can't peek at some http traffic

[Antoine Martin]

Fedora 31 server running trunk (v4).

Instead of detecting the http traffic, peek returns no data and we fail with:

io_in_cb(<socket.socket fd=4, family=AddressFamily.AF_INET, type=SocketKind.SOCK_STREAM, proto=0, laddr=('0.0.0.0', 10000)>, <flags G_IO_IN of type GLib.IOCondition>)
(..)
socket tcp socket: 192.168.0.10:10000 <- 192.168.0.106:47438 peek: got 0 bytes
may_wrap_socket: no data, not wrapping
(..)
New tcp connection received
 from '192.168.0.106:47438'
 on '0.0.0.0:10000'

That's despite waiting for 1 second polling the socket...

wireshark shows that the TCP packet does contain the HTTP GET request:

GET / HTTP/1.1
Host: 192.168.0.10:10000
Connection: keep-alive
Cache-Control: max-age=0
Upgrade-Insecure-Requests: 1
Save-Data: on
User-Agent: Mozilla/5.0 (Linux; Android 10; PH-1) AppleWebKit/537.36 (KHTML, like Gecko) Chrome/80.0.3987.87 Mobile Safari/537.36
Accept: text/html,application/xhtml+xml,application/xml;q=0.9,image/webp,image/apng,*/*;q=0.8,application/signed-exchange;v=b3;q=0.9
Accept-Encoding: gzip, deflate
Accept-Language: en-GB,en;q=0.9,fr-FR;q=0.8,fr;q=0.7,th-TH;q=0.6,th;q=0.5,en-US;q=0.4

This happens with android clients, but not with desktop browsers!?

[Antoine Martin]

Adding a recv call immediately after the empty peek shows that the data is there waiting to be read!?

[Antoine Martin]

This is a kernel / distro bug.
Every other distro works just fine. Tested: Ubuntu 18.04, Ubuntu 19.10, Debian Stretch, Debian Buster, etc..

The problem only occurs on Fedora 30 and 31. (Not tested Fedora 29 or earlier).

[Antoine Martin]

Things I've tried:

• disabling and forcing nodelay / cork
• increasing the peek timeout
• different router, different network interface

Wireshark shows:

• client -> server: SYN
• server -> client: SYN+ACK
• client -> server: "GET /"
• server -> client: RST

...
Why do we reset the connection!? We're actually doing a recv when that happens!?

Also happens with v3.

[Antoine Martin]

make all sockets peekable

[Antoine Martin]

The patch above uses the same workaround that is used with ssl sockets: emulate peek by actually reading from the socket (without PEEK) but cache the data and return it later. (also uncomment enable_peek() to use it)
Does not always help. No idea why.

[Antoine Martin]

wireshark capture of a failure

[Antoine Martin]

updated patch

[Antoine Martin]

r27660 logs more details of connection timeouts.

I really don't understand why this makes any difference, but r27662 fixes things.
(at least I haven't been able to reproduce the bug with the Android tablet or mswindows browsers)

[migration script]

this ticket has been moved to: ​https://github.com/Xpra-org/xpra/issues/2587