randomkang commented on code in PR #3144:
URL: https://github.com/apache/brpc/pull/3144#discussion_r4513071958


##########
src/brpc/policy/baidu_rpc_protocol_gpu.cpp:
##########
@@ -0,0 +1,294 @@
+// Licensed to the Apache Software Foundation (ASF) under one
+// or more contributor license agreements.  See the NOTICE file
+// distributed with this work for additional information
+// regarding copyright ownership.  The ASF licenses this file
+// to you under the Apache License, Version 2.0 (the
+// "License"); you may not use this file except in compliance
+// with the License.  You may obtain a copy of the License at
+//
+//   http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing,
+// software distributed under the License is distributed on an
+// "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+// KIND, either express or implied.  See the License for the
+// specific language governing permissions and limitations
+// under the License.
+
+#if BRPC_WITH_GDR
+
+#include <google/protobuf/descriptor.h>         // MethodDescriptor
+#include <google/protobuf/message.h>            // Message
+#include <google/protobuf/io/zero_copy_stream_impl_lite.h>
+#include <google/protobuf/io/coded_stream.h>
+#include <google/protobuf/text_format.h>
+
+#include "butil/gpu/gpu_block_pool.h"
+#include "butil/iobuf.h"                         // butil::IOBuf
+#include "butil/logging.h"                       // LOG()
+#include "butil/memory/scope_guard.h"
+#include "butil/raw_pack.h"                      // RawPacker RawUnpacker
+#include "butil/strings/string_util.h"
+
+#include "json2pb/json_to_pb.h"
+#include "json2pb/pb_to_json.h"
+#include "brpc/controller.h"                    // Controller
+#include "brpc/socket.h"                        // Socket
+#include "brpc/server.h"                        // Server
+#include "brpc/span.h"
+#include "brpc/compress.h"                      // ParseFromCompressedData
+#include "brpc/checksum.h"
+#include "brpc/stream_impl.h"
+#include "brpc/rpc_dump.h"                      // SampledRequest
+#include "brpc/rpc_pb_message_factory.h"
+#include "brpc/policy/baidu_rpc_meta.pb.h"      // RpcRequestMeta
+#include "brpc/policy/baidu_rpc_protocol.h"
+#include "brpc/policy/most_common_message.h"
+#include "brpc/policy/streaming_rpc_protocol.h"
+#include "brpc/details/usercode_backup_pool.h"
+#include "brpc/details/controller_private_accessor.h"
+#include "brpc/details/server_private_accessor.h"
+
+namespace brpc {
+namespace policy {
+
+// Notes:
+// 1. 12-byte header [PRPC][body_size][meta_size]
+// 2. body_size and meta_size are in network byte order
+// 3. Use service->full_name() + method_name to specify the method to call
+// 4. `attachment_size' is set iff request/response has attachment
+// 5. Not supported: chunk_info
+
+// Pack header into `buf'
+
+const int header_size = 12;
+// if we recv data into gpu, the header/meta/body will be copied to cpu and 
processed.
+// in to to limit the count of d2h, we will prefetch some bytes from gpu to 
cpu.
+// if header_size + meta_size + body_size(without attachment) is less than the
+// prefetch size, then one d2h is enough for one rpc.
+// The prefetch size is derived from the CPU block pool's block size so that 
the
+// allocation never exceeds the allocator's single-block limit and stays 
effective
+// if the block size changes.
+
+static const int kDefaultPrefetchD2HSize = 512;
+
+static int GetPrefetchD2HSize(butil::gdr::BlockPoolAllocator* host_allocator) {
+    const int64_t block_size = host_allocator->get_block_size();
+    return block_size > 0 ? static_cast<int>(block_size)
+                          : kDefaultPrefetchD2HSize;
+}
+
+ParseResult ParseRpcMessageGpu(butil::IOBuf* source, Socket* socket,
+                            bool /*read_eof*/, const void*) {
+
+    char header_buf[12];
+    size_t n = 0;
+    uint32_t body_size;
+    uint32_t meta_size;
+    ParseError pe = PARSE_OK;
+
+    void* prefetch_d2h_data = NULL;
+    bool is_gpu_memory = source->is_gpu_memory();
+    if (!is_gpu_memory) {
+        // A non-GPU buffer here means the message arrived through a non-GDR
+        // path; reject it instead of aborting the whole process.
+        LOG(ERROR) << "RpcMessage is not in gpu!!!";
+        return MakeParseError(PARSE_ERROR_ABSOLUTELY_WRONG);
+    }
+    butil::gdr::BlockPoolAllocator* host_allocator = 
butil::gdr::BlockPoolAllocators::GetInstance()->get_cpu_allocator();
+    const int prefetch_d2h_size = GetPrefetchD2HSize(host_allocator);
+    prefetch_d2h_data = host_allocator->AllocateRaw(prefetch_d2h_size);
+    if (prefetch_d2h_data == NULL) {
+        // Pool exhausted or single-block limit hit: ask the caller to retry
+        // instead of crashing the whole server.
+        LOG(ERROR) << "alloc host data for prefetch d2h failed, size="
+                   << prefetch_d2h_size;
+        return MakeParseError(PARSE_ERROR_NOT_ENOUGH_DATA);
+    }
+
+    // n is the bytes we real frefetch, n maybe less than prefetch_d2h_size;
+    n = source->copy_from_gpu(prefetch_d2h_data, prefetch_d2h_size);
+    size_t copy_size = n > 12 ? 12 : n;
+    memcpy(header_buf, prefetch_d2h_data, copy_size);
+
+    do {
+        // NOTE: unlike the CPU ParseRpcMessage, malformed GPU frames must NOT
+        // return PARSE_ERROR_TRY_OTHERS: InputMessengerProcessor would then
+        // dispatch the same GPU-resident IOBuf to non-GDR parsers whose
+        // copy_to cannot dereference device memory and would crash. A GDR
+        // connection only speaks baidu_std, so treat this as terminal.
+        if (n >= 4) {
+            void* dummy = header_buf;
+            if (*(const uint32_t*)dummy != *(const uint32_t*)"PRPC") {
+                pe = PARSE_ERROR_ABSOLUTELY_WRONG;
+                break;
+            }
+        } else {
+            if (memcmp(header_buf, "PRPC", n) != 0) {
+                pe = PARSE_ERROR_ABSOLUTELY_WRONG;
+                break;
+            }
+        }
+        if (n < sizeof(header_buf)) {
+            pe = PARSE_ERROR_NOT_ENOUGH_DATA;
+            break;
+        }
+        butil::RawUnpacker(header_buf + 
4).unpack32(body_size).unpack32(meta_size);
+        if (body_size > FLAGS_max_body_size) {
+            // We need this log to report the body_size to give users some 
clues
+            // which is not printed in InputMessenger.
+            LOG(ERROR) << "body_size=" << body_size << " from "
+                       << socket->remote_side() << " is too large";
+            pe = PARSE_ERROR_TOO_BIG_DATA;
+            break;
+        } else if (source->length() < sizeof(header_buf) + body_size) {
+            pe = PARSE_ERROR_NOT_ENOUGH_DATA;
+            break;
+        }
+        if (meta_size > body_size) {
+            LOG(ERROR) << "meta_size=" << meta_size << " is bigger than 
body_size="
+                       << body_size;
+            // Pop the message
+            source->pop_front(sizeof(header_buf) + body_size);
+            pe = PARSE_ERROR_ABSOLUTELY_WRONG;
+            break;
+        }
+    } while (0);
+
+    if (pe != PARSE_OK) {
+        host_allocator->DeallocateRaw(prefetch_d2h_data);
+        return MakeParseError(pe);
+    }
+
+    source->pop_front(sizeof(header_buf));
+    MostCommonMessage* msg = MostCommonMessage::Get();
+
+    if (header_size + meta_size <= n) {
+        auto deleter = [host_allocator, prefetch_d2h_data](void* data) { 
host_allocator->DeallocateRaw(prefetch_d2h_data); };
+        // `n` is the number of bytes actually prefetched from GPU. It is 
stored
+        // in the meta's `data_meta' so that FillReqBufGpu/FillResBufGpu can
+        // decide whether the prefetched host buffer also covers the body (and
+        // thus avoid a second D2H copy). This block is host memory 
(prefetched),
+        // NOT GPU memory, so we deliberately use append_user_data_with_meta
+        // (not append_user_data_gpu) to keep is_gpu_memory() returning false.
+        if (msg->meta.append_user_data_with_meta(
+                (char*)prefetch_d2h_data + header_size, meta_size, deleter, n) 
!= 0) {
+            // append failed (internal malloc): the prefetch buffer is NOT
+            // owned by anyone; release it back to the pool instead of leaking.
+            LOG(ERROR) << "Fail to append prefetched meta to IOBuf";
+            host_allocator->DeallocateRaw(prefetch_d2h_data);
+            return MakeParseError(PARSE_ERROR_NOT_ENOUGH_DATA);

Review Comment:
   OK,the error has been changed from PARSE_ERROR_NOT_ENOUGH_DATA to 
PARSE_ERROR_ABSOLUTELY_WRONG.



-- 
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.

To unsubscribe, e-mail: [email protected]

For queries about this service, please contact Infrastructure at:
[email protected]


---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]

Reply via email to