Apache HTTP Сервер Версия 2.0
Known Problems in Clients
Warning:
This document has not been fully updated to take into account changes made in the 2.0 version of the Apache HTTP Server. Some of the information may still be relevant, but please use it with care.
Over time the Apache Group has discovered or been notified of problems with various clients which we have had to work around, or explain. This document describes these problems and the workarounds available. It's not arranged in any particular order. Some familiarity with the standards is assumed, but not necessary.
For brevity, Navigator will refer to Netscape's Navigator product (which in later versions was renamed "Communicator" and various other names), and MSIE will refer to Microsoft's Internet Explorer product. All trademarks and copyrights belong to their respective companies. We welcome input from the various client authors to correct inconsistencies in this paper, or to provide us with exact version numbers where things are broken/fixed.
For reference, RFC1945 defines HTTP/1.0, and RFC2068 defines HTTP/1.1. Apache as of version 1.2 is an HTTP/1.1 server (with an optional HTTP/1.0 proxy).
Various of these workarounds are triggered by environment
variables. The admin typically controls which are set, and for
which clients, by using mod_browser
. Unless
otherwise noted all of these workarounds exist in versions 1.2
and later.
- Trailing CRLF on POSTs
- Broken KeepAlive
- Incorrect interpretation of
HTTP/1.1
in response - Requests use HTTP/1.1 but responses must be in HTTP/1.0
- Boundary problems with header parsing
- Multipart responses and Quoted Boundary Strings
- Byterange Requests
-
Set-Cookie
header is unmergeable -
Expires
headers and GIF89A animations -
POST
withoutContent-Length
- JDK 1.2 betas lose parts of responses.
-
Content-Type
change is not noticed after reload - MSIE Cookie problem with expiry date in the year 2000
- Lynx incorrectly asking for transparent content negotiation
- MSIE 4.0 mishandles Vary response header
Trailing CRLF on POSTs
This is a legacy issue. The CERN webserver required
POST
data to have an extra CRLF
following it. Thus many clients send an extra CRLF
that is not included in the Content-Length
of the
request. Apache works around this problem by eating any empty
lines which appear before a request.
Broken KeepAlive
Various clients have had broken implementations of keepalive (persistent connections). In particular the Windows versions of Navigator 2.0 get very confused when the server times out an idle connection. The workaround is present in the default config files:
BrowserMatch Mozilla/2 nokeepalive
Note that this matches some earlier versions of MSIE, which began the practice of calling themselves Mozilla in their user-agent strings just like Navigator.
MSIE 4.0b2, which claims to support HTTP/1.1, does not
properly support keepalive when it is used on 301 or 302
(redirect) responses. Unfortunately Apache's
nokeepalive
code prior to 1.2.2 would not work
with HTTP/1.1 clients. You must apply
BrowserMatch "MSIE 4\.0b2;" nokeepalive
Incorrect interpretation of
HTTP/1.1
in response
To quote from section 3.1 of RFC1945:
Since Apache is an HTTP/1.1 server, it indicates so as part of its response. Many client authors mistakenly treat this part of the response as an indication of the protocol that the response is in, and then refuse to accept the response.
The first major indication of this problem was with AOL's
proxy servers. When Apache 1.2 went into beta it was the first
wide-spread HTTP/1.1 server. After some discussion, AOL fixed
their proxies. In anticipation of similar problems, the
force-response-1.0
environment variable was added
to Apache. When present Apache will indicate "HTTP/1.0" in
response to an HTTP/1.0 client, but will not in any other way
change the response.
The pre-1.1 Java Development Kit (JDK) that is used in many clients (including Navigator 3.x and MSIE 3.x) exhibits this problem. As do some of the early pre-releases of the 1.1 JDK. We think it is fixed in the 1.1 JDK release. In any event the workaround:
BrowserMatch Java/1.0 force-response-1.0
BrowserMatch JDK/1.0 force-response-1.0
RealPlayer 4.0 from Progressive Networks also exhibits this
problem. However they have fixed it in version 4.01 of the
player, but version 4.01 uses the same User-Agent
as version 4.0. The workaround is still:
BrowserMatch "RealPlayer 4.0" force-response-1.0
Requests use HTTP/1.1 but responses must be in HTTP/1.0
MSIE 4.0b2 has this problem. Its Java VM makes requests in HTTP/1.1 format but the responses must be in HTTP/1.0 format (in particular, it does not understand chunked responses). The workaround is to fool Apache into believing the request came in HTTP/1.0 format.
BrowserMatch "MSIE 4\.0b2;" downgrade-1.0
force-response-1.0
This workaround is available in 1.2.2, and in a
Boundary problems with header parsing
All versions of Navigator from 2.0 through 4.0b2 (and possibly later) have a problem if the trailing CRLF of the response header starts at offset 256, 257 or 258 of the response. A BrowserMatch for this would match on nearly every hit, so the workaround is enabled automatically on all responses. The workaround implemented detects when this condition would occur in a response and adds extra padding to the header to push the trailing CRLF past offset 258 of the response.
Multipart responses and Quoted Boundary Strings
On multipart responses some clients will not accept quotes (") around the boundary string. The MIME standard recommends that such quotes be used. But the clients were probably written based on one of the examples in RFC2068, which does not include quotes. Apache does not include quotes on its boundary strings to workaround this problem.
Byterange Requests
A byterange request is used when the client wishes to retrieve a portion of an object, not necessarily the entire object. There was a very old draft which included these byteranges in the URL. Old clients such as Navigator 2.0b1 and MSIE 3.0 for the MAC exhibit this behaviour, and it will appear in the servers' access logs as (failed) attempts to retrieve a URL with a trailing ";xxx-yyy". Apache does not attempt to implement this at all.
A subsequent draft of this standard defines a header
Request-Range
, and a response type
multipart/x-byteranges
. The HTTP/1.1 standard
includes this draft with a few fixes, and it defines the header
Range
and type
multipart/byteranges
.
Navigator (versions 2 and 3) sends both Range
and Request-Range
headers (with the same value),
but does not accept a multipart/byteranges
response. The response must be
multipart/x-byteranges
. As a workaround, if Apache
receives a Request-Range
header it considers it
"higher priority" than a Range
header and in
response uses multipart/x-byteranges
.
The Adobe Acrobat Reader plugin makes extensive use of
byteranges and prior to version 3.01 supports only the
multipart/x-byterange
response. Unfortunately
there is no clue that it is the plugin making the request. If
the plugin is used with Navigator, the above workaround works
fine. But if the plugin is used with MSIE 3 (on Windows) the
workaround won't work because MSIE 3 doesn't give the
Range-Request
clue that Navigator does. To
workaround this, Apache special cases "MSIE 3" in the
User-Agent
and serves
multipart/x-byteranges
. Note that the necessity
for this with MSIE 3 is actually due to the Acrobat plugin, not
due to the browser.
Netscape Communicator appears to not issue the non-standard
Request-Range
header. When an Acrobat plugin prior
to version 3.01 is used with it, it will not properly
understand byteranges. The user must upgrade their Acrobat
reader to 3.01.
Set-Cookie
header is
unmergeable
The HTTP specifications say that it is legal to merge
headers with duplicate names into one (separated by commas).
Some browsers that support Cookies don't like merged headers
and prefer that each Set-Cookie
header is sent
separately. When parsing the headers returned by a CGI, Apache
will explicitly avoid merging any Set-Cookie
headers.
Expires
headers
and GIF89A animations
Navigator versions 2 through 4 will erroneously re-request
GIF89A animations on each loop of the animation if the first
response included an Expires
header. This happens
regardless of how far in the future the expiry time is set.
There is no workaround supplied with Apache, however there are
hacks for
POST
without
Content-Length
In certain situations Navigator 3.01 through 3.03 appear to
incorrectly issue a POST without the request body. There is no
known workaround. It has been fixed in Navigator 3.04,
Netscapes provides some
JDK 1.2 betas lose parts of responses.
The http client in the JDK1.2beta2 and beta3 will throw away the first part of the response body when both the headers and the first part of the body are sent in the same network packet AND keep-alive's are being used. If either condition is not met then it works fine.
See also Bug-ID's 4124329 and 4125538 at the java developer connection.
If you are seeing this bug yourself, you can add the following BrowserMatch directive to work around it:
BrowserMatch "Java1\.2beta[23]" nokeepalive
We don't advocate this though since bending over backwards for beta software is usually not a good idea; ideally it gets fixed, new betas or a final release comes out, and no one uses the broken old software anymore. In theory.
Content-Type
change is not noticed after reload
Navigator (all versions?) will cache the
content-type
for an object "forever". Using reload
or shift-reload will not cause Navigator to notice a
content-type
change. The only work-around is for
the user to flush their caches (memory and disk). By way of an
example, some folks may be using an old mime.types
file which does not map .htm
to
text/html
, in this case Apache will default to
sending text/plain
. If the user requests the page
and it is served as text/plain
. After the admin
fixes the server, the user will have to flush their caches
before the object will be shown with the correct
text/html
type.
MSIE Cookie problem with expiry date in the year 2000
MSIE versions 3.00 and 3.02 (without the Y2K patch) do not handle cookie expiry dates in the year 2000 properly. Years after 2000 and before 2000 work fine. This is fixed in IE4.01 service pack 1, and in the Y2K patch for IE3.02. Users should avoid using expiry dates in the year 2000.
Lynx incorrectly asking for transparent content negotiation
The Lynx browser versions 2.7 and 2.8 send a "negotiate: trans" header in their requests, which is an indication the browser supports transparent content negotiation (TCN). However the browser does not support TCN. As of version 1.3.4, Apache supports TCN, and this causes problems with these versions of Lynx. As a workaround future versions of Apache will ignore this header when sent by the Lynx client.
MSIE 4.0 mishandles Vary response header
MSIE 4.0 does not handle a Vary header properly. The Vary
header is generated by mod_rewrite in apache 1.3. The result is
an error from MSIE saying it cannot download the requested
file. There are more details in
A workaround is to add the following to your server's configuration files:
BrowserMatch "MSIE 4\.0" force-no-vary
(This workaround is only available with releases after 1.3.6 of the Apache Web server.)