| Age | Commit message (Collapse) | Author |
|---|
|
Add command line option "-b <interface>" which binds the ports
(and addresses) in the following "-p" and "-s" options to the specified
network interface with SO_BINDTODEVICE.
Using SO_BINDTODEVICE means uhttpd will only accept incoming traffic from
the specified network interface.
Signed-off-by: Mikael Magnusson <mikma@users.sourceforge.net>
|
|
When we run into an error like a 404 Not Found the request body is not
read and will be parsed as part of the next request. The next Request
will then fail because there is unexpected data in it.
When we run into such a problem with a request body close return an
error and close the connection. This should be easier than trying to
recover the state.
We saw this problem when /ubus/ was not installed, but the browser tried
to access it. Then uhttpd returned a 404, but the next request done in
this connection also failed with a HTTP 400, bad request.
Fixes: FS#3378
Signed-off-by: Hauke Mehrtens <hauke@hauke-m.de>
|
|
Unregister ubus subscriber in notification remove callback.
Without this call, uhttpd crashes when client tries to subscribe to
the ubus object after the object was unregistred and registered again.
It is bacuse the reference to ubus subscriber is not freed but
the memory is cleared in the uh_request_done function.
Signed-off-by: Wojciech Jowsa <wojciech.jowsa@gmail.com>
|
|
The legacy ubus protocol must not include an empty object in the result
array if the invoked ubus procedure yielded no response.
This fixes compatibility with existing legacy ubus api clients that expect
this behaviour, LuCI's fs.js in particular.
Signed-off-by: Jo-Philipp Wich <jo@mein.io>
|
|
When an uhttpd dispatch_handler provides a data_done callback which is
synchroneously finishing the request through ops->request_done(), the
calling client_poll_post_data() procedure incorrectly resets the resulting
client state from CLIENT_STATE_INIT to CLIENT_STATE_DONE which causes the
next uh_client_read_cb() invocation to bail out since no callback is
available for the CLIENT_STATE_DONE state, effectively discarding the
just received inbound request and sending the persistent connection state
into a deadlock sitation where the http client waits for a response to
its just sent request and uhttpd for further data to read.
Fix this issue by only setting CLIENT_STATE_DONE if the data_done callback
has not modified the state in the meanwhile.
Signed-off-by: Jo-Philipp Wich <jo@mein.io>
|
|
After specified time of network inactivity uhttpd is meant to close
connection. It doesn't seem to work thought. After timeout client
doesn't receive any more data but connection it still opened.
This change fixes that.
Signed-off-by: Rafał Miłecki <rafal@milecki.pl>
|
|
Complex GET requests (e.g. those with custom headers) require browsers
to send preflight OPTIONS request with:
Access-Control-Request-Method: GET
It's important to reply to such requests with the header
Access-Control-Allow-Origin (and optionally others) to allow CORS
requests.
Adding GET to the Access-Control-Allow-Methods is cosmetical as
according to the Fetch standard:
> If request’s method is not in methods, request’s method is not a
> CORS-safelisted method, and request’s credentials mode is "include" or
> methods does not contain `*`, then return a network error.
It basically means that Access-Control-Allow-Methods value is ignored
for GET, HEAD and POST methods.
Signed-off-by: Rafał Miłecki <rafal@milecki.pl>
|
|
With this change ubus will allow users with access to the object pseudo
method ":subscribe" to subscribe for notifications.
1. Move uh_ubus_allowed() up in the code
2. Export "Authorization" parsing code to the uh_ubus_get_auth()
3. Check for ":subscribe" method access
Right now this depends on "Authorization" HTTP header which browsers
don't allow setting for the EventSource. An alternative method of
submitting session token remains to be implemented.
Signed-off-by: Rafał Miłecki <rafal@milecki.pl>
|
|
Initial uhttpd ubus API was fully based on JSON-RPC. That restricted it
from supporting ubus notifications that don't fit its model.
Notifications require protocol that allows server to send data without
being polled. There are two candidates for that:
1. Server-sent events
2. WebSocket
The later one is overcomplex for this simple task so ideally uhttps ubus
should support text-based server-sent events. It's not possible with
JSON-RPC without violating it. Specification requires server to reply
with Response object. Replying with text/event-stream is not allowed.
All above led to designing new API that:
1. Uses GET and POST requests
2. Makes use of RESTful URLs
3. Uses JSON-RPC in cleaner form and only for calling ubus methods
This new API allows:
1. Listing all ubus objects and their methods using GET <prefix>/list
2. Listing object methods using GET <prefix>/list/<path>
3. Listening to object notifications with GET <prefix>/subscribe/<path>
4. Calling ubus methods using POST <prefix>/call/<path>
JSON-RPC custom protocol was also simplified to:
1. Use "method" member for ubus object method name
It was possible thanks to using RESTful URLs. Previously "method"
had to be "list" or "call".
2. Reply with Error object on ubus method call error
This simplified "result" member format as it doesn't need to contain
ubus result code anymore.
This patch doesn't break or change the old API. The biggest downside of
the new API is no support for batch requests. It's cost of using RESTful
URLs. It should not matter much as uhttpd supports keep alive.
Example usages:
1. Getting all objects and their methods:
$ curl http://192.168.1.1/ubus/list
{
"dhcp": {
"ipv4leases": {
},
"ipv6leases": {
}
},
"log": {
"read": {
"lines": "number",
"stream": "boolean",
"oneshot": "boolean"
},
"write": {
"event": "string"
}
}
}
2. Getting object methods:
$ curl http://192.168.1.1/ubus/list/log
{
"read": {
"lines": "number",
"stream": "boolean",
"oneshot": "boolean"
},
"write": {
"event": "string"
}
}
3. Subscribing to notifications:
$ curl http://192.168.1.1/ubus/subscribe/foo
event: status
data: {"count":5}
4. Calling ubus object method:
$ curl -d '{
"jsonrpc": "2.0",
"id": 1,
"method": "login",
"params": {"username": "root", "password": "password" }
}' http://192.168.1.1/ubus/call/session
{
"jsonrpc": "2.0",
"id": 1,
"result": {
"ubus_rpc_session": "01234567890123456789012345678901",
(...)
}
}
$ curl -H 'Authorization: Bearer 01234567890123456789012345678901' -d '{
"jsonrpc": "2.0",
"id": 1,
"method": "write",
"params": {"event": "Hello world" }
}' http://192.168.1.1/ubus/call/log
{
"jsonrpc": "2.0",
"id": 1,
"result": null
}
Signed-off-by: Rafał Miłecki <rafal@milecki.pl>
|
|
Initializing buffer in the uh_ubus_handle_request() didn't handle
batched requests correctly. It resulted in reusing buffer and generating
malformed replies.
Call blob_buf_init() before every usage of the global buf variable.
While at it make two functions take blob_buf pointer as argument:
1. uh_ubus_send_response()
2. uh_ubus_init_json_rpc_response()
This helps following global "buf" variable usage and will help avoiding
similar bugs in the future.
Fixes: 628341fae412 ("ubus: use local "blob_buf" in uh_ubus_handle_request_object()")
Signed-off-by: Rafał Miłecki <rafal@milecki.pl>
|
|
Use "_json_rpc_" in their names so it's clear they are related to the
JSON-RPC format. This cleans up code a bit and will allow adding more
formats in the future.
Signed-off-by: Rafał Miłecki <rafal@milecki.pl>
|
|
This follows two other functions logic: uh_ubus_send_request() and
uh_ubus_allowed(). Thanks to this change global "buf" variable is used
only for replies and doesn't require state tracking & reinitialization.
Signed-off-by: Rafał Miłecki <rafal@milecki.pl>
|
|
According to the JSON-RPC 2.0 specification "params" value can be either
an Array or Object. This change makes parse_json_rpc() accept both.
Type validation should be handled by a function that actually reads
"params" depending on expected format. This doesn't change existing
behaviour but allows adding more methods (that expect Object) in the
future.
Signed-off-by: Rafał Miłecki <rafal@milecki.pl>
|
|
Both: uh_ubus_send_request() and uh_ubus_send_list() don't use it.
Signed-off-by: Rafał Miłecki <rafal@milecki.pl>
|
|
There is no point in parsing "call" specific params for other ("list")
method calls. This is a minor cleanup that doesn't change uhttpd ubus
behaviour.
Signed-off-by: Rafał Miłecki <rafal@milecki.pl>
Acked-by: Jo-Philipp Wich <jo@mein.io>
|
|
Before this change if the cgi script hangs after writing headers
then the process will never be killed. Let's only cancel the timeout
if the process ends.
Signed-off-by: Santiago Piccinini <spiccinini@altermundi.net>
Signed-off-by: Daniel Golle <daniel@makrotopia.org>
|
|
Allow POST requests via persistent connections to improve performance
especially when using HTTPS on older devices.
After this change, average page load times in LuCI improve significantly
once the TLS connections are initiated.
When testing an ar71xx 19.07.2 build on an ethernet connected TL-WR1043nd
using luci-ssl-openssl and the ustream-openssl backend, the average page
load time for the main status page decreased to 1.3s compared to 4.7s
before, the interface and wireless configuration pages loaded in 1.2s
seconds each compared to the 4.2s and 4.9s respectively before.
Signed-off-by: Jo-Philipp Wich <jo@mein.io>
|
|
Introduce a new `-P` option which allows specifying a colon separated list
of accepted TLS ciphers.
Depending on the underlying ustream-ssl provider, the list either follows
OpenSSL's cipher string format or, in case of mbedTLS, is a simple colon
separated cipher whitelist.
Signed-off-by: Jo-Philipp Wich <jo@mein.io>
|
|
When we're starting a deferred request, the related input ustream might
have gone into read_blocked mode because incoming client request data
exhausted the ustreams internal buffer space. When this happens, edge
triggered uloop read events are "lost" and never re-triggered causing
the script input to never complete.
In order to avoid that deadlock situation, manually poke the input
ustream using ustream_poll() after invoking client_poll_post_data()
which should have drained (some) of the buffered input ustream contents.
Signed-off-by: Jo-Philipp Wich <jo@mein.io>
|
|
An invalid data access can be triggered with an HTTP POST request to a CGI
script specifying both `Transfer-Encoding: chunked` and a large negative
`Content-Length`.
The negative content length is assigned to `r->content_length` in
`client_parse_header` and passed as a negative read length to
`ustream_consume` in `client_poll_post_data` which will set the internal
ustream buffer pointer to an invalid address, causing out of bounds memory
reads later on in the code flow.
A similar implicit unsigned to signed conversion happens when parsing
chunk sizes emitted by a CGI program.
Address these issues by rejecting negative values in `r->content_length`
after assigning the `strtoul()` result.
Reported-by: Jan-Niklas Sohn <jan-niklas.sohn@gmx.de>
Signed-off-by: Jo-Philipp Wich <jo@mein.io>
|
|
Signed-off-by: Jo-Philipp Wich <jo@mein.io>
|
|
After format string checks were activated in libubox the compiler
started to complain about multiple missuses in uhttpd. This fixes the
format strings without changing the behavior.
blobmsg_get_string() just checks if the parameter is not NULL and then
calls blobmsg_data() and casts the result.
I think non of these problem is security relevant.
Signed-off-by: Hauke Mehrtens <hauke@hauke-m.de>
|
|
Escape the untrusted request URL input in the permission denied HTML output.
This fixes certain XSS vulnerabilities which can be leveraged to further
exploit the system.
Signed-off-by: Jo-Philipp Wich <jo@mein.io>
|
|
Adds ifdefs to fix building without TLS and Lua support
Signed-off-by: Paul Willoughby <paulw@spacemonkey.com>
|
|
It's one of the parameters used by default in LuCI, so it should be
included in the help output.
Signed-off-by: Karl Palsson <karlp@etactica.com>
|
|
Fixes: 77b774b ("build: avoid redefining _DEFAULT_SOURCE")
Signed-off-by: Jo-Philipp Wich <jo@mein.io>
|
|
Work around further glibc toolchain annoyances.
Signed-off-by: Jo-Philipp Wich <jo@mein.io>
|
|
Allow -l / -L arguments to be repeated to register multiple Lua prefix
handlers in the same process.
Signed-off-by: Jo-Philipp Wich <jo@mein.io>
|
|
Add _DEFAULT_SOURCE FTM in order to avoid warnings with recent glibc.
Signed-off-by: Jo-Philipp Wich <jo@mein.io>
|
|
Signed-off-by: Jo-Philipp Wich <jo@mein.io>
|
|
When the outer SSL ustream triggers a change notification due to
encountering EOF, the inner connection ustream might still have
pending data buffered.
Previously, such a condition led to truncated files delivered by
uhttpd via HTTPS and could be triggered by requesting large resources
via slow network links.
Mitigate the problem by propagating the EOF status indicator from
the outer ustream to the inner one and by deferring the client
connection shutdown until the inner ustream output buffer has been
completely drained.
Signed-off-by: Jo-Philipp Wich <jo@mein.io>
|
|
Map the "Origin:" header as $HTTP_ORIGIN environment variable for use by
request handling processes.
Signed-off-by: Jo-Philipp Wich <jo@mein.io>
|
|
Escape untrusted input like the request URL or filesystem paths in HTML
outputs such as the directory listing or 404 error messages.
This fixes certain XSS vulnerabilities which can be leveraged to further
exploit the system.
Signed-off-by: Jo-Philipp Wich <jo@mein.io>
|
|
The uh_htmlescape() function returns a copy of the given string with the
HTML special characters `<`, `>`, `"` and `'` replaced by HTML entities in
hexadecimal notation.
Signed-off-by: Jo-Philipp Wich <jo@mein.io>
|
|
This reverts commit ccd9717ba5d501b45fda957f0ea41c4660ef414c.
|
|
When a request handler accepting post data is too slow in consuming stdin,
uhttpd might deadlock with the master process stuck in a blocking write()
to the child and the child stuck with a blocking write() to the master.
Avoid this issue by putting the master side write end of the child pipe
into nonblocking mode right away and by raising the data_blocked flag
when attempts to write to the child yield EAGAIN.
Setting the flag ensures that client_poll_post_data() does not immediately
trigger a write attempt again, which effectively yields the master write
cycle so that the relay ustream has a chance to consume output of the
client process, thus solving the deadlock.
Signed-off-by: Jo-Philipp Wich <jo@mein.io>
|
|
The existing implementation incorrectly attempted to read the entire stdin
instead of fetching at most the given amount of bytes.
While we're at it, also make the size argument optional and let it default
to Luas internal buffer size.
Suggested-by: Bryan Mayland <bmayland+lede@capnbry.net>
Signed-off-by: Jo-Philipp Wich <jo@mein.io>
|
|
Instead of storing a pointer to the beginning of the query string within the
request url, store a copy in a static buffer instead. This aligns handling
the query string portion of the url with other elements like physical path
or path info information.
Since the URL is usually kept in the per-client blob buffer which might
change its memory location due to reallocations triggered by blobmsg_add_*,
it is not safe to point to it early in the request life cycle.
This fixes invalid memory access usually manifesting itself as corrupted
query string data in CGI scripts.
Reported-by: P. Wassi <p.wassi@gmx.at>
Suggested-by: Felix Fietkau <nbd@nbd.name>
Signed-off-by: Jo-Philipp Wich <jo@mein.io>
|
|
Add "text/cache-manifest" mimetype support to enable the possibility of
using Application Cache.
Signed-off-by: Adrian Panella <ianchi74@outlook.com>
|
|
Previous refactoring of the basic auth handling code broke the logic in
such a way that basic auth was only performed if a client sent an
Authorization header in its request, but it was never prompted for by
the server.
Signed-off-by: Jo-Philipp Wich <jo@mein.io>
|
|
Signed-off-by: Jo-Philipp Wich <jo@mein.io>
|
|
Mimic other web servers like Nginx or Apache and expose the parsed basic
auth information as HTTP_AUTH_USER and HTTP_AUTH_PASS environment variables
to CGI processes.
This also restores login-from-basic-auth functionality in LuCI.
Signed-off-by: Jo-Philipp Wich <jo@mein.io>
|
|
Store the parsed username and password information as HTTP headers in the
clients header blob buffer for later use by proc.c
Signed-off-by: Jo-Philipp Wich <jo@mein.io>
|
|
If a HTTP header variable has no corresponding value, then do not set it
to the empty string but to NULL, so that cgi.c will later skip it when
setting up the process environment.
Signed-off-by: Jo-Philipp Wich <jo@mein.io>
|
|
Provides a small speedup when resuming the connection.
Signed-off by: Rosen Penev <rosenp@gmail.com>
|
|
When calculating the matching prefix length, make sure to not take the trailing
slash into account in order to ensure that the resulting PATH_INFO string
always starts with a slash.
This ensures that an url like "/foo" against the matching prefix "/" or
"/foo/bar" against "/foo/" result in "/foo" and "/bar" respectively.
Signed-off-by: Jo-Philipp Wich <jo@mein.io>
|
|
The special prefix of "/" should match any url by definition but the final
assertion which ensures that the matched prefix ends in '\0' or '/' is causing
matches against the "/" prefix to fail.
Add some extra code to handle this special case to implemented the expected
behaviour.
Signed-off-by: Jo-Philipp Wich <jo@mein.io>
|
|
The check_cgi_path() function would segfault if we ever support running
uhttpd without any CGI prefix.
Add a check to prevent running uh_patch_match() when the prefix is unset.
Signed-off-by: Jo-Philipp Wich <jow@openwrt.org>
|
|
This allows the request handler to add extra headers to the response
even in the redirect case.
Signed-off-by: Felix Fietkau <nbd@nbd.name>
|
|
Add a CMake FIND_PATH and INCLUDE_DIRECTORIES searching for libubox/usock.h.
Some external toolchains which do not include standard locations would fail to
find the header otherwise.
Signed-off-by: Florian Fainelli <f.fainelli@gmail.com>
|
