UNIT_TEST_SUITE(TIndexTabletTest_Sessions)
         features.SetAllowHandlelessIO(true);
         features.SetZeroCopyWriteEnabled(true);
         features.SetGuestHandleKillPrivV2Enabled(true);
+        features.SetZeroCopyReadEnabled(true);
 
         DoTestShouldReturnFeaturesInCreateSessionResponse(config, features);
     }
diff --git a/cloud/filestore/libs/vfs_fuse/config.cpp b/cloud/filestore/libs/vfs_fuse/config.cpp
@@ -51,6 +51,7 @@ namespace {
     xxx(ZeroCopyWriteEnabled,         bool,     false                         )\
     xxx(FSyncQueueDisabled,           bool,     false                         )\
     xxx(GuestHandleKillPrivV2Enabled, bool,     false                         )\
+    xxx(ZeroCopyReadEnabled,          bool,     false                         )\
 // FILESTORE_FUSE_CONFIG
 
 #define FILESTORE_FILESYSTEM_DECLARE_CONFIG(name, type, value)                 \
diff --git a/cloud/filestore/libs/vfs_fuse/config.h b/cloud/filestore/libs/vfs_fuse/config.h
@@ -62,6 +62,7 @@ struct TFileSystemConfig
     ui32 GetMaxFuseLoopThreads() const;
 
     bool GetZeroCopyWriteEnabled() const;
+    bool GetZeroCopyReadEnabled() const;
 
     bool GetFSyncQueueDisabled() const;
 
diff --git a/cloud/filestore/libs/vfs_fuse/fs_impl_data.cpp b/cloud/filestore/libs/vfs_fuse/fs_impl_data.cpp
@@ -353,6 +353,57 @@ void TFileSystem::Read(
     request->SetOffset(offset);
     request->SetLength(size);
 
+    const bool isZeroCopyRead =
+        Config->GetZeroCopyReadEnabled() && !WriteBackCache;
+    if (isZeroCopyRead) {
+        struct iovec* iov = NULL;
+        int count = 0;
+        int ret = fuse_out_buf(req, &iov, &count);
+        if (ret == -1 || count <= 1) {
+            STORAGE_ERROR(
+                "Invalid fuse out buffers, ret=%d, count=%d",
+                ret,
+                count);
+            ReplyError(
+                *callContext,
+                MakeError(E_FS_INVAL, "Invalid fuse out buffers"),
+                req,
+                EINVAL);
+            return;
+        }
+
+        auto* iovecs = request->MutableIovecs();
+        iovecs->Reserve(count);
+
+        size_t remainingSize = request->GetLength();
+        // skip first fuse out iovec where headers are kept rest of the iovecs
+        // contain pointers to data buffers
+        for (int index = 1; index < count; index++) {
+            if (remainingSize == 0) {
+                break;
+            }
+            auto dataSize = std::min(remainingSize, iov[index].iov_len);
+            auto* iovec = iovecs->Add();
+            iovec->SetBase(reinterpret_cast<ui64>(iov[index].iov_base));
+            iovec->SetLength(dataSize);
+            remainingSize -= dataSize;
+        }
+
+        if (remainingSize != 0) {
+            STORAGE_WARN(
+                "Read request length exceeds fuse buffer space, remainingSize="
+                << remainingSize);
+            ReplyError(
+                *callContext,
+                MakeError(
+                    E_FS_INVAL,
+                    "request length exceeds fuse buffer space"),
+                req,
+                EINVAL);
+            return;
+        }
+    }
+
     TFuture<NProto::TReadDataResponse> future;
     if (WriteBackCache) {
         future = WriteBackCache.ReadData(callContext, std::move(request));
@@ -369,13 +420,22 @@ void TFileSystem::Read(
         const auto& response = future.GetValue();
         if (CheckResponse(self, *callContext, req, response)) {
             const auto& buffer = response.GetBuffer();
-            ui32 bufferOffset = response.GetBufferOffset();
-            self->ReplyBuf(
-                *callContext,
-                response.GetError(),
-                req,
-                buffer.data() + bufferOffset,
-                buffer.size() - bufferOffset);
+            if (buffer.empty()) {
+                self->ReplyBuf(
+                    *callContext,
+                    response.GetError(),
+                    req,
+                    nullptr,
+                    response.GetLength());
+            } else {
+                ui32 bufferOffset = response.GetBufferOffset();
+                self->ReplyBuf(
+                    *callContext,
+                    response.GetError(),
+                    req,
+                    buffer.data() + bufferOffset,
+                    buffer.size() - bufferOffset);
+            }
         }
     });
 }
diff --git a/cloud/filestore/libs/vfs_fuse/loop.cpp b/cloud/filestore/libs/vfs_fuse/loop.cpp
@@ -1193,6 +1193,7 @@ class TFileSystemLoop final
         config.SetMaxFuseLoopThreads(features.GetMaxFuseLoopThreads());
 
         config.SetZeroCopyWriteEnabled(features.GetZeroCopyWriteEnabled());
+        config.SetZeroCopyReadEnabled(features.GetZeroCopyReadEnabled());
 
         config.SetFSyncQueueDisabled(features.GetFSyncQueueDisabled());
 
diff --git a/cloud/filestore/public/api/protos/fs.proto b/cloud/filestore/public/api/protos/fs.proto
@@ -44,6 +44,7 @@ message TFileStoreFeatures
     bool DirectoryHandlesStorageEnabled = 29;
     uint64 DirectoryHandlesTableSize = 30;
     bool GuestHandleKillPrivV2Enabled = 31;
+    bool ZeroCopyReadEnabled = 32;
 }
 
 message TFileStore
diff --git a/cloud/filestore/tests/fio/qemu-kikimr-zero-copy-write-fallback-test/nfs-patch.txt b/cloud/filestore/tests/fio/qemu-kikimr-zero-copy-write-fallback-test/nfs-patch.txt
@@ -1 +1,2 @@
 ZeroCopyWriteEnabled: true
+ZeroCopyReadEnabled: true
diff --git a/cloud/filestore/tests/fio/qemu-kikimr-zero-copy-write-test/nfs-patch.txt b/cloud/filestore/tests/fio/qemu-kikimr-zero-copy-write-test/nfs-patch.txt
@@ -2,3 +2,5 @@ ZeroCopyWriteEnabled: true
 ThreeStageWriteEnabled: true
 UnalignedThreeStageWriteEnabled: true
 MultiTabletForwardingEnabled: true
+ZeroCopyReadEnabled: true
+TwoStageReadEnabled: true

Original file line number	Diff line number	Diff line change
`@@ -97,4 +97,7 @@ message TFileSystemConfig`
`97`	`97`
`98`	`98`	`// Enable FUSE_HANDLE_KILLPRIV_V2 on guest (FUSE client)`
`99`	`99`	`optional bool GuestHandleKillPrivV2Enabled = 29;`
	`100`	`+`
	`101`	`+ // Enable zero-copy for read operations in the VFS FUSE layer`
	`102`	`+ optional bool ZeroCopyReadEnabled = 30;`
`100`	`103`	`}`
Original file line number	Diff line number	Diff line change
`@@ -203,6 +203,9 @@ message TLocalServiceConfig`
`203`	`203`
`204`	`204`	`// Enable FUSE_HANDLE_KILLPRIV_V2 on guest (FUSE client)`
`205`	`205`	`optional bool GuestHandleKillPrivV2Enabled = 36;`
	`206`	`+`
	`207`	`+ // Enable zero-copy for read operations in the VFS FUSE layer`
	`208`	`+ optional bool ZeroCopyReadEnabled = 37;`
`206`	`209`	`}`
`207`	`210`
`208`	`211`	`////////////////////////////////////////////////////////////////////////////////`
Original file line number	Diff line number	Diff line change
`@@ -89,6 +89,8 @@ void FillFeatures(`
`89`	`89`
`90`	`90`	`features->SetGuestHandleKillPrivV2Enabled(`
`91`	`91`	`config.GetGuestHandleKillPrivV2Enabled());`
	`92`	`+`
	`93`	`+ features->SetZeroCopyReadEnabled(config.GetZeroCopyReadEnabled());`
`92`	`94`	`}`
`93`	`95`
`94`	`96`	`////////////////////////////////////////////////////////////////////////////////`