Instead of sending data always at the beginning of the request wait until the
server will respond with "HTTP/1.1 100 Continue".
Before this patch server do send "HTTP/1.1 100 Continue" but client always send
post data even without waiting server response.
P.S. this patch also touches some not 100% related tab-align issues.
Covered-by: http/data_length_constraints
Covered-by: http/read_on_write_error
Since otherwise we can have nasty bugs with part of previous *request* in
current *request* and hence some parsing errors.
And now we have failures:
http/non_lingering_close: [forking] [err] ../http.c:1326: Assertion !evbuffer_drain(tmp, -1) failed in ../http.c
Also since after this patch code became more generic, we now respond with
HTTP_ENTITYTOOLARGE even without "Expect: 100-Continue", which is correct by
RFC.
Refs: #321
v2: remove EVHTTP_CON_ABOUT_TO_CLOSE
By lingering close I mean something what nginx have for this name, by this term
I mean that we need to read all the body even if it's size greater then
`max_body_size`, otherwise browsers on win32 (including chrome) failed read the
http status - entity-too-large (while on linux chrome for instance are good),
and also this includes badly written http clients.
Refs: #321
v2: do this only under EVHTTP_SERVER_LINGERING_CLOSE
Otherwise if we will try to write more data than server can accept
(see `evhttp_set_max_body_size()` for libevent server) we will get `EPIPE` and
will not try to read server's response which must contain 400 error for now
(which is not strictly correct though, it must 413).
```
$ strace regress --no-fork http/data_length_constraints
...
connect(10, {sa_family=AF_INET, sin_port=htons(43988), sin_addr=inet_addr("127.0.0.1")}, 16) = -1 EINPROGRESS (Operation now in progress)
...
writev(10, [{"POST / HTTP/1.1\r\nHost: somehost\r"..., 60}, {"aaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaa"..., 16324}], 2) = 16384
epoll_wait(5, [{EPOLLOUT, {u32=10, u64=10}}, {EPOLLIN, {u32=11, u64=11}}], 32, 50000) = 2
writev(10, [{"aaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaa"..., 16384}], 1) = 16384
ioctl(11, FIONREAD, [32768]) = 0
readv(11, [{"POST / HTTP/1.1\r\nHost: somehost\r"..., 4096}], 1) = 4096
epoll_ctl(5, EPOLL_CTL_DEL, 11, 0x7fff09d41e50) = 0
epoll_ctl(5, EPOLL_CTL_ADD, 11, {EPOLLOUT, {u32=11, u64=11}}) = 0
epoll_wait(5, [{EPOLLOUT, {u32=10, u64=10}}, {EPOLLOUT, {u32=11, u64=11}}], 32, 50000) = 2
writev(10, [{"aaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaa"..., 16384}], 1) = 16384
writev(11, [{"HTTP/1.1 400 Bad Request\r\nConten"..., 129}, {"<HTML><HEAD>\n<TITLE>400 Bad Requ"..., 94}], 2) = 223
epoll_ctl(5, EPOLL_CTL_DEL, 11, 0x7fff09d42080) = 0
shutdown(11, SHUT_WR) = 0
close(11) = 0
epoll_wait(5, [{EPOLLOUT|EPOLLERR|EPOLLHUP, {u32=10, u64=10}}], 32, 50000) = 1
writev(10, [{"aaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaa"..., 16384}], 1) = -1 EPIPE (Broken pipe)
--- SIGPIPE {si_signo=SIGPIPE, si_code=SI_USER, si_pid=13954, si_uid=1000} ---
epoll_ctl(5, EPOLL_CTL_DEL, 10, 0x7fff09d42010) = 0
shutdown(10, SHUT_WR) = -1 ENOTCONN (Transport endpoint is not connected)
close(10) = 0
write(1, "\n FAIL ../test/regress_http.c:3"..., 37
```
Careful reader can ask why it send error even when it didn't read
`evcon->max_body_size`, and the answer will be checks for `evcon->max_body_size
against `Content-Length` header, which contains ~8MB (-2 bytes).
And also if we will not drain the output buffer than we will send buffer that
we didn't send in previous request and instead of sending method via
`evhttp_make_header()`.
Fixes: http/data_length_constraints
Refs: #321
v2: do this only under EVHTTP_CON_READ_ON_WRITE_ERROR flag
And we can't make them continuous, since the latest is a public API, and
otherwise we will break binary compatibility.
Also extra check for EVHTTP_CON_PUBLIC_FLAGS_END, in case somebody forgot about
this (implementer I mean).
Refs: #182
Though CloudABI implements a very large part of POSIX, it does not
provide these header files, for the reason that there is no raw device
access, no resource limiting and no access to the global process table
through wait().
It looks like these header files are not actually needed in theory.
There don't seem to be any constructs in these source files that use
these features, but I suspect they might still be required on some
systems.
Before this patch http server don't knows when client disconnected until it
will try to write to it, IOW to detect is client still alive you need to write
something to client socket, however it is not convenient since it requires to
store all clients somewhere and poll them periodically, and I don't see any
regressions if we will leave EV_READ always (like libevhtp do), since we
already reset read callback in evhttp_write_buffer() (see
http/write_during_read).
Also since we don't disable EV_READ anymore we don't need some enable EV_READ,
so we will reduce number of epoll_ctl() calls.
Covered-by: http/terminate_chunked_oneshot
Covered-by: http/write_during_read
Fixes: #78
Before this patch every time we are retrying our request we resolve
domain, but we could optimize this (since UDP is slow) by using cached
conn_address value, so do this.
In http the only case when when we could store it is when we already
connected, *but* if we are doing request using domain name, then we need
to do request to nameserver to get IP address, and this is handled by
bufferevent.
So when we have IP address (from nameserver) and don't have connection
to this IP address, we could already cache it to avoid extra DNS
requests (since UDP is slow), and we can't do this from http layer, only
from bufferevent.
This will fix some invalid read/write:
==556== Invalid read of size 8
==556== at 0x4E4EEC6: event_queue_remove_timeout (minheap-internal.h:178)
==556== by 0x4E508AA: event_del_nolock_ (event.c:2764)
==556== by 0x4E53535: event_base_loop (event.c:3088)
==556== by 0x406FCFA: dispatch (libcrawl.c:271)
==556== by 0x402863: main (crawler.c:49)
==556== Address 0x68a3f18 is 152 bytes inside a block of size 400 free'd
==556== at 0x4C29C97: free (in /usr/local/lib/valgrind/vgpreload_memcheck-amd64-linux.so)
==556== by 0x406F140: renew (libcrawl.c:625)
==556== by 0x4E6CDE9: evhttp_connection_cb_cleanup (http.c:1331)
==556== by 0x4E6E2B2: evhttp_connection_cb (http.c:1424)
==556== by 0x4E4DF2D: bufferevent_writecb (bufferevent_sock.c:310)
==556== by 0x4E52D1D: event_process_active_single_queue (event.c:1584)
==556== by 0x4E53676: event_base_loop (event.c:1676)
==556== by 0x406FCFA: dispatch (libcrawl.c:271)
==556== by 0x402863: main (crawler.c:49)
But this one because of some invalid write before (I guess).
It is 100% reproduced during massive crawling (because this process
has many different servers), but after spending some time for trying to
reproduce this using some simple tests/utils I gave up for a few days (I
have a lot of work to do), but I'm sending this patch as a reminder.
Just in case, I've tried next tests:
- mixing timeouts/retries
- shutdown http server and return it back
- slow dns server for first request
- sleep before accept
- hacking libevent sources to change the behaviour of http layer (so it
will go into that function which I'm insterested in).
This patch provides the ability to receive a callback on the completion of a
request. The callback takes place immediately before the request's resources
are released.
evhttp_write_buffer() used by evhttp_send_reply_chunk() can take callback
executed when (part of) the buffer has been written. Using this callback to
schedule the next chunk avoids buffering large amounts of data in memory.
Basically tcp final handshake looks like this:
(C - client, S - server)
ACK[C] - FIN/ACK[S] - FIN/ACK[S] - ACK [C]
However there are servers, that didn't close connection like this,
while it is still _considered_ as valid, and using libevent http layer
we can do requests to such servers.
Modified handshake:
(C - client, S - server)
ACK[C] - RST/ACK[S] - RST/ACK[S]
And in this case we can't extract IP address from socket, because it is
already closed, and getpeername() will return: "transport endpoint is not connected".
So we need to store address that we are connecting to, after we know it,
and that is what this patch do.
I have reproduced it, however it have some extra packages.
(I will try to fix it)
https://github.com/azat/nfq-examples/blob/master/nfqnl_rst_fin.c
The evhttp_send_reply method invokes evhttp_write_buffer with a
callback that may release the underlying request object and
bufferevent upon completion. This cleanup callback is invoked by the
underlying bufferevent's write callback. Improperly enabling write
events before referencing the bufferevent could lead to use after free
and memory corruption.
It is useful to know why you callback called with NULL (i.e. it failed),
for example if you set max_body with evhttp_connection_set_max_body_size()
you must know that it failed because of body was longer than this size.
(Commit message tweaked by Nick)
This patch add check in evhttp_decode_uri_internal() that next 2 symbols
are exists in array of chars for decoding, if don't have two next 2
symbols don't try to decode '%FF'
Before this patch socket created before domain was resolved, and it
always create with AF_INET (ipv4), but we must create socket only after
domain was resolved to understad which protocol family have domain
address.
Thank to Patrick Pelletier, who found this bug.
