ref: 137d309c0c5217c618f0c51e5cbaa71716dd9bcb
dir: /src/http.c/
/********************************************************************
* *
* THIS FILE IS PART OF THE libopusfile SOFTWARE CODEC SOURCE CODE. *
* USE, DISTRIBUTION AND REPRODUCTION OF THIS LIBRARY SOURCE IS *
* GOVERNED BY A BSD-STYLE SOURCE LICENSE INCLUDED WITH THIS SOURCE *
* IN 'COPYING'. PLEASE READ THESE TERMS BEFORE DISTRIBUTING. *
* *
* THE libopusfile SOURCE CODE IS (C) COPYRIGHT 2012 *
* by the Xiph.Org Foundation and contributors http://www.xiph.org/ *
* *
********************************************************************/
#include "internal.h"
#include <ctype.h>
#include <errno.h>
#include <limits.h>
#include <string.h>
/*RFCs referenced in this file:
RFC 761: DOD Standard Transmission Control Protocol
RFC 1535: A Security Problem and Proposed Correction With Widely Deployed DNS
Software
RFC 1738: Uniform Resource Locators (URL)
RFC 1945: Hypertext Transfer Protocol -- HTTP/1.0
RFC 2068: Hypertext Transfer Protocol -- HTTP/1.1
RFC 2145: Use and Interpretation of HTTP Version Numbers
RFC 2246: The TLS Protocol Version 1.0
RFC 2459: Internet X.509 Public Key Infrastructure Certificate and
Certificate Revocation List (CRL) Profile
RFC 2616: Hypertext Transfer Protocol -- HTTP/1.1
RFC 2617: HTTP Authentication: Basic and Digest Access Authentication
RFC 2817: Upgrading to TLS Within HTTP/1.1
RFC 2818: HTTP Over TLS
RFC 3492: Punycode: A Bootstring encoding of Unicode for Internationalized
Domain Names in Applications (IDNA)
RFC 3986: Uniform Resource Identifier (URI): Generic Syntax
RFC 3987: Internationalized Resource Identifiers (IRIs)
RFC 4343: Domain Name System (DNS) Case Insensitivity Clarification
RFC 5894: Internationalized Domain Names for Applications (IDNA):
Background, Explanation, and Rationale
RFC 6066: Transport Layer Security (TLS) Extensions: Extension Definitions
RFC 6125: Representation and Verification of Domain-Based Application Service
Identity within Internet Public Key Infrastructure Using X.509 (PKIX)
Certificates in the Context of Transport Layer Security (TLS)*/
typedef struct OpusParsedURL OpusParsedURL;
typedef struct OpusStringBuf OpusStringBuf;
typedef struct OpusHTTPConn OpusHTTPConn;
typedef struct OpusHTTPStream OpusHTTPStream;
static char *op_string_range_dup(const char *_start,const char *_end){
size_t len;
char *ret;
OP_ASSERT(_start<=_end);
len=_end-_start;
/*This is to help avoid overflow elsewhere, later.*/
if(OP_UNLIKELY(len>=INT_MAX))return NULL;
ret=(char *)_ogg_malloc(sizeof(*ret)*(len+1));
if(OP_LIKELY(ret!=NULL)){
memcpy(ret,_start,sizeof(*ret)*(len));
ret[len]='\0';
}
return ret;
}
static char *op_string_dup(const char *_s){
return op_string_range_dup(_s,_s+strlen(_s));
}
static char *op_string_tolower(char *_s){
int i;
for(i=0;_s[i]!='\0';i++){
int c;
c=_s[i];
if(c>='A'&&c<='Z')c+='a'-'A';
_s[i]=(char)c;
}
return _s;
}
/*URI character classes (from RFC 3986).*/
#define OP_URL_ALPHA \
"ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz"
#define OP_URL_DIGIT "0123456789"
#define OP_URL_HEXDIGIT "0123456789ABCDEFabcdef"
/*Not a character class, but the characters allowed in <scheme>.*/
#define OP_URL_SCHEME OP_URL_ALPHA OP_URL_DIGIT "+-."
#define OP_URL_GEN_DELIMS "#/:?@[]"
#define OP_URL_SUB_DELIMS "!$&'()*+,;="
#define OP_URL_RESERVED OP_URL_GEN_DELIMS OP_URL_SUB_DELIMS
#define OP_URL_UNRESERVED OP_URL_ALPHA OP_URL_DIGIT "-._~"
/*Not a character class, but the characters allowed in <pct-encoded>.*/
#define OP_URL_PCT_ENCODED "%"
/*Not a character class or production rule, but for convenience.*/
#define OP_URL_PCHAR_BASE \
OP_URL_UNRESERVED OP_URL_PCT_ENCODED OP_URL_SUB_DELIMS
#define OP_URL_PCHAR OP_URL_PCHAR_BASE ":@"
/*Not a character class, but the characters allowed in <userinfo> and
<IP-literal>.*/
#define OP_URL_PCHAR_NA OP_URL_PCHAR_BASE ":"
/*Not a character class, but the characters allowed in <segment-nz-nc>.*/
#define OP_URL_PCHAR_NC OP_URL_PCHAR_BASE "@"
/*Not a character clsss, but the characters allowed in <path>.*/
#define OP_URL_PATH OP_URL_PCHAR "/"
/*Not a character class, but the characters allowed in <query> / <fragment>.*/
#define OP_URL_QUERY_FRAG OP_URL_PCHAR "/?"
/*Check the <% HEXDIG HEXDIG> escapes of a URL for validity.
Return: 0 if valid, or a negative value on failure.*/
static int op_validate_url_escapes(const char *_s){
int i;
for(i=0;_s[i];i++){
if(_s[i]=='%'){
if(OP_UNLIKELY(!isxdigit(_s[i+1]))
||OP_UNLIKELY(!isxdigit(_s[i+2]))
/*RFC 3986 says %00 "should be rejected if the application is not
expecting to receive raw data within a component."*/
||OP_UNLIKELY(_s[i+1]=='0'&&_s[i+2]=='0')){
return OP_FALSE;
}
i+=2;
}
}
return 0;
}
/*Convert a hex digit to its actual value.
_c: The hex digit to convert.
Presumed to be valid ('0'...'9', 'A'...'F', or 'a'...'f').
Return: The value of the digit, in the range [0,15].*/
static int op_hex_value(int _c){
return _c>='a'?_c-'a'+10:_c>='A'?_c-'A'+10:_c-'0';
}
/*Unescape all the <% HEXDIG HEXDIG> sequences in a string in-place.
This does no validity checking.*/
static char *op_unescape_url_component(char *_s){
int i;
int j;
for(i=j=0;_s[i];i++,j++){
if(_s[i]=='%'){
_s[i]=(char)(op_hex_value(_s[i+1])<<4|op_hex_value(_s[i+2]));
i+=2;
}
}
return _s;
}
/*Parse a file: URL.
This code is not meant to be fast: strspn() with large sets is likely to be
slow, but it is very convenient.
It is meant to be RFC 1738-compliant (as updated by RFC 3986).*/
static const char *op_parse_file_url(const char *_src){
const char *scheme_end;
const char *path;
const char *path_end;
scheme_end=_src+strspn(_src,OP_URL_SCHEME);
if(OP_UNLIKELY(*scheme_end!=':')
||scheme_end-_src!=4||op_strncasecmp(_src,"file",4)!=0){
/*Unsupported protocol.*/
return NULL;
}
/*Make sure all escape sequences are valid to simplify unescaping later.*/
if(OP_UNLIKELY(op_validate_url_escapes(scheme_end+1)<0))return NULL;
if(scheme_end[1]=='/'&&scheme_end[2]=='/'){
const char *host;
/*file: URLs can have a host!
Yeah, I was surprised, too, but that's what RFC 1738 says.
It also says, "The file URL scheme is unusual in that it does not specify
an Internet protocol or access method for such files; as such, its
utility in network protocols between hosts is limited," which is a mild
understatement.*/
host=scheme_end+3;
/*The empty host is what we expect.*/
if(OP_LIKELY(*host=='/'))path=host;
else{
const char *host_end;
char host_buf[28];
/*RFC 1738 says localhost "is interpreted as `the machine from which the
URL is being interpreted,'" so let's check for it.*/
host_end=host+strspn(host,OP_URL_PCHAR_BASE);
/*No <port> allowed.
This also rejects IP-Literals.*/
if(*host_end!='/')return NULL;
/*An escaped "localhost" can take at most 27 characters.*/
if(OP_UNLIKELY(host_end-host>27))return NULL;
memcpy(host_buf,host,sizeof(*host_buf)*(host_end-host));
host_buf[host_end-host]='\0';
op_unescape_url_component(host_buf);
op_string_tolower(host_buf);
/*Some other host: give up.*/
if(OP_UNLIKELY(strcmp(host_buf,"localhost")!=0))return NULL;
path=host_end;
}
}
else path=scheme_end+1;
path_end=path+strspn(path,OP_URL_PATH);
/*This will reject a <query> or <fragment> component, too.
I don't know what to do with queries, but a temporal fragment would at
least make sense.
RFC 1738 pretty clearly defines a <searchpart> that's equivalent to the
RFC 3986 <query> component for other schemes, but not the file: scheme,
so I'm going to just reject it.*/
if(*path_end!='\0')return NULL;
return path;
}
#if defined(OP_ENABLE_HTTP)
# include <sys/ioctl.h>
# include <sys/types.h>
# include <sys/socket.h>
# include <sys/timeb.h>
# include <arpa/inet.h>
# include <netinet/in.h>
# include <netinet/tcp.h>
# include <fcntl.h>
# include <netdb.h>
# include <poll.h>
# include <unistd.h>
# include <openssl/ssl.h>
# include <openssl/x509v3.h>
/*The maximum number of simultaneous connections.
RFC 2616 says this SHOULD NOT be more than 2, but everyone on the modern web
ignores that (e.g., IE 8 bumped theirs up from 2 to 6, Firefox uses 15).
If it makes you feel better, we'll only ever actively read from one of these
at a time.
The others are kept around mainly to avoid slow-starting a new connection
when seeking, and time out rapidly.*/
# define OP_NCONNS_MAX (4)
/*The number of redirections at which we give up.
The value here is the current default in Firefox.
RFC 2068 mandated a maximum of 5, but RFC 2616 relaxed that to "a client
SHOULD detect infinite redirection loops."
Fortunately, 20 is less than infinity.*/
# define OP_REDIRECT_LIMIT (20)
/*The initial size of the buffer used to read a response message (before the
body).*/
# define OP_RESPONSE_SIZE_MIN (510)
/*The maximum size of a response message (before the body).
Responses larger than this will be discarded.
I've seen a real server return 20 kB of data for a 302 Found response.
Increasing this beyond 32kB will cause problems on platforms with a 16-bit
int.*/
# define OP_RESPONSE_SIZE_MAX (32766)
/*The number of milliseconds we will allow a connection to sit idle before we
refuse to resurrect it.
Apache as of 2.2 has reduced its default timeout to 5 seconds (from 15), so
that's what we'll use here.*/
# define OP_CONNECTION_IDLE_TIMEOUT_MS (5*1000)
/*The number of milliseconds we will wait to send or receive data before giving
up.*/
# define OP_POLL_TIMEOUT_MS (30*1000)
/*We will always attempt to read ahead at least this much in preference to
opening a new connection.*/
# define OP_READAHEAD_THRESH_MIN (32*(opus_int32)1024)
/*The amount of data to request after a seek.
This is a trade-off between read throughput after a seek vs. the the ability
to quickly perform another seek with the same connection.*/
# define OP_PIPELINE_CHUNK_SIZE (32*(opus_int32)1024)
/*Subsequent chunks are requested with larger and larger sizes until they pass
this threshold, after which we just ask for the rest of the resource.*/
# define OP_PIPELINE_CHUNK_SIZE_MAX (1024*(opus_int32)1024)
/*This is the maximum number of requests we'll make with a single connection.
Many servers will simply disconnect after we attempt some number of requests,
possibly without sending a Connection: close header, meaning we won't
discover it until we try to read beyond the end of the current chunk.
We can reconnect when that happens, but this is slow.
Instead, we impose a limit ourselves (set to the default for Apache
installations and thus likely the most common value in use).*/
# define OP_PIPELINE_MAX_REQUESTS (100)
/*This should be the number of requests, starting from a chunk size of
OP_PIPELINE_CHUNK_SIZE and doubling each time, until we exceed
OP_PIPELINE_CHUNK_SIZE_MAX and just request the rest of the file.
We won't reuse a connection when seeking unless it has at least this many
requests left, to reduce the chances we'll have to open a new connection
while reading forward afterwards.*/
# define OP_PIPELINE_MIN_REQUESTS (7)
/*Is this an https URL?
For now we can simply check the last letter of the scheme.*/
# define OP_URL_IS_SSL(_url) ((_url)->scheme[4]=='s')
/*Does this URL use the default port for its scheme?*/
# define OP_URL_IS_DEFAULT_PORT(_url) \
(!OP_URL_IS_SSL(_url)&&(_url)->port==80 \
||OP_URL_IS_SSL(_url)&&(_url)->port==443)
struct OpusParsedURL{
/*Either "http" or "https".*/
char *scheme;
/*The user name from the <userinfo> component, or NULL.*/
char *user;
/*The password from the <userinfo> component, or NULL.*/
char *pass;
/*The <host> component.
This may not be NULL.*/
char *host;
/*The <path> and <query> components.
This may not be NULL.*/
char *path;
/*The <port> component.
This is set to the default port if the URL did not contain one.*/
unsigned port;
};
/*Parse a URL.
This code is not meant to be fast: strspn() with large sets is likely to be
slow, but it is very convenient.
It is meant to be RFC 3986-compliant.
We currently do not support IRIs (Internationalized Resource Identifiers,
RFC 3987).
Callers should translate them to URIs first.*/
static int op_parse_url_impl(OpusParsedURL *_dst,const char *_src){
const char *scheme_end;
const char *authority;
const char *userinfo_end;
const char *user;
const char *user_end;
const char *pass;
const char *hostport;
const char *hostport_end;
const char *host_end;
const char *port;
opus_int32 port_num;
const char *port_end;
const char *path;
const char *path_end;
const char *uri_end;
scheme_end=_src+strspn(_src,OP_URL_SCHEME);
if(OP_UNLIKELY(*scheme_end!=':')
||OP_UNLIKELY(scheme_end-_src<4)||OP_UNLIKELY(scheme_end-_src>5)
||OP_UNLIKELY(op_strncasecmp(_src,"https",scheme_end-_src)!=0)){
/*Unsupported protocol.*/
return OP_EIMPL;
}
if(OP_UNLIKELY(scheme_end[1]!='/')||OP_UNLIKELY(scheme_end[2]!='/')){
/*We require an <authority> component.*/
return OP_EINVAL;
}
authority=scheme_end+3;
/*Make sure all escape sequences are valid to simplify unescaping later.*/
if(OP_UNLIKELY(op_validate_url_escapes(authority)<0))return OP_EINVAL;
/*Look for a <userinfo> component.*/
userinfo_end=authority+strspn(authority,OP_URL_PCHAR_NA);
if(*userinfo_end=='@'){
/*Found one.*/
user=authority;
/*Look for a password (yes, clear-text passwords are deprecated, I know,
but what else are people supposed to use? use SSL if you care).*/
user_end=authority+strspn(authority,OP_URL_PCHAR_BASE);
if(*user_end==':')pass=user_end+1;
else pass=NULL;
hostport=userinfo_end+1;
}
else{
/*We shouldn't have to initialize user_end, but gcc is too dumb to figure
out that user!=NULL below means we didn't take this else branch.*/
user=user_end=NULL;
pass=NULL;
hostport=authority;
}
/*Try to figure out where the <host> component ends.*/
if(hostport[0]=='['){
hostport++;
/*We have an <IP-literal>, which can contain colons.*/
hostport_end=host_end=hostport+strspn(hostport,OP_URL_PCHAR_NA);
if(OP_UNLIKELY(*hostport_end++!=']'))return OP_EINVAL;
}
/*Currently we don't support IDNA (RFC 5894), because I don't want to deal
with the policy about which domains should not be internationalized to
avoid confusing similarities.
Give this API Punycode (RFC 3492) domain names instead.*/
else hostport_end=host_end=hostport+strspn(hostport,OP_URL_PCHAR_BASE);
/*TODO: Validate host.*/
/*Is there a port number?*/
port_num=-1;
if(*hostport_end==':'){
int i;
port=hostport_end+1;
port_end=port+strspn(port,OP_URL_DIGIT);
path=port_end;
/*Not part of RFC 3986, but require port numbers in the range 0...65535.*/
if(OP_LIKELY(port_end-port>0)){
while(*port=='0')port++;
if(OP_UNLIKELY(port_end-port>5))return OP_EINVAL;
port_num=0;
for(i=0;i<port_end-port;i++)port_num=port_num*10+port[i]-'0';
if(OP_UNLIKELY(port_num>65535))return OP_EINVAL;
}
}
else path=hostport_end;
path_end=path+strspn(path,OP_URL_PATH);
/*If the path is not empty, it must begin with a '/'.*/
if(OP_LIKELY(path_end>path)&&OP_UNLIKELY(path[0]!='/'))return OP_EINVAL;
/*Consume the <query> component, if any (right now we don't split this out
from the <path> component).*/
if(*path_end=='?')path_end=path_end+strspn(path_end,OP_URL_QUERY_FRAG);
/*Discard the <fragment> component, if any.
This doesn't get sent to the server.
Some day we should add support for Media Fragment URIs
<http://www.w3.org/TR/media-frags/>.*/
if(*path_end=='#')uri_end=path_end+1+strspn(path_end+1,OP_URL_QUERY_FRAG);
else uri_end=path_end;
/*If there's anything left, this was not a valid URL.*/
if(OP_UNLIKELY(*uri_end!='\0'))return OP_EINVAL;
_dst->scheme=op_string_range_dup(_src,scheme_end);
if(OP_UNLIKELY(_dst->scheme==NULL))return OP_EFAULT;
op_string_tolower(_dst->scheme);
if(user!=NULL){
_dst->user=op_string_range_dup(user,user_end);
if(OP_UNLIKELY(_dst->user==NULL))return OP_EFAULT;
op_unescape_url_component(_dst->user);
/*Unescaping might have created a ':' in the username.
That's not allowed by RFC 2617's Basic Authentication Scheme.*/
if(OP_UNLIKELY(strchr(_dst->user,':')!=NULL))return OP_EINVAL;
}
else _dst->user=NULL;
if(pass!=NULL){
_dst->pass=op_string_range_dup(pass,userinfo_end);
if(OP_UNLIKELY(_dst->pass==NULL))return OP_EFAULT;
op_unescape_url_component(_dst->pass);
}
else _dst->pass=NULL;
_dst->host=op_string_range_dup(hostport,host_end);
if(OP_UNLIKELY(_dst->host==NULL))return OP_EFAULT;
if(port_num<0){
if(_src[4]=='s')port_num=443;
else port_num=80;
}
_dst->port=(unsigned)port_num;
/*RFC 2616 says an empty <abs-path> component is equivalent to "/", and we
MUST use the latter in the Request-URI.
Reserve space for the slash here.*/
if(path==path_end||path[0]=='?')path--;
_dst->path=op_string_range_dup(path,path_end);
if(OP_UNLIKELY(_dst->path==NULL))return OP_EFAULT;
/*And force-set it here.*/
_dst->path[0]='/';
return 0;
}
static void op_parsed_url_init(OpusParsedURL *_url){
memset(_url,0,sizeof(*_url));
}
static void op_parsed_url_clear(OpusParsedURL *_url){
_ogg_free(_url->scheme);
_ogg_free(_url->user);
_ogg_free(_url->pass);
_ogg_free(_url->host);
_ogg_free(_url->path);
}
static int op_parse_url(OpusParsedURL *_dst,const char *_src){
OpusParsedURL url;
int ret;
op_parsed_url_init(&url);
ret=op_parse_url_impl(&url,_src);
if(OP_UNLIKELY(ret<0))op_parsed_url_clear(&url);
else *_dst=*&url;
return ret;
}
/*A buffer to hold growing strings.
The main purpose of this is to consolidate allocation checks and simplify
cleanup on a failed allocation.*/
struct OpusStringBuf{
char *buf;
int nbuf;
int cbuf;
};
static void op_sb_init(OpusStringBuf *_sb){
_sb->buf=NULL;
_sb->nbuf=0;
_sb->cbuf=0;
}
static void op_sb_clear(OpusStringBuf *_sb){
_ogg_free(_sb->buf);
}
static int op_sb_ensure_capacity(OpusStringBuf *_sb,int _capacity){
char *buf;
int cbuf;
buf=_sb->buf;
cbuf=_sb->cbuf;
if(_capacity>=cbuf-1){
if(OP_UNLIKELY(cbuf>INT_MAX-1>>1))return OP_EFAULT;
if(OP_UNLIKELY(_capacity>=INT_MAX-1))return OP_EFAULT;
cbuf=OP_MAX(2*cbuf+1,_capacity+1);
buf=_ogg_realloc(buf,sizeof(*buf)*cbuf);
if(OP_UNLIKELY(buf==NULL))return OP_EFAULT;
_sb->buf=buf;
_sb->cbuf=cbuf;
}
return 0;
}
static int op_sb_grow(OpusStringBuf *_sb,int _max_size){
char *buf;
int cbuf;
buf=_sb->buf;
cbuf=_sb->cbuf;
OP_ASSERT(_max_size<=INT_MAX-1);
cbuf=cbuf<=_max_size-1>>1?2*cbuf+1:_max_size+1;
buf=_ogg_realloc(buf,sizeof(*buf)*cbuf);
if(OP_UNLIKELY(buf==NULL))return OP_EFAULT;
_sb->buf=buf;
_sb->cbuf=cbuf;
return 0;
}
static int op_sb_append(OpusStringBuf *_sb,const char *_s,int _len){
char *buf;
int nbuf;
int ret;
nbuf=_sb->nbuf;
if(OP_UNLIKELY(nbuf>INT_MAX-_len))return OP_EFAULT;
ret=op_sb_ensure_capacity(_sb,nbuf+_len);
if(OP_UNLIKELY(ret<0))return ret;
buf=_sb->buf;
memcpy(buf+nbuf,_s,sizeof(*buf)*_len);
nbuf+=_len;
buf[nbuf]='\0';
_sb->nbuf=nbuf;
return 0;
}
static int op_sb_append_string(OpusStringBuf *_sb,const char *_s){
return op_sb_append(_sb,_s,strlen(_s));
}
static int op_sb_append_port(OpusStringBuf *_sb,unsigned _port){
char port_buf[7];
OP_ASSERT(_port<=65535U);
sprintf(port_buf,":%u",_port);
return op_sb_append_string(_sb,port_buf);
}
static int op_sb_append_nonnegative_int64(OpusStringBuf *_sb,opus_int64 _i){
char digit;
int nbuf_start;
int ret;
OP_ASSERT(_i>=0);
nbuf_start=_sb->nbuf;
ret=0;
do{
digit='0'+_i%10;
ret|=op_sb_append(_sb,&digit,1);
_i/=10;
}
while(_i>0);
if(OP_LIKELY(ret>=0)){
char *buf;
int nbuf_end;
buf=_sb->buf;
nbuf_end=_sb->nbuf-1;
/*We've added the digits backwards.
Reverse them.*/
while(nbuf_start<nbuf_end){
digit=buf[nbuf_start];
buf[nbuf_start]=buf[nbuf_end];
buf[nbuf_end]=digit;
nbuf_start++;
nbuf_end--;
}
}
return ret;
}
static struct addrinfo *op_resolve(const char *_host,unsigned _port){
struct addrinfo *addrs;
struct addrinfo hints;
char service[6];
memset(&hints,0,sizeof(hints));
hints.ai_socktype=SOCK_STREAM;
hints.ai_flags=AI_NUMERICSERV;
OP_ASSERT(_port<=65535U);
sprintf(service,"%u",_port);
if(OP_LIKELY(!getaddrinfo(_host,service,&hints,&addrs)))return addrs;
return NULL;
}
static int op_sock_set_nonblocking(int _fd,int _nonblocking){
int flags;
flags=fcntl(_fd,F_GETFL);
if(OP_UNLIKELY(flags<0))return flags;
if(_nonblocking)flags|=O_NONBLOCK;
else flags&=~O_NONBLOCK;
return fcntl(_fd,F_SETFL,flags);
}
/*Disable/enable write coalescing if we can.
We always send whole requests at once and always parse the response headers
before sending another one, so normally write coalescing just causes added
delay.*/
static void op_sock_set_tcp_nodelay(int _fd,int _nodelay){
# if defined(TCP_NODELAY)&&(defined(IPPROTO_TCP)||defined(SOL_TCP))
# if defined(IPPROTO_TCP)
# define OP_SO_LEVEL IPPROTO_TCP
# else
# define OP_SO_LEVEL SOL_TCP
# endif
/*It doesn't really matter if this call fails, but it would be interesting
to hit a case where it does.*/
OP_ALWAYS_TRUE(!setsockopt(_fd,OP_SO_LEVEL,TCP_NODELAY,
&_nodelay,sizeof(_nodelay)));
# endif
}
/*A single physical connection to an HTTP server.
We may have several of these open at once.*/
struct OpusHTTPConn{
/*The current position indicator for this connection.*/
opus_int64 pos;
/*The position where the current request will end, or -1 if we're reading
until EOF (an unseekable stream or the initial HTTP/1.0 request).*/
opus_int64 end_pos;
/*The position where next request we've sent will start, or -1 if we haven't
sent the next request yet.*/
opus_int64 next_pos;
/*The end of the next request or -1 if we requested the rest of the resource.
This is only set to a meaningful value if next_pos is not -1.*/
opus_int64 next_end;
/*The SSL connection, if this is https.*/
SSL *ssl_conn;
/*The next connection in either the LRU or free list.*/
OpusHTTPConn *next;
/*The last time we blocked for reading from this connection.*/
struct timeb read_time;
/*The number of bytes we've read since the last time we blocked.*/
opus_int64 read_bytes;
/*The estimated throughput of this connection, in bytes/s.*/
opus_int64 read_rate;
/*The socket we're reading from.*/
int fd;
/*The number of remaining requests we are allowed on this connection.*/
int nrequests_left;
/*The chunk size to use for pipelining requests.*/
opus_int32 chunk_size;
};
static void op_http_conn_init(OpusHTTPConn *_conn){
_conn->next_pos=-1;
_conn->ssl_conn=NULL;
_conn->next=NULL;
_conn->fd=-1;
}
static void op_http_conn_clear(OpusHTTPConn *_conn){
if(_conn->ssl_conn!=NULL)SSL_free(_conn->ssl_conn);
/*SSL frees the BIO for us.*/
if(_conn->fd>=0)close(_conn->fd);
}
/*The global stream state.*/
struct OpusHTTPStream{
/*The list of connections.*/
OpusHTTPConn conns[OP_NCONNS_MAX];
/*The context object used as a framework for TLS/SSL functions.*/
SSL_CTX *ssl_ctx;
/*The cached session to reuse for future connections.*/
SSL_SESSION *ssl_session;
/*The LRU list (ordered from MRU to LRU) of currently connected
connections.*/
OpusHTTPConn *lru_head;
/*The free list.*/
OpusHTTPConn *free_head;
/*The URL to connect to.*/
OpusParsedURL url;
/*Information about the address we connected to.*/
struct addrinfo addr_info;
/*The address we connected to.*/
union{
struct sockaddr s;
struct sockaddr_in v4;
struct sockaddr_in6 v6;
} addr;
/*A buffer used to build HTTP requests.*/
OpusStringBuf request;
/*A buffer used to build proxy CONNECT requests.*/
OpusStringBuf proxy_connect;
/*A buffer used to receive the response headers.*/
OpusStringBuf response;
/*The Content-Length, if specified, or -1 otherwise.
This will always be specified for seekable streams.*/
opus_int64 content_length;
/*The position indicator used when no connection is active.*/
opus_int64 pos;
/*The connection we're currently reading from.
This can be -1 if no connection is active.*/
int cur_conni;
/*Whether or not the server supports range requests.*/
int seekable;
/*Whether or not the server supports HTTP/1.1 with persistent connections.*/
int pipeline;
/*Whether or not we should skip certificate checks.*/
int skip_certificate_check;
/*The offset of the tail of the request.
Only the offset in the Range: header appears after this, allowing us to
quickly edit the request to ask for a new range.*/
int request_tail;
/*The estimated time required to open a new connection, in milliseconds.*/
opus_int32 connect_rate;
};
static void op_http_stream_init(OpusHTTPStream *_stream){
OpusHTTPConn **pnext;
int ci;
pnext=&_stream->free_head;
for(ci=0;ci<OP_NCONNS_MAX;ci++){
op_http_conn_init(_stream->conns+ci);
*pnext=_stream->conns+ci;
pnext=&_stream->conns[ci].next;
}
_stream->ssl_ctx=NULL;
_stream->ssl_session=NULL;
_stream->lru_head=NULL;
op_parsed_url_init(&_stream->url);
op_sb_init(&_stream->request);
op_sb_init(&_stream->proxy_connect);
op_sb_init(&_stream->response);
_stream->seekable=0;
}
/*Close the connection and move it to the free list.
_stream: The stream containing the free list.
_conn: The connection to close.
_penxt: The linked-list pointer currently pointing to this connection.
_gracefully: Whether or not to shut down cleanly.*/
static void op_http_conn_close(OpusHTTPStream *_stream,OpusHTTPConn *_conn,
OpusHTTPConn **_pnext,int _gracefully){
/*If we don't shut down gracefully, the server MUST NOT re-use our session
according to RFC 2246, because it can't tell the difference between an
abrupt close and a truncation attack.
So we shut down gracefully if we can.
However, we will not wait if this would block (it's not worth the savings
from session resumption to do so).
Clients (that's us) MAY resume a TLS session that ended with an incomplete
close, according to RFC 2818, so that's no reason to make sure the server
shut things down gracefully.
It also says "client implementations MUST treat any premature closes as
errors and the data received as potentially truncated," but libopusfile
treats errors and potentially truncated data in unseekable streams just
like a normal EOF.
We warn about this in the docs, and give some suggestions if you truly want
to avoid truncation attacks.*/
if(_gracefully&&_conn->ssl_conn!=NULL)SSL_shutdown(_conn->ssl_conn);
op_http_conn_clear(_conn);
_conn->next_pos=-1;
_conn->ssl_conn=NULL;
_conn->fd=-1;
OP_ASSERT(*_pnext==_conn);
*_pnext=_conn->next;
_conn->next=_stream->free_head;
_stream->free_head=_conn;
}
static void op_http_stream_clear(OpusHTTPStream *_stream){
while(_stream->lru_head!=NULL){
op_http_conn_close(_stream,_stream->lru_head,&_stream->lru_head,0);
}
if(_stream->ssl_session!=NULL)SSL_SESSION_free(_stream->ssl_session);
if(_stream->ssl_ctx!=NULL)SSL_CTX_free(_stream->ssl_ctx);
op_sb_clear(&_stream->response);
op_sb_clear(&_stream->proxy_connect);
op_sb_clear(&_stream->request);
op_parsed_url_clear(&_stream->url);
}
static int op_http_conn_write_fully(OpusHTTPConn *_conn,
const char *_buf,int _buf_size){
struct pollfd fd;
SSL *ssl_conn;
fd.fd=_conn->fd;
ssl_conn=_conn->ssl_conn;
while(_buf_size>0){
int err;
if(ssl_conn!=NULL){
int ret;
ret=SSL_write(ssl_conn,_buf,_buf_size);
if(ret>0){
/*Wrote some data.*/
_buf+=ret;
_buf_size-=ret;
continue;
}
/*Connection closed.*/
else if(ret==0)return OP_FALSE;
err=SSL_get_error(ssl_conn,ret);
/*Yes, renegotiations can cause SSL_write() to block for reading.*/
if(err==SSL_ERROR_WANT_READ)fd.events=POLLIN;
else if(err==SSL_ERROR_WANT_WRITE)fd.events=POLLOUT;
else return OP_FALSE;
}
else{
ssize_t ret;
errno=0;
ret=write(fd.fd,_buf,_buf_size);
if(ret>0){
_buf+=ret;
_buf_size-=ret;
continue;
}
err=errno;
if(err!=EAGAIN&&err!=EWOULDBLOCK)return OP_FALSE;
fd.events=POLLOUT;
}
if(poll(&fd,1,OP_POLL_TIMEOUT_MS)<=0)return OP_FALSE;
}
return 0;
}
static int op_http_conn_estimate_available(OpusHTTPConn *_conn){
int available;
int ret;
ret=ioctl(_conn->fd,FIONREAD,&available);
if(ret<0)available=0;
/*This requires the SSL read_ahead flag to be unset to work.
We ignore partial records as well as the protocol overhead for any pending
bytes.
This means we might return somewhat less than can truly be read without
blocking (if there's a partial record).
This is okay, because we're using this value to estimate network transfer
time, and we _have_ already received those bytes.
We also might return slightly more (due to protocol overhead), but that's
small enough that it probably doesn't matter.*/
if(_conn->ssl_conn!=NULL)available+=SSL_pending(_conn->ssl_conn);
return available;
}
static opus_int32 op_time_diff_ms(const struct timeb *_end,
const struct timeb *_start){
opus_int64 dtime;
dtime=_end->time-_start->time;
OP_ASSERT(_end->millitm<1000);
OP_ASSERT(_start->millitm<1000);
if(OP_UNLIKELY(dtime>(0x7FFFFFFF-1000)/1000))return 0x7FFFFFFF;
if(OP_UNLIKELY(dtime<(-0x7FFFFFFF+999)/1000))return -0x7FFFFFFF-1;
return (opus_int32)dtime*1000+_end->millitm-_start->millitm;
}
/*Update the read rate estimate for this connection.*/
static void op_http_conn_read_rate_update(OpusHTTPConn *_conn){
struct timeb read_time;
opus_int32 read_delta_ms;
opus_int64 read_delta_bytes;
opus_int64 read_rate;
read_delta_bytes=_conn->read_bytes;
if(read_delta_bytes<=0)return;
OP_ALWAYS_TRUE(!ftime(&read_time));
read_delta_ms=op_time_diff_ms(&read_time,&_conn->read_time);
read_rate=_conn->read_rate;
read_delta_ms=OP_MAX(read_delta_ms,1);
read_rate+=read_delta_bytes*1000/read_delta_ms-read_rate+4>>3;
*&_conn->read_time=*&read_time;
_conn->read_bytes=0;
_conn->read_rate=read_rate;
}
/*Tries to read from the given connection.
[out] _buf: Returns the data read.
_buf_size: The size of the buffer.
_blocking: Whether or not to block until some data is retrieved.
Return: A positive number of bytes read on success.
0: The read would block, or the connection was closed.
OP_EREAD: There was a fatal read error.*/
static int op_http_conn_read(OpusHTTPConn *_conn,
char *_buf,int _buf_size,int _blocking){
struct pollfd fd;
SSL *ssl_conn;
int nread;
int nread_unblocked;
fd.fd=_conn->fd;
ssl_conn=_conn->ssl_conn;
nread=nread_unblocked=0;
do{
int err;
if(ssl_conn!=NULL){
int ret;
ret=SSL_read(ssl_conn,_buf+nread,_buf_size-nread);
OP_ASSERT(ret<=_buf_size-nread);
if(ret>0){
/*Read some data.
Keep going to see if there's more.*/
nread+=ret;
nread_unblocked+=ret;
continue;
}
/*If we already read some data, return it right now.*/
if(nread>0)break;
err=SSL_get_error(ssl_conn,ret);
if(ret==0){
/*Connection close.
Check for a clean shutdown to prevent truncation attacks.
This check always succeeds for SSLv2, as it has no "close notify"
message and thus can't verify an orderly shutdown.*/
return err==SSL_ERROR_ZERO_RETURN?0:OP_EREAD;
}
if(err==SSL_ERROR_WANT_READ)fd.events=POLLIN;
/*Yes, renegotiations can cause SSL_read() to block for writing.*/
else if(err==SSL_ERROR_WANT_WRITE)fd.events=POLLOUT;
/*Some other error.*/
else return OP_EREAD;
}
else{
ssize_t ret;
errno=0;
ret=read(fd.fd,_buf+nread,_buf_size-nread);
OP_ASSERT(ret<=_buf_size-nread);
if(ret>0){
/*Read some data.
Keep going to see if there's more.*/
nread+=ret;
nread_unblocked+=ret;
continue;
}
/*If we already read some data or the connection was closed, return
right now.*/
if(ret==0||nread>0)break;
err=errno;
if(err!=EAGAIN&&err!=EWOULDBLOCK)return OP_EREAD;
fd.events=POLLIN;
}
_conn->read_bytes+=nread_unblocked;
op_http_conn_read_rate_update(_conn);
nread_unblocked=0;
if(!_blocking)break;
/*Need to wait to get any data at all.*/
if(poll(&fd,1,OP_POLL_TIMEOUT_MS)<=0)return OP_EREAD;
}
while(nread<_buf_size);
_conn->read_bytes+=nread_unblocked;
return nread;
}
/*Tries to look at the pending data for a connection without consuming it.
[out] _buf: Returns the data at which we're peeking.
_buf_size: The size of the buffer.*/
static int op_http_conn_peek(OpusHTTPConn *_conn,
char *_buf,int _buf_size){
struct pollfd fd;
SSL *ssl_conn;
int ret;
fd.fd=_conn->fd;
ssl_conn=_conn->ssl_conn;
for(;;){
int err;
if(ssl_conn!=NULL){
ret=SSL_peek(ssl_conn,_buf,_buf_size);
/*Either saw some data or the connection was closed.*/
if(ret>=0)return ret;
err=SSL_get_error(ssl_conn,ret);
if(err==SSL_ERROR_WANT_READ)fd.events=POLLIN;
/*Yes, renegotiations can cause SSL_peek() to block for writing.*/
else if(err==SSL_ERROR_WANT_WRITE)fd.events=POLLOUT;
else return 0;
}
else{
errno=0;
ret=(int)recv(fd.fd,_buf,_buf_size,MSG_PEEK);
/*Either saw some data or the connection was closed.*/
if(ret>=0)return ret;
err=errno;
if(err!=EAGAIN&&err!=EWOULDBLOCK)return 0;
fd.events=POLLIN;
}
/*Need to wait to get any data at all.*/
if(poll(&fd,1,OP_POLL_TIMEOUT_MS)<=0)return 0;
}
}
/*When parsing response headers, RFC 2616 mandates that all lines end in CR LF.
However, even in the year 2012, I have seen broken servers use just a LF.
This is the evil that Postel's advice from RFC 761 breeds.*/
/*Reads the entirety of a response to an HTTP request into the response buffer.
Actual parsing and validation is done later.
Return: The number of bytes in the response on success, OP_EREAD if the
connection was closed before reading any data, or another negative
value on any other error.*/
static int op_http_conn_read_response(OpusHTTPConn *_conn,
OpusStringBuf *_response){
int ret;
_response->nbuf=0;
ret=op_sb_ensure_capacity(_response,OP_RESPONSE_SIZE_MIN);
if(OP_UNLIKELY(ret<0))return ret;
for(;;){
char *buf;
int size;
int capacity;
int read_limit;
int terminated;
size=_response->nbuf;
capacity=_response->cbuf-1;
if(OP_UNLIKELY(size>=capacity)){
ret=op_sb_grow(_response,OP_RESPONSE_SIZE_MAX);
if(OP_UNLIKELY(ret<0))return ret;
capacity=_response->cbuf-1;
/*The response was too large.
This prevents a bad server from running us out of memory.*/
if(OP_UNLIKELY(size>=capacity))return OP_EIMPL;
}
buf=_response->buf;
ret=op_http_conn_peek(_conn,buf+size,capacity-size);
if(OP_UNLIKELY(ret<=0))return size<=0?OP_EREAD:OP_FALSE;
/*We read some data.*/
/*Make sure the starting characters are "HTTP".
Otherwise we could wind up waiting forever for a response from
something that is not an HTTP server.*/
if(size<4&&op_strncasecmp(buf,"HTTP",OP_MIN(size+ret,4))!=0){
return OP_FALSE;
}
/*How far can we read without passing the "\r\n\r\n" terminator?*/
buf[size+ret]='\0';
terminated=0;
for(read_limit=OP_MAX(size-3,0);read_limit<size+ret;read_limit++){
/*We don't look for the leading '\r' thanks to broken servers.*/
if(buf[read_limit]=='\n'){
if(buf[read_limit+1]=='\r'&&OP_LIKELY(buf[read_limit+2]=='\n')){
terminated=3;
break;
}
/*This case is for broken servers.*/
else if(OP_UNLIKELY(buf[read_limit+1]=='\n')){
terminated=2;
break;
}
}
}
read_limit+=terminated;
OP_ASSERT(size<=read_limit);
OP_ASSERT(read_limit<=size+ret);
/*Actually consume that data.*/
ret=op_http_conn_read(_conn,buf+size,read_limit-size,1);
if(OP_UNLIKELY(ret<=0))return OP_FALSE;
size+=ret;
buf[size]='\0';
_response->nbuf=size;
/*We found the terminator and read all the data up to and including it.*/
if(terminated&&OP_LIKELY(size>=read_limit))return size;
}
return OP_EIMPL;
}
# define OP_HTTP_DIGIT "0123456789"
/*The Reason-Phrase is not allowed to contain control characters, except
horizontal tab (HT: \011).*/
# define OP_HTTP_CREASON_PHRASE \
"\001\002\003\004\005\006\007\010\012\013\014\015\016\017\020\021" \
"\022\023\024\025\026\027\030\031\032\033\034\035\036\037\177"
# define OP_HTTP_CTLS \
"\001\002\003\004\005\006\007\010\011\012\013\014\015\016\017\020" \
"\021\022\023\024\025\026\027\030\031\032\033\034\035\036\037\177"
/*This also includes '\t', but we get that from OP_HTTP_CTLS.*/
# define OP_HTTP_SEPARATORS " \"(),/:;<=>?@[\\]{}"
/*TEXT can also include LWS, but that has structure, so we parse it
separately.*/
# define OP_HTTP_CTOKEN OP_HTTP_CTLS OP_HTTP_SEPARATORS
/*Return: The amount of linear white space (LWS) at the start of _s.*/
static int op_http_lwsspn(const char *_s){
int i;
for(i=0;;){
if(_s[0]=='\r'&&_s[1]=='\n'&&(_s[2]=='\t'||_s[2]==' '))i+=3;
/*This case is for broken servers.*/
else if(_s[0]=='\n'&&(_s[1]=='\t'||_s[1]==' '))i+=2;
else if(_s[i]=='\t'||_s[i]==' ')i++;
else return i;
}
}
static char *op_http_parse_status_line(int *_v1_1_compat,
char **_status_code,char *_response){
char *next;
char *status_code;
int v1_1_compat;
size_t d;
/*RFC 2616 Section 6.1 does not say that the tokens in the Status-Line cannot
be separated by optional LWS, but since it specifically calls out where
spaces are to be placed and that CR and LF are not allowed except at the
end, I am assuming this to be true.*/
/*We already validated that this starts with "HTTP"*/
OP_ASSERT(op_strncasecmp(_response,"HTTP",4)==0);
next=_response+4;
if(OP_UNLIKELY(*next++!='/'))return NULL;
d=strspn(next,OP_HTTP_DIGIT);
/*"Leading zeros MUST be ignored by recipients."*/
while(*next=='0'){
next++;
OP_ASSERT(d>0);
d--;
}
/*We only support version 1.x*/
if(OP_UNLIKELY(d!=1)||OP_UNLIKELY(*next++!='1'))return NULL;
if(OP_UNLIKELY(*next++!='.'))return NULL;
d=strspn(next,OP_HTTP_DIGIT);
if(OP_UNLIKELY(d<=0))return NULL;
/*"Leading zeros MUST be ignored by recipients."*/
while(*next=='0'){
next++;
OP_ASSERT(d>0);
d--;
}
/*We don't need to parse the version number.
Any non-zero digit means it's greater than 1.*/
v1_1_compat=d>0;
next+=d;
if(OP_UNLIKELY(*next++!=' '))return NULL;
status_code=next;
d=strspn(next,OP_HTTP_DIGIT);
if(OP_UNLIKELY(d!=3))return NULL;
next+=d;
/*The Reason-Phrase can be empty, but the space must be here.*/
if(OP_UNLIKELY(*next++!=' '))return NULL;
next+=strcspn(next,OP_HTTP_CREASON_PHRASE);
/*We are not mandating this be present thanks to broken servers.*/
if(OP_LIKELY(*next=='\r'))next++;
if(OP_UNLIKELY(*next++!='\n'))return NULL;
if(_v1_1_compat!=NULL)*_v1_1_compat=v1_1_compat;
*_status_code=status_code;
return next;
}
/*Get the next response header.
[out] _header: The header token, NUL-terminated, with leading and trailing
whitespace stripped, and converted to lower case (to simplify
case-insensitive comparisons), or NULL if there are no more
response headers.
[out] _cdr: The remaining contents of the header, excluding the initial
colon (':') and the terminating CRLF ("\r\n"),
NUL-terminated, and with leading and trailing whitespace
stripped, or NULL if there are no more response headers.
[inout] _s: On input, this points to the start of the current line of the
response headers.
On output, it points to the start of the first line following
this header, or NULL if there are no more response headers.
Return: 0 on success, or a negative value on failure.*/
static int op_http_get_next_header(char **_header,char **_cdr,char **_s){
char *header;
char *header_end;
char *cdr;
char *cdr_end;
char *next;
size_t d;
next=*_s;
/*The second case is for broken servers.*/
if(next[0]=='\r'&&next[1]=='\n'||OP_UNLIKELY(next[0]=='\n')){
/*No more headers.*/
*_header=NULL;
*_cdr=NULL;
*_s=NULL;
return 0;
}
header=next+op_http_lwsspn(next);
d=strcspn(header,OP_HTTP_CTOKEN);
if(OP_UNLIKELY(d<=0))return OP_FALSE;
header_end=header+d;
next=header_end+op_http_lwsspn(header_end);
if(OP_UNLIKELY(*next++!=':'))return OP_FALSE;
next+=op_http_lwsspn(next);
cdr=next;
do{
cdr_end=next+strcspn(next,OP_HTTP_CTLS);
next=cdr_end+op_http_lwsspn(cdr_end);
}
while(next>cdr_end);
/*We are not mandating this be present thanks to broken servers.*/
if(OP_LIKELY(*next=='\r'))next++;
if(OP_UNLIKELY(*next++!='\n'))return OP_FALSE;
*header_end='\0';
*cdr_end='\0';
/*Field names are case-insensitive.*/
op_string_tolower(header);
*_header=header;
*_cdr=cdr;
*_s=next;
return 0;
}
static opus_int64 op_http_parse_nonnegative_int64(const char **_next,
const char *_cdr){
const char *next;
opus_int64 content_length;
int i;
next=_cdr+strspn(_cdr,OP_HTTP_DIGIT);
*_next=next;
if(OP_UNLIKELY(next<=_cdr))return OP_FALSE;
while(*_cdr=='0')_cdr++;
if(OP_UNLIKELY(next-_cdr>19))return OP_EIMPL;
content_length=0;
for(i=0;i<next-_cdr;i++){
int digit;
digit=_cdr[i]-'0';
/*Check for overflow.*/
if(OP_UNLIKELY(content_length>(OP_INT64_MAX-9)/10+(digit<=7))){
return OP_EIMPL;
}
content_length=content_length*10+digit;
}
return content_length;
}
static opus_int64 op_http_parse_content_length(const char *_cdr){
const char *next;
opus_int64 content_length;
content_length=op_http_parse_nonnegative_int64(&next,_cdr);
if(OP_UNLIKELY(*next!='\0'))return OP_FALSE;
return content_length;
}
static int op_http_parse_content_range(opus_int64 *_first,opus_int64 *_last,
opus_int64 *_length,const char *_cdr){
opus_int64 first;
opus_int64 last;
opus_int64 length;
size_t d;
if(OP_UNLIKELY(op_strncasecmp(_cdr,"bytes",5)!=0))return OP_FALSE;
_cdr+=5;
d=op_http_lwsspn(_cdr);
if(OP_UNLIKELY(d<=0))return OP_FALSE;
_cdr+=d;
if(*_cdr!='*'){
first=op_http_parse_nonnegative_int64(&_cdr,_cdr);
if(OP_UNLIKELY(first<0))return (int)first;
_cdr+=op_http_lwsspn(_cdr);
if(*_cdr++!='-')return OP_FALSE;
_cdr+=op_http_lwsspn(_cdr);
last=op_http_parse_nonnegative_int64(&_cdr,_cdr);
if(OP_UNLIKELY(last<0))return (int)last;
_cdr+=op_http_lwsspn(_cdr);
}
else{
/*This is for a 416 response (Requested range not satisfiable).*/
first=last=-1;
_cdr++;
}
if(OP_UNLIKELY(*_cdr++!='/'))return OP_FALSE;
if(*_cdr!='*'){
length=op_http_parse_nonnegative_int64(&_cdr,_cdr);
if(OP_UNLIKELY(length<0))return (int)length;
}
else{
/*The total length is unspecified.*/
_cdr++;
length=-1;
}
if(OP_UNLIKELY(*_cdr!='\0'))return OP_FALSE;
if(OP_UNLIKELY(last<first))return OP_FALSE;
if(length>=0&&OP_UNLIKELY(last>=length))return OP_FALSE;
*_first=first;
*_last=last;
*_length=length;
return 0;
}
/*Parse the Connection response header and look for a "close" token.
Return: 1 if a "close" token is found, 0 if it's not found, and a negative
value on error.*/
static int op_http_parse_connection(char *_cdr){
size_t d;
int ret;
ret=0;
for(;;){
d=strcspn(_cdr,OP_HTTP_CTOKEN);
if(OP_UNLIKELY(d<=0))return OP_FALSE;
if(op_strncasecmp(_cdr,"close",(int)d)==0)ret=1;
/*We're supposed to strip and ignore any headers mentioned in the
Connection header if this response is from an HTTP/1.0 server (to
work around forwarding of hop-by-hop headers by old proxies), but the
only hop-by-hop header we look at is Connection itself.
Everything else is a well-defined end-to-end header, and going back and
undoing the things we did based on already-examined headers would be
hard (since we only scan them once, in a destructive manner).
Therefore we just ignore all the other tokens.*/
_cdr+=d;
d=op_http_lwsspn(_cdr);
if(d<=0)break;
_cdr+=d;
}
return OP_UNLIKELY(*_cdr!='\0')?OP_FALSE:ret;
}
typedef int (*op_ssl_step_func)(SSL *_ssl_conn);
/*Try to run an SSL function to completion (blocking if necessary).*/
static int op_do_ssl_step(SSL *_ssl_conn,int _fd,op_ssl_step_func _step){
struct pollfd fd;
fd.fd=_fd;
for(;;){
int ret;
int err;
ret=(*_step)(_ssl_conn);
if(ret>=0)return ret;
err=SSL_get_error(_ssl_conn,ret);
if(err==SSL_ERROR_WANT_READ)fd.events=POLLIN;
else if(err==SSL_ERROR_WANT_WRITE)fd.events=POLLOUT;
else return OP_FALSE;
if(poll(&fd,1,OP_POLL_TIMEOUT_MS)<=0)return OP_FALSE;
}
}
/*Implement a BIO type that just indicates every operation should be retried.
We use this when initializing an SSL connection via a proxy to allow the
initial handshake to proceed all the way up to the first read attempt, and
then return.
This allows the TLS client hello message to be pipelined with the HTTP
CONNECT request.*/
static int op_bio_retry_write(BIO *_b,const char *_buf,int _num){
(void)_buf;
(void)_num;
BIO_clear_retry_flags(_b);
BIO_set_retry_write(_b);
return -1;
}
static int op_bio_retry_read(BIO *_b,char *_buf,int _num){
(void)_buf;
(void)_num;
BIO_clear_retry_flags(_b);
BIO_set_retry_read(_b);
return -1;
}
static int op_bio_retry_puts(BIO *_b,const char *_str){
return op_bio_retry_write(_b,_str,0);
}
static long op_bio_retry_ctrl(BIO *_b,int _cmd,long _num,void *_ptr){
long ret;
(void)_b;
(void)_num;
(void)_ptr;
ret=0;
switch(_cmd){
case BIO_CTRL_RESET:
case BIO_C_RESET_READ_REQUEST:{
BIO_clear_retry_flags(_b);
/*Fall through.*/
}
case BIO_CTRL_EOF:
case BIO_CTRL_SET:
case BIO_CTRL_SET_CLOSE:
case BIO_CTRL_FLUSH:
case BIO_CTRL_DUP:{
ret=1;
}break;
}
return ret;
}
static int op_bio_retry_new(BIO *_b){
_b->init=1;
_b->num=0;
_b->ptr=NULL;
return 1;
}
static int op_bio_retry_free(BIO *_b){
return _b!=NULL;
}
/*This is not const because OpenSSL doesn't allow it, even though it won't
write to it.*/
static BIO_METHOD op_bio_retry_method={
BIO_TYPE_NULL,
"retry",
op_bio_retry_write,
op_bio_retry_read,
op_bio_retry_puts,
NULL,
op_bio_retry_ctrl,
op_bio_retry_new,
op_bio_retry_free,
NULL
};
/*Establish a CONNECT tunnel and pipeline the start of the TLS handshake for
proxying https URL requests.*/
int op_http_conn_establish_tunnel(OpusHTTPStream *_stream,
OpusHTTPConn *_conn,int _fd,SSL *_ssl_conn,BIO *_ssl_bio){
BIO *retry_bio;
char *status_code;
char *next;
int ret;
_conn->ssl_conn=NULL;
_conn->fd=_fd;
OP_ASSERT(_stream->proxy_connect.nbuf>0);
ret=op_http_conn_write_fully(_conn,
_stream->proxy_connect.buf,_stream->proxy_connect.nbuf);
if(OP_UNLIKELY(ret<0))return ret;
retry_bio=BIO_new(&op_bio_retry_method);
if(OP_UNLIKELY(retry_bio==NULL))return OP_EFAULT;
SSL_set_bio(_ssl_conn,retry_bio,_ssl_bio);
SSL_set_connect_state(_ssl_conn);
/*This shouldn't succeed, since we can't read yet.*/
OP_ALWAYS_TRUE(SSL_connect(_ssl_conn)<0);
SSL_set_bio(_ssl_conn,_ssl_bio,_ssl_bio);
/*Only now do we disable write coalescing, to allow the CONNECT
request and the start of the TLS handshake to be combined.*/
op_sock_set_tcp_nodelay(_fd,1);
ret=op_http_conn_read_response(_conn,&_stream->response);
if(OP_UNLIKELY(ret<0))return ret;
next=op_http_parse_status_line(NULL,&status_code,_stream->response.buf);
/*According to RFC 2817, "Any successful (2xx) response to a
CONNECT request indicates that the proxy has established a
connection to the requested host and port.*/
if(OP_UNLIKELY(next==NULL)||OP_UNLIKELY(status_code[0]!='2'))return OP_FALSE;
return 0;
}
/*Match a host name against a host with a possible wildcard pattern according
to the rules of RFC 6125 Section 6.4.3.
Return: 0 if the pattern doesn't match, and a non-zero value if it does.*/
static int op_http_hostname_match(const char *_host,size_t _host_len,
ASN1_STRING *_pattern){
const char *pattern;
size_t host_label_len;
size_t host_suffix_len;
size_t pattern_len;
size_t pattern_label_len;
size_t pattern_prefix_len;
size_t pattern_suffix_len;
pattern=(const char *)ASN1_STRING_data(_pattern);
pattern_len=strlen(pattern);
/*Check the pattern for embedded NULs.*/
if(OP_UNLIKELY(pattern_len!=(size_t)ASN1_STRING_length(_pattern)))return 0;
pattern_label_len=strcspn(pattern,".");
OP_ASSERT(pattern_label_len<=pattern_len);
pattern_prefix_len=strcspn(pattern,"*");
if(pattern_prefix_len>=pattern_label_len){
/*"The client SHOULD NOT attempt to match a presented identifier in which
the wildcard character comprises a label other than the left-most label
(e.g., do not match bar.*.example.net)."*/
if(pattern_prefix_len<pattern_len)return 0;
/*If the pattern does not contain a wildcard in the first element, do an
exact match.
Don't use the system strcasecmp here, as that uses the locale and
RFC 4343 makes clear that DNS's case-insensitivity only applies to
the ASCII range.*/
return _host_len==pattern_len&&op_strncasecmp(_host,pattern,_host_len)==0;
}
/*"However, the client SHOULD NOT attempt to match a presented identifier
where the wildcard character is embedded within an A-label or U-label of
an internationalized domain name.*/
if(op_strncasecmp(pattern,"xn--",4)==0)return 0;
host_label_len=strcspn(_host,".");
/*Make sure the host has at least two dots, to prevent the wildcard match
from being ridiculously wide.
We should have already checked to ensure it had at least one.*/
if(OP_UNLIKELY(_host[host_label_len]!='.')
||strchr(_host+host_label_len+1,'.')==NULL){
return 0;
}
OP_ASSERT(host_label_len<_host_len);
/*"If the wildcard character is the only character of the left-most label in
the presented identifier, the client SHOULD NOT compare against anything
but the left-most label of the reference identifier (e.g., *.example.com
would match foo.example.com but not bar.foo.example.com)."
This is really confusingly worded, as we check this by actually comparing
the rest of the pattern for an exact match.
We also use the fact that the wildcard must match at least one character,
so the left-most label of the hostname must be at least as large as the
left-most label of the pattern.*/
if(host_label_len<pattern_label_len)return 0;
OP_ASSERT(pattern[pattern_prefix_len]=='*');
/*"The client MAY match a presented identifier in which the wildcard
character is not the only character of the label (e.g., baz*.example.net
and *baz.example.net and b*z.example.net would be taken to match
baz1.example.net and foobaz.example.net and buzz.example.net,
respectively)."*/
pattern_suffix_len=pattern_len-pattern_prefix_len-1;
host_suffix_len=_host_len-host_label_len
+pattern_label_len-pattern_prefix_len-1;
return pattern_suffix_len==host_suffix_len
&&op_strncasecmp(_host,pattern,pattern_prefix_len)==0
&&op_strncasecmp(_host+_host_len-host_suffix_len,
pattern+pattern_prefix_len+1,host_suffix_len)==0;
}
/*Convert a host to a numeric address, if possible.
Return: A struct addrinfo containing the address, if it was numeric, and NULL
otherise.*/
static struct addrinfo *op_inet_pton(const char *_host){
struct addrinfo *addrs;
struct addrinfo hints;
memset(&hints,0,sizeof(hints));
hints.ai_socktype=SOCK_STREAM;
hints.ai_flags=AI_NUMERICHOST;
if(!getaddrinfo(_host,NULL,&hints,&addrs))return addrs;
return NULL;
}
/*Verify the server's hostname matches the certificate they presented using
the procedure from Section 6 of RFC 6125.
Return: 0 if the certificate doesn't match, and a non-zero value if it does.*/
static int op_http_verify_hostname(OpusHTTPStream *_stream,
SSL *_ssl_conn){
X509 *peer_cert;
STACK_OF(GENERAL_NAME) *san_names;
char *host;
size_t host_len;
int ret;
host=_stream->url.host;
host_len=strlen(host);
peer_cert=SSL_get_peer_certificate(_ssl_conn);
/*We set VERIFY_PEER, so we shouldn't get here without a certificate.*/
if(OP_UNLIKELY(peer_cert==NULL))return 0;
ret=0;
OP_ASSERT(host_len<INT_MAX);
/*RFC 2818 says (after correcting for Eratta 1077): "If a subjectAltName
extension of type dNSName is present, that MUST be used as the identity.
Otherwise, the (most specific) Common Name field in the Subject field of
the certificate MUST be used.
Although the use of the Common Name is existing practice, it is deprecated
and Certification Authorities are encouraged to use the dNSName
instead."
"Matching is performed using the matching rules specified by RFC 2459.
If more than one identity of a given type is present in the certificate
(e.g., more than one dNSName name), a match in any one of the set is
considered acceptable.
Names may contain the wildcard character * which is condered to match any
single domain name component or component fragment.
E.g., *.a.com matches foo.a.com but not bar.foo.a.com.
f*.com matches foo.com but not bar.com."
"In some cases, the URI is specified as an IP address rather than a
hostname.
In this case, the iPAddress subjectAltName must be present in the
certificate and must exactly match the IP in the URI."*/
san_names=X509_get_ext_d2i(peer_cert,NID_subject_alt_name,NULL,NULL);
if(san_names!=NULL){
struct addrinfo *addr;
unsigned char *ip;
int ip_len;
int nsan_names;
int sni;
/*Check to see if the host was specified as a simple IP address.*/
addr=op_inet_pton(host);
ip=NULL;
ip_len=0;
if(addr!=NULL){
switch(addr->ai_family){
case AF_INET:{
struct sockaddr_in *s;
s=(struct sockaddr_in *)addr->ai_addr;
OP_ASSERT(addr->ai_addrlen>=sizeof(*s));
ip=(unsigned char *)&s->sin_addr;
ip_len=sizeof(s->sin_addr);
}break;
case AF_INET6:{
struct sockaddr_in6 *s;
s=(struct sockaddr_in6 *)addr->ai_addr;
OP_ASSERT(addr->ai_addrlen>=sizeof(*s));
ip=(unsigned char *)&s->sin6_addr;
ip_len=sizeof(s->sin6_addr);
}break;
}
}
/*We can only verify fully-qualified domain names.
To quote RFC 6125: "The extracted data MUST include only information that
can be securely parsed out of the inputs (e.g., parsing the fully
qualified DNS domain name out of the "host" component (or its
equivalent) of a URI or deriving the application service type from the
scheme of a URI) ..."
We don't have a way to check (without relying on DNS records, which might
be subverted), if this address is fully-qualified.
This is particularly problematic when using a CONNECT tunnel, as it is
the server that does DNS lookup, not us.
However, we are certain that if the hostname has no '.', it is definitely
not a fully-qualified domain name (with the exception of crazy TLDs that
actually resolve, like "uz", but I am willing to ignore those).
RFC 1535 says "...in any event where a '.' exists in a specified name it
should be assumed to be a fully qualified domain name (FQDN) and SHOULD
be tried as a rooted name first."
That doesn't give us any security guarantees, of course (a subverted DNS
could fail the original query and our resolver might still retry with a
local domain appended).
If we don't have a FQDN, just set the number of names to 0, so we'll fail
and clean up any resources we allocated.*/
if(ip==NULL&&strchr(host,'.')==NULL)nsan_names=0;
/*RFC 2459 says there MUST be at least one, but we don't depend on it.*/
else nsan_names=sk_GENERAL_NAME_num(san_names);
for(sni=0;sni<nsan_names;sni++){
const GENERAL_NAME *name;
name=sk_GENERAL_NAME_value(san_names,sni);
if(ip==NULL){
if(name->type==GEN_DNS
&&op_http_hostname_match(host,host_len,name->d.dNSName)){
ret=1;
break;
}
}
else if(name->type==GEN_IPADD){
unsigned char *cert_ip;
/*If we do have an IP address, compare it directly.
RFC 6125: "When the reference identity is an IP address, the identity
MUST be converted to the 'network byte order' octet string
representation.
For IP Version 4, as specified in RFC 791, the octet string will
contain exactly four octets.
For IP Version 6, as specified in RFC 2460, the octet string will
contain exactly sixteen octets.
This octet string is then compared against subjectAltName values of
type iPAddress.
A match occurs if the reference identity octet string and the value
octet strings are identical."*/
cert_ip=ASN1_STRING_data(name->d.iPAddress);
if(ip_len==ASN1_STRING_length(name->d.iPAddress)
&&memcmp(ip,cert_ip,ip_len)==0){
ret=1;
break;
}
}
}
sk_GENERAL_NAME_pop_free(san_names,GENERAL_NAME_free);
if(addr!=NULL)freeaddrinfo(addr);
}
/*Do the same FQDN check we did above.
We don't do this once in advance for both cases, because in the
subjectAltName case we might have an IPv6 address without a dot.*/
else if(strchr(host,'.')!=NULL){
int last_cn_loc;
int cn_loc;
/*If there is no subjectAltName, match against commonName.
RFC 6125 says that at least one significant CA is known to issue certs
with multiple CNs, although it SHOULD NOT.
It also says: "The server's identity may also be verified by comparing
the reference identity to the Common Name (CN) value in the last
Relative Distinguished Name (RDN) of the subject field of the server's
certificate (where "last" refers to the DER-encoded order...)."
So find the last one and check it.*/
cn_loc=-1;
do{
last_cn_loc=cn_loc;
cn_loc=X509_NAME_get_index_by_NID(X509_get_subject_name(peer_cert),
NID_commonName,last_cn_loc);
}
while(cn_loc>=0);
ret=last_cn_loc>=0
&&op_http_hostname_match(host,host_len,
X509_NAME_ENTRY_get_data(
X509_NAME_get_entry(X509_get_subject_name(peer_cert),last_cn_loc)));
}
X509_free(peer_cert);
return ret;
}
/*Perform the TLS handshake on a new connection.*/
int op_http_conn_start_tls(OpusHTTPStream *_stream,OpusHTTPConn *_conn,
int _fd,SSL *_ssl_conn){
SSL_SESSION *ssl_session;
BIO *ssl_bio;
int skip_certificate_check;
int ret;
ssl_bio=BIO_new_socket(_fd,BIO_NOCLOSE);
if(OP_LIKELY(ssl_bio==NULL))return OP_FALSE;
# if !defined(OPENSSL_NO_TLSEXT)
/*Support for RFC 6066 Server Name Indication.*/
SSL_set_tlsext_host_name(_ssl_conn,_stream->url.host);
# endif
/*Resume a previous session if available.*/
if(_stream->ssl_session!=NULL){
SSL_set_session(_ssl_conn,_stream->ssl_session);
}
/*If we're proxying, establish the CONNECT tunnel.*/
if(_stream->proxy_connect.nbuf>0){
ret=op_http_conn_establish_tunnel(_stream,_conn,
_fd,_ssl_conn,ssl_bio);
if(OP_UNLIKELY(ret<0))return ret;
}
else{
/*Otherwise, just use this socket directly.*/
op_sock_set_tcp_nodelay(_fd,1);
SSL_set_bio(_ssl_conn,ssl_bio,ssl_bio);
SSL_set_connect_state(_ssl_conn);
}
ret=op_do_ssl_step(_ssl_conn,_fd,SSL_connect);
if(OP_UNLIKELY(ret<=0))return OP_FALSE;
ssl_session=_stream->ssl_session;
skip_certificate_check=_stream->skip_certificate_check;
if(ssl_session==NULL||!skip_certificate_check){
ret=op_do_ssl_step(_ssl_conn,_fd,SSL_do_handshake);
if(OP_UNLIKELY(ret<=0))return OP_FALSE;
/*OpenSSL does not do hostname verification, despite the fact that we just
passed it the hostname above in the call to SSL_set_tlsext_host_name(),
because they are morons.
Do it for them.*/
if(!skip_certificate_check&&!op_http_verify_hostname(_stream,_ssl_conn)){
return OP_FALSE;
}
if(ssl_session==NULL){
/*Save the session for later resumption.*/
_stream->ssl_session=SSL_get1_session(_ssl_conn);
}
}
_conn->ssl_conn=_ssl_conn;
_conn->fd=_fd;
_conn->nrequests_left=OP_PIPELINE_MAX_REQUESTS;
return 0;
}
/*Try to start a connection to the next address in the given list of a given
type.
_fd: The socket to connect with.
[inout] _addr: A pointer to the list of addresses.
This will be advanced to the first one that matches the given
address family (possibly the current one).
_ai_family: The address family to connect to.
Return: 1 If the connection was successful.
0 If the connection is in progress.
OP_FALSE If the connection failed and there were no more addresses
left to try.
*_addr will be set to NULL in this case.*/
static int op_sock_connect_next(int _fd,
struct addrinfo **_addr,int _ai_family){
struct addrinfo *addr;
addr=*_addr;
for(;;){
/*Move to the next address of the requested type.*/
for(;addr!=NULL&&addr->ai_family!=_ai_family;addr=addr->ai_next);
*_addr=addr;
/*No more: failure.*/
if(addr==NULL)return OP_FALSE;
if(connect(_fd,addr->ai_addr,addr->ai_addrlen)>=0)return 1;
if(OP_LIKELY(errno==EINPROGRESS))return 0;
addr=addr->ai_next;
}
}
/*The number of address families to try connecting to simultaneously.*/
# define OP_NPROTOS (2)
static int op_http_connect(OpusHTTPStream *_stream,OpusHTTPConn *_conn,
struct addrinfo *_addrs,struct timeb *_start_time){
struct addrinfo *addr;
struct addrinfo *addrs[OP_NPROTOS];
struct pollfd fds[OP_NPROTOS];
int ai_family;
int nprotos;
int ret;
int pi;
int pj;
for(pi=0;pi<OP_NPROTOS;pi++)addrs[pi]=NULL;
addr=_addrs;
/*Try connecting via both IPv4 and IPv6 simultaneously, and keep the first
one that succeeds.*/
for(;addr!=NULL;addr=addr->ai_next){
/*Give IPv6 a slight edge by putting it first in the list.*/
if(addr->ai_family==AF_INET6){
OP_ASSERT(addr->ai_addrlen<=sizeof(struct sockaddr_in6));
if(addrs[0]==NULL)addrs[0]=addr;
}
else if(addr->ai_family==AF_INET){
OP_ASSERT(addr->ai_addrlen<=sizeof(struct sockaddr_in));
if(addrs[1]==NULL)addrs[1]=addr;
}
}
/*Consolidate the list of addresses.*/
for(pi=nprotos=0;pi<OP_NPROTOS;pi++){
if(addrs[pi]!=NULL){
addrs[nprotos]=addrs[pi];
nprotos++;
}
}
/*Pop the connection off the free list and put it on the LRU list.*/
OP_ASSERT(_stream->free_head==_conn);
_stream->free_head=_conn->next;
_conn->next=_stream->lru_head;
_stream->lru_head=_conn;
OP_ALWAYS_TRUE(!ftime(_start_time));
*&_conn->read_time=*_start_time;
_conn->read_bytes=0;
_conn->read_rate=0;
/*Try to start a connection to each protocol.*/
for(pi=0;pi<nprotos;pi++){
ai_family=addrs[pi]->ai_family;
fds[pi].fd=socket(ai_family,SOCK_STREAM,addrs[pi]->ai_protocol);
fds[pi].events=POLLOUT;
if(OP_LIKELY(fds[pi].fd>=0)){
if(OP_LIKELY(op_sock_set_nonblocking(fds[pi].fd,1)>=0)){
ret=op_sock_connect_next(fds[pi].fd,addrs+pi,ai_family);
if(OP_UNLIKELY(ret>0)){
/*It succeeded right away (technically possible), so stop.*/
nprotos=pi+1;
break;
}
/*Otherwise go on to the next protocol, and skip the clean-up below.*/
else if(ret==0)continue;
/*Tried all the addresses for this protocol.*/
}
/*Clean up the socket.*/
close(fds[pi].fd);
}
/*Remove this protocol from the list.*/
memmove(addrs+pi,addrs+pi+1,sizeof(*addrs)*(nprotos-pi-1));
nprotos--;
pi--;
}
/*Wait for one of the connections to finish.*/
while(pi>=nprotos&&nprotos>0&&poll(fds,nprotos,OP_POLL_TIMEOUT_MS)>0){
for(pi=0;pi<nprotos;pi++){
socklen_t errlen;
int err;
/*Still waiting...*/
if(!fds[pi].revents)continue;
errlen=sizeof(err);
/*Some platforms will return the pending error in &err and return 0.
Others will put it in errno and return -1.*/
ret=getsockopt(fds[pi].fd,SOL_SOCKET,SO_ERROR,&err,&errlen);
if(ret<0)err=errno;
/*Success!*/
if(err==0||err==EISCONN)break;
/*Move on to the next address for this protocol.*/
ai_family=addrs[pi]->ai_family;
addrs[pi]=addrs[pi]->ai_next;
ret=op_sock_connect_next(fds[pi].fd,addrs+pi,ai_family);
/*It succeeded right away, so stop.*/
if(ret>0)break;
/*Otherwise go on to the next protocol, and skip the clean-up below.*/
else if(ret==0)continue;
/*Tried all the addresses for this protocol.
Remove it from the list.*/
close(fds[pi].fd);
memmove(fds+pi,fds+pi+1,sizeof(*fds)*(nprotos-pi-1));
memmove(addrs+pi,addrs+pi+1,sizeof(*addrs)*(nprotos-pi-1));
nprotos--;
pi--;
}
}
/*Close all the other sockets.*/
for(pj=0;pj<nprotos;pj++)if(pi!=pj)close(fds[pj].fd);
/*If none of them succeeded, we're done.*/
if(pi>=nprotos)return OP_FALSE;
/*Save this address for future connection attempts.*/
if(addrs[pi]!=&_stream->addr_info){
memcpy(&_stream->addr_info,addrs[pi],sizeof(_stream->addr_info));
_stream->addr_info.ai_addr=&_stream->addr.s;
_stream->addr_info.ai_next=NULL;
memcpy(&_stream->addr,addrs[pi]->ai_addr,addrs[pi]->ai_addrlen);
}
if(OP_URL_IS_SSL(&_stream->url)){
SSL *ssl_conn;
/*Start the SSL connection.*/
OP_ASSERT(_stream->ssl_ctx!=NULL);
ssl_conn=SSL_new(_stream->ssl_ctx);
if(OP_LIKELY(ssl_conn!=NULL)){
ret=op_http_conn_start_tls(_stream,_conn,fds[pi].fd,ssl_conn);
if(OP_LIKELY(ret>=0))return ret;
SSL_free(ssl_conn);
}
close(fds[pi].fd);
_conn->fd=-1;
return OP_FALSE;
}
/*Just a normal non-SSL connection.*/
_conn->ssl_conn=NULL;
_conn->fd=fds[pi].fd;
_conn->nrequests_left=OP_PIPELINE_MAX_REQUESTS;
/*Disable write coalescing.
We always send whole requests at once and always parse the response headers
before sending another one.*/
op_sock_set_tcp_nodelay(fds[pi].fd,1);
return 0;
}
# define OP_BASE64_LENGTH(_len) (((_len)+2)/3*4)
static const char BASE64_TABLE[64]={
'A','B','C','D','E','F','G','H','I','J','K','L','M','N','O','P',
'Q','R','S','T','U','V','W','X','Y','Z','a','b','c','d','e','f',
'g','h','i','j','k','l','m','n','o','p','q','r','s','t','u','v',
'w','x','y','z','0','1','2','3','4','5','6','7','8','9','+','/'
};
static char *op_base64_encode(char *_dst,const char *_src,int _len){
unsigned s0;
unsigned s1;
unsigned s2;
int ngroups;
int i;
ngroups=_len/3;
for(i=0;i<ngroups;i++){
s0=_src[3*i+0];
s1=_src[3*i+1];
s2=_src[3*i+2];
_dst[4*i+0]=BASE64_TABLE[s0>>2];
_dst[4*i+1]=BASE64_TABLE[s0&3<<4|s1>>4];
_dst[4*i+2]=BASE64_TABLE[s1&15<<2|s2>>6];
_dst[4*i+3]=BASE64_TABLE[s2&63];
}
_len-=3*i;
if(_len==1){
s0=_src[3*i+0];
_dst[4*i+0]=BASE64_TABLE[s0>>2];
_dst[4*i+1]=BASE64_TABLE[s0&3<<4];
_dst[4*i+2]='=';
_dst[4*i+3]='=';
i++;
}
else if(_len==2){
s0=_src[3*i+0];
s1=_src[3*i+1];
_dst[4*i+0]=BASE64_TABLE[s0>>2];
_dst[4*i+1]=BASE64_TABLE[s0&3<<4|s1>>4];
_dst[4*i+2]=BASE64_TABLE[s1&15<<2];
_dst[4*i+3]='=';
i++;
}
_dst[4*i]='\0';
return _dst+4*i;
}
/*Construct an HTTP authorization header using RFC 2617's Basic Authentication
Scheme and append it to the given string buffer.*/
static int op_sb_append_basic_auth_header(OpusStringBuf *_sb,
const char *_header,const char *_user,const char *_pass){
int user_len;
int pass_len;
int user_pass_len;
int base64_len;
int nbuf_total;
int ret;
ret=op_sb_append_string(_sb,_header);
ret|=op_sb_append(_sb,": Basic ",8);
user_len=strlen(_user);
pass_len=strlen(_pass);
if(OP_UNLIKELY(pass_len>INT_MAX-user_len))return OP_EFAULT;
if(OP_UNLIKELY(user_len+pass_len>(INT_MAX>>2)*3-3))return OP_EFAULT;
user_pass_len=user_len+1+pass_len;
base64_len=OP_BASE64_LENGTH(user_pass_len);
/*Stick "user:pass" at the end of the buffer so we can Base64 encode it
in-place.*/
nbuf_total=_sb->nbuf;
if(OP_UNLIKELY(base64_len>INT_MAX-nbuf_total))return OP_EFAULT;
nbuf_total+=base64_len;
ret|=op_sb_ensure_capacity(_sb,nbuf_total);
if(OP_UNLIKELY(ret<0))return ret;
_sb->nbuf=nbuf_total-user_pass_len;
OP_ALWAYS_TRUE(!op_sb_append(_sb,_user,user_len));
OP_ALWAYS_TRUE(!op_sb_append(_sb,":",1));
OP_ALWAYS_TRUE(!op_sb_append(_sb,_pass,pass_len));
op_base64_encode(_sb->buf+nbuf_total-base64_len,
_sb->buf+nbuf_total-user_pass_len,user_pass_len);
return op_sb_append(_sb,"\r\n",2);
}
static int op_http_allow_pipelining(const char *_server){
/*Servers known to do bad things with pipelined requests.
This list is taken from Gecko's nsHttpConnection::SupportsPipelining() (in
netwerk/protocol/http/nsHttpConnection.cpp).*/
static const char *BAD_SERVERS[]={
"EFAServer/",
"Microsoft-IIS/4.",
"Microsoft-IIS/5.",
"Netscape-Enterprise/3.",
"Netscape-Enterprise/4.",
"Netscape-Enterprise/5.",
"Netscape-Enterprise/6.",
"WebLogic 3.",
"WebLogic 4.",
"WebLogic 5.",
"WebLogic 6.",
"Winstone Servlet Engine v0."
};
# define NBAD_SERVERS ((int)(sizeof(BAD_SERVERS)/sizeof(*BAD_SERVERS)))
if(*_server>='E'&&*_server<='W'){
int si;
for(si=0;si<NBAD_SERVERS;si++){
if(strncmp(_server,BAD_SERVERS[si],strlen(BAD_SERVERS[si]))==0){
return 0;
}
}
}
return 1;
# undef NBAD_SERVERS
}
static int op_http_stream_open(OpusHTTPStream *_stream,const char *_url,
int _skip_certificate_check,const char *_proxy_host,unsigned _proxy_port,
const char *_proxy_user,const char *_proxy_pass){
struct addrinfo *addrs;
const char *last_host;
unsigned last_port;
int nredirs;
int ret;
if(_proxy_host!=NULL&&OP_UNLIKELY(_proxy_port>65535U))return OP_EINVAL;
last_host=NULL;
/*We shouldn't have to initialize last_port, but gcc is too dumb to figure
out that last_host!=NULL implies we've already taken one trip through the
loop.*/
last_port=0;
ret=op_parse_url(&_stream->url,_url);
if(OP_UNLIKELY(ret<0))return ret;
for(nredirs=0;nredirs<OP_REDIRECT_LIMIT;nredirs++){
struct timeb start_time;
struct timeb end_time;
char *next;
char *status_code;
const char *host;
unsigned port;
int minor_version_pos;
int v1_1_compat;
if(_proxy_host==NULL){
host=_stream->url.host;
port=_stream->url.port;
}
else{
host=_proxy_host;
port=_proxy_port;
}
/*If connecting to the same place as last time, don't re-resolve it.*/
addrs=NULL;
if(last_host!=NULL){
if(strcmp(last_host,host)==0&&last_port==port)addrs=&_stream->addr_info;
else if(_stream->ssl_session!=NULL){
/*Forget any cached SSL session from the last host.*/
SSL_SESSION_free(_stream->ssl_session);
_stream->ssl_session=NULL;
}
if(last_host!=_proxy_host)_ogg_free((void *)last_host);
}
last_host=host;
last_port=port;
/*Initialize the SSL library if necessary.*/
if(OP_URL_IS_SSL(&_stream->url)&&_stream->ssl_ctx==NULL){
SSL_CTX *ssl_ctx;
# if !defined(OPENSSL_NO_LOCKING)
/*The documentation says SSL_library_init() is not reentrant.
We don't want to add our own depenencies on a threading library, and it
appears that it's safe to call OpenSSL's locking functions before the
library is initialized, so that's what we'll do (really OpenSSL should
do this for us).
This doesn't guarantee that _other_ threads in the application aren't
calling SSL_library_init() at the same time, but there's not much we
can do about that.*/
CRYPTO_w_lock(CRYPTO_LOCK_SSL);
# endif
SSL_library_init();
/*Needed to get SHA2 algorithms with old OpenSSL versions.*/
OpenSSL_add_ssl_algorithms();
# if !defined(OPENSSL_NO_LOCKING)
CRYPTO_w_unlock(CRYPTO_LOCK_SSL);
# endif
ssl_ctx=SSL_CTX_new(SSLv23_client_method());
if(ssl_ctx==NULL)return OP_EFAULT;
if(!_skip_certificate_check){
/*We don't do anything if this fails, since it just means we won't load
any certificates (and thus all checks will fail).
However, as that is probably the result of a system
mis-configuration, assert here to make it easier to identify.*/
OP_ALWAYS_TRUE(SSL_CTX_set_default_verify_paths(ssl_ctx));
SSL_CTX_set_verify(ssl_ctx,SSL_VERIFY_PEER,NULL);
}
_stream->ssl_ctx=ssl_ctx;
_stream->skip_certificate_check=_skip_certificate_check;
if(_proxy_host!=NULL){
/*We need to establish a CONNECT tunnel to handle https proxying.
Build the request we'll send to do so.*/
ret=op_sb_append(&_stream->proxy_connect,"CONNECT ",8);
ret|=op_sb_append_string(&_stream->proxy_connect,_stream->url.host);
ret|=op_sb_append_port(&_stream->proxy_connect,_stream->url.port);
/*CONNECT requires at least HTTP 1.1.*/
ret|=op_sb_append(&_stream->proxy_connect," HTTP/1.1\r\n",11);
ret|=op_sb_append(&_stream->proxy_connect,"Host: ",6);
ret|=op_sb_append_string(&_stream->proxy_connect,_stream->url.host);
/*The example in RFC 2817 Section 5.2 specifies an explicit port even
when connecting to the default port.
Given that the proxy doesn't know whether we're trying to connect to
an http or an https URL except by the port number, this seems like a
good idea.*/
ret|=op_sb_append_port(&_stream->proxy_connect,_stream->url.port);
ret|=op_sb_append(&_stream->proxy_connect,"\r\n",2);
ret|=op_sb_append(&_stream->proxy_connect,"User-Agent: .\r\n",15);
if(_proxy_user!=NULL&&_proxy_pass!=NULL){
ret|=op_sb_append_basic_auth_header(&_stream->proxy_connect,
"Proxy-Authorization",_proxy_user,_proxy_pass);
}
/*For backwards compatibility.*/
ret|=op_sb_append(&_stream->proxy_connect,
"Proxy-Connection: keep-alive\r\n",30);
ret|=op_sb_append(&_stream->proxy_connect,"\r\n",2);
if(OP_UNLIKELY(ret<0))return ret;
}
}
/*Actually make the connection.*/
if(addrs!=&_stream->addr_info){
addrs=op_resolve(host,port);
if(OP_UNLIKELY(addrs==NULL))return OP_FALSE;
}
ret=op_http_connect(_stream,_stream->conns+0,addrs,&start_time);
if(addrs!=&_stream->addr_info)freeaddrinfo(addrs);
if(OP_UNLIKELY(ret<0))return ret;
/*Build the request to send.*/
_stream->request.nbuf=0;
ret=op_sb_append(&_stream->request,"GET ",4);
ret|=op_sb_append_string(&_stream->request,
_proxy_host!=NULL?_url:_stream->url.path);
/*Send HTTP/1.0 by default for maximum compatibility (so we don't have to
re-try if HTTP/1.1 fails, though it shouldn't, even for a 1.0 server).
This means we aren't conditionally compliant with RFC 2145, because we
violate the requirement that "An HTTP client SHOULD send a request
version equal to the highest version for which the client is at least
conditionally compliant...".
According to RFC 2145, that means we can't claim any compliance with any
IETF HTTP specification.*/
ret|=op_sb_append(&_stream->request," HTTP/1.0\r\n",11);
/*Remember where this is so we can upgrade to HTTP/1.1 if the server
supports it.*/
minor_version_pos=_stream->request.nbuf-3;
ret|=op_sb_append(&_stream->request,"Host: ",6);
ret|=op_sb_append_string(&_stream->request,_stream->url.host);
if(!OP_URL_IS_DEFAULT_PORT(&_stream->url)){
ret|=op_sb_append_port(&_stream->request,_stream->url.port);
}
ret|=op_sb_append(&_stream->request,"\r\n",2);
/*User-Agents have been a bad idea, so send as little as possible.
RFC 2616 requires at least one token in the User-Agent, which must have
at least one character.*/
ret|=op_sb_append(&_stream->request,"User-Agent: .\r\n",15);
if(_proxy_host!=NULL&&!OP_URL_IS_SSL(&_stream->url)
&&_proxy_user!=NULL&&_proxy_pass!=NULL){
ret|=op_sb_append_basic_auth_header(&_stream->request,
"Proxy-Authorization",_proxy_user,_proxy_pass);
}
if(_stream->url.user!=NULL&&_stream->url.pass!=NULL){
ret|=op_sb_append_basic_auth_header(&_stream->request,
"Authorization",_stream->url.user,_stream->url.pass);
}
/*Always send a Referer [sic] header.
It's common to refuse to serve a resource unless one is present.
We just use the relative "/" URI to suggest we came from the same domain,
as this is the most common check.
This might violate RFC 2616's mandate that the field "MUST NOT be sent if
the Request-URI was obtained from a source that does not have its own
URI, such as input from the user keyboard," but we don't really have any
way to know.*/
/*TODO: Should we update this on redirects?*/
ret|=op_sb_append(&_stream->request,"Referer: /\r\n",12);
/*Always send a Range request header to find out if we're seekable.
This requires an HTTP/1.1 server to succeed, but we'll still get what we
want with an HTTP/1.0 server that ignores this request header.*/
ret|=op_sb_append(&_stream->request,"Range: bytes=0-\r\n",17);
/*Remember where this is so we can append offsets to it later.*/
_stream->request_tail=_stream->request.nbuf-4;
ret|=op_sb_append(&_stream->request,"\r\n",2);
if(OP_UNLIKELY(ret<0))return ret;
ret=op_http_conn_write_fully(_stream->conns+0,
_stream->request.buf,_stream->request.nbuf);
if(OP_UNLIKELY(ret<0))return ret;
ret=op_http_conn_read_response(_stream->conns+0,&_stream->response);
if(OP_UNLIKELY(ret<0))return ret;
OP_ALWAYS_TRUE(!ftime(&end_time));
next=op_http_parse_status_line(&v1_1_compat,&status_code,
_stream->response.buf);
if(OP_UNLIKELY(next==NULL))return OP_FALSE;
if(status_code[0]=='2'){
opus_int64 content_length;
opus_int64 range_length;
int pipeline;
/*We only understand 20x codes.*/
if(status_code[1]!='0')return OP_FALSE;
content_length=-1;
range_length=-1;
/*Pipelining is disabled by default.*/
pipeline=0;
for(;;){
char *header;
char *cdr;
ret=op_http_get_next_header(&header,&cdr,&next);
if(OP_UNLIKELY(ret<0))return ret;
if(header==NULL)break;
if(strcmp(header,"content-length")==0){
/*Two Content-Length headers?*/
if(OP_UNLIKELY(content_length>=0))return OP_FALSE;
content_length=op_http_parse_content_length(cdr);
if(OP_UNLIKELY(content_length<0))return (int)content_length;
/*Make sure the Content-Length and Content-Range headers match.*/
if(range_length>=0&&OP_UNLIKELY(content_length!=range_length)){
return OP_FALSE;
}
}
else if(strcmp(header,"content-range")==0){
opus_int64 range_first;
opus_int64 range_last;
/*Two Content-Range headers?*/
if(OP_UNLIKELY(range_length>=0))return OP_FALSE;
ret=op_http_parse_content_range(&range_first,&range_last,
&range_length,cdr);
if(OP_UNLIKELY(ret<0))return ret;
/*"A response with satus code 206 (Partial Content) MUST NOTE
include a Content-Range field with a byte-range-resp-spec of
'*'."*/
if(status_code[2]=='6'
&&(OP_UNLIKELY(range_first<0)||OP_UNLIKELY(range_last<0))){
return OP_FALSE;
}
/*We asked for the entire resource.*/
if(range_length>=0){
/*Quit if we didn't get it.*/
if(range_last>=0&&OP_UNLIKELY(range_last!=range_length-1)){
return OP_FALSE;
}
}
/*If there was no length, use the end of the range.*/
else if(range_last>=0)range_length=range_last+1;
/*Make sure the Content-Length and Content-Range headers match.*/
if(content_length>=0&&OP_UNLIKELY(content_length!=range_length)){
return OP_FALSE;
}
}
else if(strcmp(header,"connection")==0){
/*According to RFC 2616, if an HTTP/1.1 application does not support
pipelining, it "MUST include the 'close' connection option in
every message."
Therefore, if we receive one in the initial response, disable
pipelining entirely.
The server still might support it (e.g., we might just have hit the
request limit for a temporary child process), but if it doesn't
and we assume it does, every time we cross a chunk boundary we'll
error out and reconnect, adding lots of latency.*/
ret=op_http_parse_connection(cdr);
if(OP_UNLIKELY(ret<0))return ret;
pipeline-=ret;
}
else if(strcmp(header,"server")){
/*If we got a Server response header, and it wasn't from a known-bad
server, enable pipelining, as long as it's at least HTTP/1.1.
According to RFC 2145, the server is supposed to respond with the
highest minor version number it supports unless it is known or
suspected that we incorrectly implement the HTTP specification.
So it should send back at least HTTP/1.1, despite our HTTP/1.0
request.*/
pipeline+=v1_1_compat&&op_http_allow_pipelining(cdr);
}
}
switch(status_code[2]){
/*200 OK*/
case '0':break;
/*203 Non-Authoritative Information*/
case '3':break;
/*204 No Content*/
case '4':{
if(content_length>=0&&OP_UNLIKELY(content_length!=0)){
return OP_FALSE;
}
}break;
/*206 Partial Content*/
case '6':{
/*No Content-Range header.*/
if(OP_UNLIKELY(range_length<0))return OP_FALSE;
content_length=range_length;
/*The server supports range requests for this resource.
We can seek.*/
_stream->seekable=1;
}break;
/*201 Created: the response "SHOULD include an entity containing a list
of resource characteristics and location(s)," but not an Opus file.
202 Accepted: the response "SHOULD include an indication of request's
current status and either a pointer to a status monitor or some
estimate of when the user can expect the request to be fulfilled,"
but not an Opus file.
205 Reset Content: this "MUST NOT include an entity," meaning no Opus
file.
207...209 are not yet defined, so we don't know how to handle them.*/
default:return OP_FALSE;
}
_stream->content_length=content_length;
_stream->pipeline=pipeline>0;
/*Pipelining requires HTTP/1.1 persistent connections.*/
if(pipeline)_stream->request.buf[minor_version_pos]='1';
_stream->conns[0].pos=0;
_stream->conns[0].end_pos=_stream->seekable?content_length:-1;
_stream->conns[0].chunk_size=-1;
_stream->cur_conni=0;
_stream->connect_rate=op_time_diff_ms(&end_time,&start_time);
_stream->connect_rate=OP_MAX(_stream->connect_rate,1);
/*The URL has been successfully opened.*/
return 0;
}
/*Shouldn't get 1xx; 4xx and 5xx are both failures (and we don't retry).
Everything else is undefined.*/
else if(status_code[0]!='3')return OP_FALSE;
/*We have some form of redirect request.*/
/*We only understand 30x codes.*/
if(status_code[1]!='0')return OP_FALSE;
switch(status_code[2]){
/*300 Multiple Choices: "If the server has a preferred choice of
representation, it SHOULD include the specific URI for that
representation in the Location field," otherwise we'll fail.*/
case '0':
/*301 Moved Permanently*/
case '1':
/*302 Found*/
case '2':
/*307 Temporary Redirect*/
case '7':break;
/*305 Use Proxy: "The Location field gives the URI of the proxy."
TODO: This shouldn't actually be that hard to do.*/
case '5':return OP_EIMPL;
/*303 See Other: "The new URI is not a substitute reference for the
originally requested resource."
304 Not Modified: "The 304 response MUST NOT contain a message-body."
306 (Unused)
308...309 are not yet defined, so we don't know how to handle them.*/
default:return OP_FALSE;
}
_url=NULL;
for(;;){
char *header;
char *cdr;
ret=op_http_get_next_header(&header,&cdr,&next);
if(OP_UNLIKELY(ret<0))return ret;
if(header==NULL)break;
if(strcmp(header,"location")==0&&OP_LIKELY(_url==NULL))_url=cdr;
}
if(OP_UNLIKELY(_url==NULL))return OP_FALSE;
/*Don't free last_host if it came from the last URL.*/
if(last_host!=_proxy_host)_stream->url.host=NULL;
op_parsed_url_clear(&_stream->url);
ret=op_parse_url(&_stream->url,_url);
if(OP_UNLIKELY(ret<0)){
if(ret==OP_EINVAL)ret=OP_FALSE;
if(last_host!=_proxy_host)_ogg_free((void *)last_host);
return ret;
}
op_http_conn_close(_stream,_stream->conns+0,&_stream->lru_head,1);
}
/*Redirection limit reached.*/
return OP_FALSE;
}
static int op_http_conn_send_request(OpusHTTPStream *_stream,
OpusHTTPConn *_conn,opus_int64 _pos,opus_int32 _chunk_size,
int _try_not_to_block){
opus_int64 next_end;
int ret;
/*We shouldn't have another request outstanding.*/
OP_ASSERT(_conn->next_pos<0);
/*Build the request to send.*/
OP_ASSERT(_stream->request.nbuf>=_stream->request_tail);
_stream->request.nbuf=_stream->request_tail;
ret=op_sb_append_nonnegative_int64(&_stream->request,_pos);
ret|=op_sb_append(&_stream->request,"-",1);
if(_chunk_size>0&&OP_ADV_OFFSET(_pos,2*_chunk_size)<_stream->content_length){
/*We shouldn't be pipelining requests with non-HTTP/1.1 servers.*/
OP_ASSERT(_stream->pipeline);
next_end=_pos+_chunk_size;
ret|=op_sb_append_nonnegative_int64(&_stream->request,next_end-1);
/*Use a larger chunk size for our next request.*/
_chunk_size<<=1;
/*But after a while, just request the rest of the resource.*/
if(_chunk_size>OP_PIPELINE_CHUNK_SIZE_MAX)_chunk_size=-1;
}
else{
/*Either this was a non-pipelined request or we were close enough to the
end to just ask for the rest.*/
next_end=-1;
_chunk_size=-1;
}
ret|=op_sb_append(&_stream->request,"\r\n\r\n",4);
if(OP_UNLIKELY(ret<0))return ret;
/*If we don't want to block, check to see if there's enough space in the send
queue.
There's still a chance we might block, even if there is enough space, but
it's a much slimmer one.
Blocking at all is pretty unlikely, as we won't have any requests queued
when _try_not_to_block is set, so if FIONSPACE isn't available (e.g., on
Linux), just skip the test.*/
if(_try_not_to_block){
# if defined(FIONSPACE)
int available;
ret=ioctl(_conn->fd,FIONSPACE,&available);
if(ret<0||available<_stream->request.nbuf)return 1;
# endif
}
ret=op_http_conn_write_fully(_conn,
_stream->request.buf,_stream->request.nbuf);
if(OP_UNLIKELY(ret<0))return ret;
_conn->next_pos=_pos;
_conn->next_end=next_end;
/*Save the chunk size to use for the next request.*/
_conn->chunk_size=_chunk_size;
_conn->nrequests_left--;
return ret;
}
/*Handles the response to all requests after the first one.
Return: 1 if the connection was closed or timed out, 0 on success, or a
negative value on any other error.*/
static int op_http_conn_handle_response(OpusHTTPStream *_stream,
OpusHTTPConn *_conn){
char *next;
char *status_code;
opus_int64 range_length;
opus_int64 next_pos;
opus_int64 next_end;
int ret;
ret=op_http_conn_read_response(_conn,&_stream->response);
/*If the server just closed the connection on us, we may have just hit a
connection re-use limit, so we might want to retry.*/
if(OP_UNLIKELY(ret<0))return ret==OP_EREAD?1:ret;
next=op_http_parse_status_line(NULL,&status_code,_stream->response.buf);
if(OP_UNLIKELY(next==NULL))return OP_FALSE;
/*We _need_ a 206 Partial Content response.
Nothing else will do.*/
if(strncmp(status_code,"206",3)!=0){
/*But on a 408 Request Timeout, we might want to re-try.*/
return strncmp(status_code,"408",3)==0?1:OP_FALSE;
}
next_pos=_conn->next_pos;
next_end=_conn->next_end;
range_length=-1;
for(;;){
char *header;
char *cdr;
ret=op_http_get_next_header(&header,&cdr,&next);
if(OP_UNLIKELY(ret<0))return ret;
if(header==NULL)break;
if(strcmp(header,"content-range")==0){
opus_int64 range_first;
opus_int64 range_last;
/*Two Content-Range headers?*/
if(OP_UNLIKELY(range_length>=0))return OP_FALSE;
ret=op_http_parse_content_range(&range_first,&range_last,
&range_length,cdr);
if(OP_UNLIKELY(ret<0))return ret;
/*"A response with satus code 206 (Partial Content) MUST NOT
include a Content-Range field with a byte-range-resp-spec of
'*'."*/
if(OP_UNLIKELY(range_first<0)||OP_UNLIKELY(range_last<0))return OP_FALSE;
/*We also don't want range_last to overflow.*/
if(OP_UNLIKELY(range_last>=OP_INT64_MAX))return OP_FALSE;
range_last++;
/*Quit if we didn't get the offset we asked for.*/
if(range_first!=next_pos)return OP_FALSE;
if(next_end<0){
/*We asked for the rest of the resource.*/
if(range_length>=0){
/*Quit if we didn't get it.*/
if(OP_UNLIKELY(range_last!=range_length))return OP_FALSE;
}
/*If there was no length, use the end of the range.*/
else range_length=range_last;
next_end=range_last;
}
else{
if(range_last!=next_end)return OP_FALSE;
/*If there was no length, use the larger of the content length or the
end of this chunk.*/
if(range_length<0){
range_length=OP_MAX(range_last,_stream->content_length);
}
}
}
else if(strcmp(header,"content-length")==0){
opus_int64 content_length;
/*Validate the Content-Length header, if present, against the request we
made.*/
content_length=op_http_parse_content_length(cdr);
if(OP_UNLIKELY(content_length<0))return (int)content_length;
if(next_end<0){
/*If we haven't seen the Content-Range header yet and we asked for the
rest of the resource, set next_end, so we can make sure they match
when we do find the Content-Range header.*/
if(OP_UNLIKELY(next_pos>OP_INT64_MAX-content_length))return OP_FALSE;
next_end=next_pos+content_length;
}
/*Otherwise, make sure they match now.*/
else if(OP_UNLIKELY(next_end-next_pos!=content_length))return OP_FALSE;
}
else if(strcmp(header,"connection")==0){
ret=op_http_parse_connection(cdr);
if(OP_UNLIKELY(ret<0))return ret;
/*If the server told us it was going to close the connection, don't make
any more requests.*/
if(OP_UNLIKELY(ret>0))_conn->nrequests_left=0;
}
}
/*No Content-Range header.*/
if(OP_UNLIKELY(range_length<0))return OP_FALSE;
/*Update the content_length if necessary.*/
_stream->content_length=range_length;
_conn->pos=next_pos;
_conn->end_pos=next_end;
_conn->next_pos=-1;
return 0;
}
/*Open a new connection that will start reading at byte offset _pos.
_pos: The byte offset to start readiny from.
_chunk_size: The number of bytes to ask for in the initial request, or -1 to
request the rest of the resource.
This may be more bytes than remain, in which case it will be
converted into a request for the rest.*/
static int op_http_conn_open_pos(OpusHTTPStream *_stream,
OpusHTTPConn *_conn,opus_int64 _pos,opus_int32 _chunk_size){
struct timeb start_time;
struct timeb end_time;
opus_int32 connect_rate;
opus_int32 connect_time;
int ret;
ret=op_http_connect(_stream,_conn,&_stream->addr_info,&start_time);
if(OP_UNLIKELY(ret<0))return ret;
ret=op_http_conn_send_request(_stream,_conn,_pos,_chunk_size,0);
if(OP_UNLIKELY(ret<0))return ret;
ret=op_http_conn_handle_response(_stream,_conn);
if(OP_UNLIKELY(ret!=0))return OP_FALSE;
OP_ALWAYS_TRUE(!ftime(&end_time));
_stream->cur_conni=_conn-_stream->conns;
OP_ASSERT(_stream->cur_conni>=0&&_stream->cur_conni<OP_NCONNS_MAX);
/*The connection has been successfully opened.
Update the connection time estimate.*/
connect_time=op_time_diff_ms(&end_time,&start_time);
connect_rate=_stream->connect_rate;
connect_rate+=OP_MAX(connect_time,1)-connect_rate+8>>4;
_stream->connect_rate=connect_rate;
return 0;
}
/*Read data from the current response body.
If we're pipelining and we get close to the end of this response, queue
another request.
If we've reached the end of this response body, parse the next response and
keep going.
[out] _buf: Returns the data read.
_buf_size: The size of the buffer.
Return: A positive number of bytes read on success.
0: The connection was closed.
OP_EREAD: There was a fatal read error.*/
static int op_http_conn_read_body(OpusHTTPStream *_stream,
OpusHTTPConn *_conn,unsigned char *_buf,int _buf_size){
opus_int64 pos;
opus_int64 end_pos;
opus_int64 next_pos;
opus_int64 content_length;
int nread;
int pipeline;
int ret;
/*Currently this function can only be called on the LRU head.
Otherwise, we'd need a _pnext pointer if we needed to close the connection,
and re-opening it would re-organize the lists.*/
OP_ASSERT(_stream->lru_head==_conn);
/*We should have filterd out empty reads by this point.*/
OP_ASSERT(_buf_size>0);
pos=_conn->pos;
end_pos=_conn->end_pos;
next_pos=_conn->next_pos;
pipeline=_stream->pipeline;
content_length=_stream->content_length;
if(end_pos>=0){
/*Have we reached the end of the current response body?*/
if(pos>=end_pos){
OP_ASSERT(content_length>=0);
/*If this was the end of the stream, we're done.
Also return early if a non-blocking read was requested (regardless of
whether we might be able to parse the next response without
blocking).*/
if(content_length<=end_pos)return 0;
/*Otherwise, start on the next response.*/
if(next_pos<0){
/*We haven't issued another request yet.*/
if(!pipeline||_conn->nrequests_left<=0){
/*There are two ways to get here: either the server told us it was
going to close the connection after the last request, or we
thought we were reading the whole resource, but it grew while we
were reading it.
The only way the latter could have happened is if content_length
changed while seeking.
Open a new request to read the rest.*/
OP_ASSERT(_stream->seekable);
/*Try to open a new connection to read another chunk.*/
op_http_conn_close(_stream,_conn,&_stream->lru_head,1);
/*If we're not pipelining, we should be requesting the rest.*/
OP_ASSERT(pipeline||_conn->chunk_size==-1);
ret=op_http_conn_open_pos(_stream,_conn,end_pos,_conn->chunk_size);
if(OP_UNLIKELY(ret<0))return OP_EREAD;
}
else{
/*Issue the request now (better late than never).*/
ret=op_http_conn_send_request(_stream,_conn,pos,_conn->chunk_size,0);
if(OP_UNLIKELY(ret<0))return OP_EREAD;
next_pos=_conn->next_pos;
OP_ASSERT(next_pos>=0);
}
}
if(next_pos>=0){
/*We shouldn't be trying to read past the current request body if we're
seeking somewhere else.*/
OP_ASSERT(next_pos==end_pos);
ret=op_http_conn_handle_response(_stream,_conn);
if(OP_UNLIKELY(ret<0))return OP_EREAD;
if(OP_UNLIKELY(ret>0)&&pipeline){
opus_int64 next_end;
next_end=_conn->next_end;
/*Our request timed out or the server closed the connection.
Try re-connecting.*/
op_http_conn_close(_stream,_conn,&_stream->lru_head,1);
/*Unless there's a bug, we should be able to convert
(next_pos,next_end) into valid (_pos,_chunk_size) parameters.*/
OP_ASSERT(next_end<0
||next_end-next_pos>=0&&next_end-next_pos<=0x7FFFFFFF);
ret=op_http_conn_open_pos(_stream,_conn,next_pos,
next_end<0?-1:(opus_int32)(next_end-next_pos));
if(OP_UNLIKELY(ret<0))return OP_EREAD;
}
else if(OP_UNLIKELY(ret!=0))return OP_EREAD;
}
pos=_conn->pos;
end_pos=_conn->end_pos;
content_length=_stream->content_length;
}
OP_ASSERT(end_pos>pos);
_buf_size=OP_MIN(_buf_size,end_pos-pos);
}
nread=op_http_conn_read(_conn,(char *)_buf,_buf_size,1);
if(OP_UNLIKELY(nread<0))return nread;
pos+=nread;
_conn->pos=pos;
OP_ASSERT(end_pos<0||content_length>=0);
/*TODO: If nrequests_left<=0, we can't make a new request, and there will be
a big pause after we hit the end of the chunk while we open a new
connection.
It would be nice to be able to start that process now, but we have no way
to do it in the background without blocking (even if we could start it, we
have no guarantee the application will return control to us in a
sufficiently timely manner to allow us to complete it, and this is
uncommon enough that it's not worth using threads just for this).*/
if(end_pos>=0&&end_pos<content_length&&next_pos<0
&&pipeline&&OP_LIKELY(_conn->nrequests_left>0)){
opus_int64 request_thresh;
opus_int32 chunk_size;
/*Are we getting close to the end of the current response body?
If so, we should request more data.*/
request_thresh=_stream->connect_rate*_conn->read_rate>>12;
/*But don't commit ourselves too quickly.*/
chunk_size=_conn->chunk_size;
if(chunk_size>=0)request_thresh=OP_MIN(chunk_size>>2,request_thresh);
if(end_pos-pos<request_thresh){
ret=op_http_conn_send_request(_stream,_conn,end_pos,_conn->chunk_size,1);
if(OP_UNLIKELY(ret<0))return OP_EREAD;
}
}
return nread;
}
static int op_http_stream_read(void *_stream,
unsigned char *_ptr,int _buf_size){
OpusHTTPStream *stream;
ptrdiff_t nread;
opus_int64 size;
opus_int64 pos;
int ci;
stream=(OpusHTTPStream *)_stream;
/*Check for an empty read.*/
if(_buf_size<=0)return 0;
ci=stream->cur_conni;
/*No current connection => EOF.*/
if(ci<0)return 0;
pos=stream->conns[ci].pos;
size=stream->content_length;
/*Check for EOF.*/
if(size>=0){
if(pos>=size)return 0;
/*Check for a short read.*/
if(_buf_size>size-pos)_buf_size=(int)(size-pos);
}
nread=op_http_conn_read_body(stream,stream->conns+ci,_ptr,_buf_size);
if(OP_UNLIKELY(nread<=0)){
/*We hit an error or EOF.
Either way, we're done with this connection.*/
op_http_conn_close(stream,stream->conns+ci,&stream->lru_head,1);
stream->cur_conni=-1;
stream->pos=pos;
}
return nread;
}
/*Discard data until we reach the _target position.
This destroys the contents of _stream->response.buf, as we need somewhere to
read this data, and that is a convenient place.
_just_read_ahead: Whether or not this is a plain fast-forward.
If 0, we need to issue a new request for a chunk at _target
and discard all the data from our current request(s).
Otherwise, we should be able to reach _target without
issuing any new requests.
_target: The stream position to which to read ahead.*/
static int op_http_conn_read_ahead(OpusHTTPStream *_stream,
OpusHTTPConn *_conn,int _just_read_ahead,opus_int64 _target){
opus_int64 pos;
opus_int64 end_pos;
opus_int64 next_pos;
opus_int64 next_end;
ptrdiff_t nread;
int ret;
pos=_conn->pos;
end_pos=_conn->end_pos;
next_pos=_conn->next_pos;
next_end=_conn->next_end;
if(!_just_read_ahead){
/*We need to issue a new pipelined request.
This is the only case where we allow more than one outstanding request
at a time, so we need to reset next_pos (we'll restore it below if we
did have an outstanding request).*/
OP_ASSERT(_stream->pipeline);
_conn->next_pos=-1;
ret=op_http_conn_send_request(_stream,_conn,_target,
OP_PIPELINE_CHUNK_SIZE,0);
if(OP_UNLIKELY(ret<0))return ret;
}
/*We can reach the target position by reading forward in the current chunk.*/
if(_just_read_ahead&&(end_pos<0||_target<end_pos))end_pos=_target;
else if(next_pos>=0){
opus_int64 next_next_pos;
opus_int64 next_next_end;
/*We already have a request outstanding.
Finish off the current chunk.*/
while(pos<end_pos){
nread=op_http_conn_read(_conn,_stream->response.buf,
(int)OP_MIN(end_pos-pos,_stream->response.cbuf),1);
/*We failed to read ahead.*/
if(nread<=0)return OP_FALSE;
pos+=nread;
}
OP_ASSERT(pos==end_pos);
if(_just_read_ahead){
next_next_pos=next_next_end=-1;
end_pos=_target;
}
else{
OP_ASSERT(_conn->next_pos==_target);
next_next_pos=_target;
next_next_end=_conn->next_end;
_conn->next_pos=next_pos;
_conn->next_end=next_end;
end_pos=next_end;
}
ret=op_http_conn_handle_response(_stream,_conn);
if(OP_UNLIKELY(ret!=0))return OP_FALSE;
_conn->next_pos=next_next_pos;
_conn->next_end=next_next_end;
}
while(pos<end_pos){
nread=op_http_conn_read(_conn,_stream->response.buf,
(int)OP_MIN(end_pos-pos,_stream->response.cbuf),1);
/*We failed to read ahead.*/
if(nread<=0)return OP_FALSE;
pos+=nread;
}
OP_ASSERT(pos==end_pos);
if(!_just_read_ahead){
ret=op_http_conn_handle_response(_stream,_conn);
if(OP_UNLIKELY(ret!=0))return OP_FALSE;
}
else _conn->pos=end_pos;
OP_ASSERT(_conn->pos==_target);
return 0;
}
static int op_http_stream_seek(void *_stream,opus_int64 _offset,int _whence){
struct timeb seek_time;
OpusHTTPStream *stream;
OpusHTTPConn *conn;
OpusHTTPConn **pnext;
OpusHTTPConn *close_conn;
OpusHTTPConn **close_pnext;
opus_int64 content_length;
opus_int64 pos;
int pipeline;
int ci;
int ret;
stream=(OpusHTTPStream *)_stream;
if(!stream->seekable)return -1;
content_length=stream->content_length;
/*If we're seekable, we should have gotten a Content-Length.*/
OP_ASSERT(content_length>=0);
ci=stream->cur_conni;
pos=ci<0?content_length:stream->conns[ci].pos;
switch(_whence){
case SEEK_SET:{
/*Check for overflow:*/
if(_offset<0)return -1;
pos=_offset;
}break;
case SEEK_CUR:{
/*Check for overflow:*/
if(_offset<-pos||_offset>OP_INT64_MAX-pos)return -1;
pos+=_offset;
}break;
case SEEK_END:{
/*Check for overflow:*/
if(_offset>content_length||_offset<content_length-OP_INT64_MAX)return -1;
pos=content_length-_offset;
}break;
default:return -1;
}
/*Mark when we deactivated the active connection.*/
if(ci>=0){
op_http_conn_read_rate_update(stream->conns+ci);
*&seek_time=*&stream->conns[ci].read_time;
}
else OP_ALWAYS_TRUE(!ftime(&seek_time));
/*If we seeked past the end of the stream, just disable the active
connection.*/
if(pos>=content_length){
stream->cur_conni=-1;
stream->pos=pos;
return 0;
}
/*First try to find a connection we can use without waiting.*/
pnext=&stream->lru_head;
conn=stream->lru_head;
while(conn!=NULL){
opus_int64 conn_pos;
opus_int64 end_pos;
int available;
/*If this connection has been dormant too long or has made too many
requests, close it.
This is to prevent us from hitting server limits/firewall timeouts.*/
if(op_time_diff_ms(&seek_time,&conn->read_time)>
OP_CONNECTION_IDLE_TIMEOUT_MS
||conn->nrequests_left<OP_PIPELINE_MIN_REQUESTS){
op_http_conn_close(stream,conn,pnext,1);
conn=*pnext;
continue;
}
available=op_http_conn_estimate_available(conn);
conn_pos=conn->pos;
end_pos=conn->end_pos;
if(conn->next_pos>=0){
OP_ASSERT(end_pos>=0);
OP_ASSERT(conn->next_pos==end_pos);
end_pos=conn->next_end;
}
OP_ASSERT(end_pos<0||conn_pos<=end_pos);
/*Can we quickly read ahead without issuing a new request or waiting for
any more data?
If we have an oustanding request, we'll over-estimate the amount of data
it has available (because we'll count the response headers, too), but
that probably doesn't matter.*/
if(conn_pos<=pos&&pos-conn_pos<=available&&(end_pos<0||pos<end_pos)){
/*Found a suitable connection to re-use.*/
ret=op_http_conn_read_ahead(stream,conn,1,pos);
if(OP_UNLIKELY(ret<0)){
/*The connection might have become stale, so close it and keep going.*/
op_http_conn_close(stream,conn,pnext,1);
conn=*pnext;
continue;
}
/*Sucessfully resurrected this connection.*/
*pnext=conn->next;
conn->next=stream->lru_head;
stream->lru_head=conn;
stream->cur_conni=conn-stream->conns;
return 0;
}
pnext=&conn->next;
conn=conn->next;
}
/*Chances are that didn't work, so now try to find one we can use by reading
ahead a reasonable amount and/or by issuing a new request.*/
close_pnext=NULL;
close_conn=NULL;
pnext=&stream->lru_head;
conn=stream->lru_head;
pipeline=stream->pipeline;
while(conn!=NULL){
opus_int64 conn_pos;
opus_int64 end_pos;
opus_int64 read_ahead_thresh;
int available;
int just_read_ahead;
/*Dividing by 2048 instead of 1000 scales this by nearly 1/2, biasing away
from connection re-use (and roughly compensating for the lag required to
reopen the TCP window of a connection that's been idle).
There's no overflow checking here, because it's vanishingly unlikely, and
all it would do is cause us to make poor decisions.*/
read_ahead_thresh=OP_MAX(OP_READAHEAD_THRESH_MIN,
stream->connect_rate*conn->read_rate>>11);
available=op_http_conn_estimate_available(conn);
conn_pos=conn->pos;
end_pos=conn->end_pos;
if(conn->next_pos>=0){
OP_ASSERT(end_pos>=0);
OP_ASSERT(conn->next_pos==end_pos);
end_pos=conn->next_end;
}
OP_ASSERT(end_pos<0||conn_pos<=end_pos);
/*Can we quickly read ahead without issuing a new request?*/
just_read_ahead=conn_pos<=pos&&pos-conn_pos-available<=read_ahead_thresh
&&(end_pos<0||pos<end_pos);
if(just_read_ahead||pipeline&&end_pos>=0
&&end_pos-conn_pos-available<=read_ahead_thresh){
/*Found a suitable connection to re-use.*/
ret=op_http_conn_read_ahead(stream,conn,just_read_ahead,pos);
if(OP_UNLIKELY(ret<0)){
/*The connection might have become stale, so close it and keep going.*/
op_http_conn_close(stream,conn,pnext,1);
conn=*pnext;
continue;
}
/*Sucessfully resurrected this connection.*/
*pnext=conn->next;
conn->next=stream->lru_head;
stream->lru_head=conn;
stream->cur_conni=conn-stream->conns;
return 0;
}
close_pnext=pnext;
close_conn=conn;
pnext=&conn->next;
conn=conn->next;
}
/*No suitable connections.
Open a new one.*/
if(stream->free_head==NULL){
/*All connections in use.
Expire one of them (we should have already picked which one when scanning
the list).*/
OP_ASSERT(close_conn!=NULL);
OP_ASSERT(close_pnext!=NULL);
op_http_conn_close(stream,close_conn,close_pnext,1);
}
OP_ASSERT(stream->free_head!=NULL);
conn=stream->free_head;
/*If we can pipeline, only request a chunk of data.
If we're seeking now, there's a good chance we will want to seek again
soon, and this avoids committing this connection to reading the rest of
the stream.
Particularly with SSL or proxies, issuing a new request on the same
connection can be substantially faster than opening a new one.
This also limits the amount of data the server will blast at us on this
connection if we later seek elsewhere and start reading from a different
connection.*/
ret=op_http_conn_open_pos(stream,conn,pos,
pipeline?OP_PIPELINE_CHUNK_SIZE:-1);
if(OP_UNLIKELY(ret<0)){
op_http_conn_close(stream,conn,&stream->lru_head,1);
return -1;
}
return 0;
}
static opus_int64 op_http_stream_tell(void *_stream){
OpusHTTPStream *stream;
int ci;
stream=(OpusHTTPStream *)_stream;
ci=stream->cur_conni;
return ci<0?stream->pos:stream->conns[ci].pos;
}
static int op_http_stream_close(void *_stream){
OpusHTTPStream *stream;
stream=(OpusHTTPStream *)_stream;
if(OP_LIKELY(stream!=NULL)){
op_http_stream_clear(stream);
_ogg_free(stream);
}
return 0;
}
static const OpusFileCallbacks OP_HTTP_CALLBACKS={
op_http_stream_read,
op_http_stream_seek,
op_http_stream_tell,
op_http_stream_close
};
#endif
/*The actual URL stream creation function.
This one isn't extensible like the application-level interface, but because
it isn't public, we're free to change it in the future.*/
static void *op_url_stream_create_impl(OpusFileCallbacks *_cb,const char *_url,
int _skip_certificate_check,const char *_proxy_host,unsigned _proxy_port,
const char *_proxy_user,const char *_proxy_pass){
const char *path;
/*Check to see if this is a valid file: URL.*/
path=op_parse_file_url(_url);
if(path!=NULL){
char *unescaped_path;
void *ret;
unescaped_path=op_string_dup(path);
if(OP_UNLIKELY(unescaped_path==NULL))return NULL;
ret=op_fopen(_cb,op_unescape_url_component(unescaped_path),"rb");
_ogg_free(unescaped_path);
return ret;
}
#if defined(OP_ENABLE_HTTP)
/*If not, try http/https.*/
else{
OpusHTTPStream *stream;
int ret;
stream=(OpusHTTPStream *)_ogg_malloc(sizeof(*stream));
if(OP_UNLIKELY(stream==NULL))return NULL;
op_http_stream_init(stream);
ret=op_http_stream_open(stream,_url,_skip_certificate_check,
_proxy_host,_proxy_port,_proxy_user,_proxy_pass);
if(OP_UNLIKELY(ret<0)){
op_http_stream_clear(stream);
_ogg_free(stream);
return NULL;
}
*_cb=*&OP_HTTP_CALLBACKS;
return stream;
}
#else
(void)_skip_certificate_check;
(void)_proxy_host;
(void)_proxy_port;
(void)_proxy_user;
(void)_proxy_pass;
return NULL;
#endif
}
void *op_url_stream_vcreate(OpusFileCallbacks *_cb,
const char *_url,va_list _ap){
int skip_certificate_check;
const char *proxy_host;
opus_int32 proxy_port;
const char *proxy_user;
const char *proxy_pass;
skip_certificate_check=0;
proxy_host=NULL;
proxy_port=8080;
proxy_user=NULL;
proxy_pass=NULL;
for(;;){
ptrdiff_t request;
request=va_arg(_ap,char *)-(char *)NULL;
/*If we hit NULL, we're done processing options.*/
if(!request)break;
switch(request){
case OP_SSL_SKIP_CERTIFICATE_CHECK_REQUEST:{
skip_certificate_check=!!va_arg(_ap,opus_int32);
}break;
case OP_HTTP_PROXY_HOST_REQUEST:{
proxy_host=va_arg(_ap,const char *);
}break;
case OP_HTTP_PROXY_PORT_REQUEST:{
proxy_port=va_arg(_ap,opus_int32);
if(proxy_port<0||proxy_port>(opus_int32)65535)return NULL;
}break;
case OP_HTTP_PROXY_USER_REQUEST:{
proxy_user=va_arg(_ap,const char *);
}break;
case OP_HTTP_PROXY_PASS_REQUEST:{
proxy_pass=va_arg(_ap,const char *);
}break;
/*Some unknown option.*/
default:return NULL;
}
}
return op_url_stream_create_impl(_cb,_url,skip_certificate_check,
proxy_host,proxy_port,proxy_user,proxy_pass);
}
void *op_url_stream_create(OpusFileCallbacks *_cb,
const char *_url,...){
va_list ap;
va_start(ap,_url);
return op_url_stream_vcreate(_cb,_url,ap);
}