示例位置: <hyperscan source>/examples/pcapscan.cc
參考:http://01org.github.io/hyperscan/dev-reference/api_files.htmlhtml
此示例實現一個簡單的數據包匹配性能測量程序。前端
pcapscan使用libpcap從pcap文件中讀取數據包,並根據一個規則文件中指定的多個正則表達式對報文進行匹配,並輸出匹配結果和一些統計信息。pcapscan使用並對比了兩種匹配模式:BLOCK和STREAM。BLOCK模式時它對單個數據包進行匹配;而STREAM模式下它經過五元組將數據包進行簡單分流,並對每條流中的數據進行匹配。STREAM模式能夠命中跨越數據包邊界的匹配數據(好比,要匹配abc,而a在前一個數據的末尾,而bc在後一個數據包的前端,這兩個數據包在一個流中,那麼STREAM模式匹配能夠命中它,而BLOCK模式不能)。c++
此示例演示瞭如下hyperscan概念:git
下面按照代碼執行的前後順序對pcapscan源碼進行簡單解讀。github
函數buildDatabase用來編譯規則文件中的多個正則表達式,參數mode指定了是BLOCK仍是STREAM模式。正則表達式
static hs_database_t *buildDatabase(const vector<const char *> &expressions, const vector<unsigned> flags, const vector<unsigned> ids, unsigned int mode) { hs_database_t *db; hs_compile_error_t *compileErr; hs_error_t err; Clock clock; clock.start(); err = hs_compile_multi(expressions.data(), flags.data(), ids.data(), expressions.size(), mode, nullptr, &db, &compileErr); clock.stop(); if (err != HS_SUCCESS) { if (compileErr->expression < 0) { // The error does not refer to a particular expression. cerr << "ERROR: " << compileErr->message << endl; } else { cerr << "ERROR: Pattern '" << expressions[compileErr->expression] << "' failed compilation with error: " << compileErr->message << endl; } // As the compileErr pointer points to dynamically allocated memory, if // we get an error, we must be sure to release it. This is not // necessary when no error is detected. hs_free_compile_error(compileErr); exit(-1); }
//...
}
其中的核心代碼是hs_compile_multi的調用,此函數用來編譯多個正則表達式,從代碼可見除了mode參數,BLOCK和STREAM模式都使用這一API。它的原型是express
hs_error_t hs_compile_multi(const char *const * expressions, const unsigned int * flags, const unsigned int * ids, unsigned int elements, unsigned int mode, const hs_platform_info_t * platform, hs_database_t ** db, hs_compile_error_t ** error)
其中,expressions是多個正則表達式字符串,flags和ids分別是expressions對應的flag和id數組;elements是表達式字符串的個數;其他參數與上一個例子中提到的hs_compile的參數涵義相同。ubuntu
這裏要注意的一個事情是參數ids,它是正則表達式的ID數組。每一個表達式都有一個惟一ID,這樣命中的時候匹配回調函數能夠獲得此ID,告訴調用者哪一個表達式命中了。若是ids傳入NULL,則全部表達式的ID都爲0。api
Benchmark構造函數中,爲接下來的匹配分配足夠的臨時數據空間(scratch space)。這裏有一個技巧:1)BLOCK和STREAM模式的匹配只需共用一個scratch;2)這個scratch足夠大,方法是調用兩次,在第2次調用時hyperscan若是發現空間不夠會進行增長。數組
public: Benchmark(const hs_database_t *streaming, const hs_database_t *block) : db_streaming(streaming), db_block(block), scratch(nullptr), matchCount(0) { // Allocate enough scratch space to handle either streaming or block // mode, so we only need the one scratch region. hs_error_t err = hs_alloc_scratch(db_streaming, &scratch); if (err != HS_SUCCESS) { cerr << "ERROR: could not allocate scratch space. Exiting." << endl; exit(-1); } // This second call will increase the scratch size if more is required // for block mode. err = hs_alloc_scratch(db_block, &scratch); if (err != HS_SUCCESS) { cerr << "ERROR: could not allocate scratch space. Exiting." << endl; exit(-1); } }
在Benchmark::readStreams方法中,從pcap文件中讀取了全部數據包(其實封裝必須是ethernet-ipv4-tcp/udp),並根據五元組進行簡單分流。主要代碼以下
while ((pktData = pcap_next(pcapHandle, &pktHeader)) != nullptr) { unsigned int offset = 0, length = 0; if (!payloadOffset(pktData, &offset, &length)) { continue; } // Valid TCP or UDP packet const struct ip *iphdr = (const struct ip *)(pktData + sizeof(struct ether_header)); const char *payload = (const char *)pktData + offset; size_t id = stream_map.insert(std::make_pair(FiveTuple(iphdr), stream_map.size())).first->second; packets.push_back(string(payload, length)); stream_ids.push_back(id); }
注意,stream_ids這個vector存儲了每個數據包對應的stream id。
因爲須要用到STREAM模式,因此在匹配前要先將流打開,見Benchmark::openStreams
// Open a Hyperscan stream for each stream in stream_ids void openStreams() { streams.resize(stream_map.size()); for (auto &stream : streams) { hs_error_t err = hs_open_stream(db_streaming, 0, &stream); if (err != HS_SUCCESS) { cerr << "ERROR: Unable to open stream. Exiting." << endl; exit(-1); } } }
其中,streams的類型是vector<hs_stream_t *>。
2.5.1 STREAM模式
在Benchmark::scanStreams中
// Scan each packet (in the ordering given in the PCAP file) through // Hyperscan using the streaming interface. void scanStreams() { for (size_t i = 0; i != packets.size(); ++i) { const std::string &pkt = packets[i]; hs_error_t err = hs_scan_stream(streams[stream_ids[i]], pkt.c_str(), pkt.length(), 0, scratch, onMatch, &matchCount); if (err != HS_SUCCESS) { cerr << "ERROR: Unable to scan packet. Exiting." << endl; exit(-1); } } }
hs_scan_stream的原型:
hs_error_t hs_scan_stream(hs_stream_t * id, const char * data, unsigned int length, unsigned int flags, hs_scratch_t * scratch, match_event_handler onEvent, void * ctxt)
其中,id是數據所屬的stream對應hs_stream_t指針,這裏叫id其實我感受不太合適; 其他參數與hs_scan相同。
這裏調用的streams[stream_ids[i]]已經在上一步打開流中初始化。
2.5.2 BLOCK模式
BLOCK模式比STREAM簡單許多,在Benchmark::scanBlock中
// Scan each packet (in the ordering given in the PCAP file) through // Hyperscan using the block-mode interface. void scanBlock() { for (size_t i = 0; i != packets.size(); ++i) { const std::string &pkt = packets[i]; hs_error_t err = hs_scan(db_block, pkt.c_str(), pkt.length(), 0, scratch, onMatch, &matchCount); if (err != HS_SUCCESS) { cerr << "ERROR: Unable to scan packet. Exiting." << endl; exit(-1); } } }
hs_scan在解讀simple中已經說過了,再也不贅述。
包括關閉流(hs_close_stream)、釋放database等。這裏要注意hs_close_stream時仍會進行匹配。
STREAM模式的用法比BLOCK模式要複雜一些,這裏簡單用僞代碼總結一下
// N是流的規格,事先已肯定好 hs_database_t* db; hs_stream_t* steams[N]; hs_scratch_t* tmp; uint8_t* pkt; // 1) 編譯多個正則表達式 hs_compile_multi(&db, HS_MODE_STREAM); // 2) 準備scratch hs_alloc_scratch(db, &tmp); // 3) 打開流 for(i=0; i<N; i++) hs_open_stream(db, &streams[i]); // 4) 收到數據包,並將其分到指定流 stream_id = classify(pkt); // 5) 流匹配 hs_scan_stream(streams[stream_id], pkt, &tmp, callBack); // 6) 清理資源, 注意hs_close_stream仍可能有匹配 for(i=0; i<N; i++) hs_close_stream(db, streams[i], &tmp, callBack);
hs_free_scrach(tmp); hs_free_database(db);
能夠經過hs_database_size()和hs_stream_size()分別得到database和每條流的stream state的大小。正則表達式的數目和複雜度會影響stream state的大小,隨着數目和複雜度的增長,可能會愈來愈大。在支持上百萬條流和複雜規則文件的系統上,stream state的內存耗費可能很大。
運行示例前要準備一個pcap文件和一個規則文件,規則文件的格式如
123:/weibo/ 456:/[f|F]ile/
每行一個正則表達式,冒號前面是表達式的ID,後面是pcre正則表達式。
如下是編譯和運行截圖,我用了一個微博流量的pcap,並匹配其中的weibo關鍵字:
zzq@ubuntu14:~/hs_demo$ g++ -o pcapscan pcapscan.cc -std=c++11 -lhs -lpcap zzq@ubuntu14:~/hs_demo$ ./pcapscan ptn weibo.pcap Pattern file: ptn Compiling Hyperscan databases with 1 patterns. Hyperscan streaming mode database compiled in 0.000236959 seconds. Hyperscan block mode database compiled in 4.8277e-05 seconds. PCAP input file: weibo.pcap 4 packets in 3 streams, totalling 3641 bytes. Average packet length: 910 bytes. Average stream length: 1213 bytes. Streaming mode Hyperscan database size : 1000 bytes. Block mode Hyperscan database size : 1000 bytes. Streaming mode Hyperscan stream state size: 25 bytes (per stream). Streaming mode: Total matches: 9 Match rate: 2.5312 matches/kilobyte Throughput (with stream overhead): 2576.33 megabits/sec Throughput (no stream overhead): 5444.49 megabits/sec Block mode: Total matches: 9 Match rate: 2.5312 matches/kilobyte Throughput: 16227.30 megabits/sec WARNING: Input PCAP file is less than 2MB in size. This test may have been too short to calculate accurate results.