分享一個hhvm使用http server方式來處理請求的問題及對應的patch。hhvm3+版本支持fastcgi模式,而以前的版本都只能用http serve模式來響應請求,因爲hhvm的http server支持的功能比較弱,如今大部分使用場景都選擇fastcgi了,而此次要說的問題僅存在於使用了http server模式的全部版本hhvm。html
咱們發現線上的hhvm會有丟失http header的極小機率問題,在單臺日請求量近千萬狀況下只有2個左右的發生機率,而且將出問題的url在線下復現,也沒有復現,瞬間沒有一點點線索了。沒有辦法,用笨辦法:看代碼+排除法,沒有辦法的辦法。但事實證實每每笨辦法也是最終的辦法。nginx
簡單的說,咱們的服務架構是nginx做爲反向代理服務器,來請求hhvm。Facebook的工程師基於libevent1.4.14b,增長了設置backlog功能和修復了一些內存泄漏問題,http的header的解析工做仍是在libevent裏獨立完成的,截取http.c部分代碼以下git
enum message_read_status evhttp_parse_headers(struct evhttp_request *req, struct evbuffer* buffer) { char *line; enum message_read_status status = MORE_DATA_EXPECTED; struct evkeyvalq* headers = req->input_headers; while ((line = evbuffer_readline(buffer)) != NULL) { char *skey, *svalue; if (*line == '\0') { /* Last header - Done */ status = ALL_DATA_READ; free(line); break; } /* Check if this is a continuation line */ if (*line == ' ' || *line == '\t') { if (evhttp_append_to_last_header(headers, line) == -1) goto error; free(line); continue; } /* Processing of header lines */ svalue = line; skey = strsep(&svalue, ":"); if (svalue == NULL) goto error; svalue += strspn(svalue, " "); if (evhttp_add_header(headers, skey, svalue) == -1) goto error; free(line); } return (status); error: free(line); return (DATA_CORRUPTED); }
能看出問題嗎,因爲定位問題過程比較曲折,這裏不羅嗦了。github
bug出如今buffer.c裏的readline函數,以下服務器
char * evbuffer_readline(struct evbuffer *buffer) { u_char *data = EVBUFFER_DATA(buffer); size_t len = EVBUFFER_LENGTH(buffer); char *line; unsigned int i; for (i = 0; i < len; i++) { if (data[i] == '\r' || data[i] == '\n') break; } if (i == len) return (NULL); if ((line = malloc(i + 1)) == NULL) { fprintf(stderr, "%s: out of memory\n", __func__); return (NULL); } memcpy(line, data, i); line[i] = '\0'; /* * Some protocols terminate a line with '\r\n', so check for * that, too. */ if ( i < len - 1 ) { char fch = data[i], sch = data[i+1]; /* Drain one more character if needed */ if ( (sch == '\r' || sch == '\n') && sch != fch ) i += 1; } evbuffer_drain(buffer, i + 1); return (line); }
此函數的功能就是截取一行出來,而後循環解析完全部header。rfc2616規定行分隔符是\r\n,而libevent認定的換行符(eof)倒是很是的寬鬆,好比\r就行。若是收包時剛好將\r和\n分離在兩個包裏,那麼後邊這個包解析時第一個字節就是\n,libevent解析時會認爲這是一行(其實仍是以前的行),而這行並無內容,即內容是\0,這會形成evhttp_parse_headers函數直接認爲header已經所有解析完畢了,也就是說丟失了header。架構
多說一句libevent的2x版本,相比以前的版本代碼變化很大,單就解析header來講,已經提供了三種級別的換行符認定標準,不過默認仍是最鬆散級別。app
因爲咱們這裏的環境都是可控的,不會存在那些亂七八糟換行符,因此就按照強約束\r\n(LRCF)來解析,針對libevent1.4.14b的patch以下socket
--- ./buffer.c 2010-06-20 21:06:04.000000000 +0800 +++ ./buffer.c 2014-04-21 14:22:25.783883798 +0800 @@ -211,45 +211,35 @@ char * evbuffer_readline(struct evbuffer *buffer) { - u_char *data = EVBUFFER_DATA(buffer); - size_t len = EVBUFFER_LENGTH(buffer); - char *line; - unsigned int i; - - for (i = 0; i < len; i++) { - if (data[i] == '\r' || data[i] == '\n') - break; - } - - if (i == len) - return (NULL); - - if ((line = malloc(i + 1)) == NULL) { - fprintf(stderr, "%s: out of memory\n", __func__); - return (NULL); - } + u_char *data = EVBUFFER_DATA(buffer); + size_t len = EVBUFFER_LENGTH(buffer); + char *line; + unsigned int i; + + for (i = 0; i < len; i++) { + if (data[i] == '\r') { + if (i + 1 < len && data[i+1] == '\n') { + break; + } + } + } + + if (i == len) + return (NULL); + + if ((line = malloc(i + 1)) == NULL) { + fprintf(stderr, "%s: out of memory\n", __func__); + return (NULL); + } - memcpy(line, data, i); - line[i] = '\0'; + memcpy(line, data, i); + line[i] = '\0'; - /* - * Some protocols terminate a line with '\r\n', so check for - * that, too. - */ - if ( i < len - 1 ) { - char fch = data[i], sch = data[i+1]; + evbuffer_drain(buffer, i + 2); - /* Drain one more character if needed */ - if ( (sch == '\r' || sch == '\n') && sch != fch ) - i += 1; - } - - evbuffer_drain(buffer, i + 1); - - return (line); + return (line); } - char * evbuffer_readln(struct evbuffer *buffer, size_t *n_read_out, enum evbuffer_eol_style eol_style)