question about BPF sk_skb_stream_verdict redirect and poll/epoll events

question about BPF sk_skb_stream_verdict redirect and poll/epoll events

[Farbod Shahinfar]

Hello,

I am doing a simple experiment in which I send a message to a TCP server 
and the server echoes the message. I am attaching a BPF sk_skb 
stream_verdict program to the server socket to redirect the message back 
to the client (redirects the SKB on the same socket but to the TX 
queue). In my test, I noticed that the user-space server, which is using 
the poll system call, is woken up, and when it reads the socket, 
receives zero as the number of bytes read.

I first want to ask if this is the intended behavior.
In case this is the intended behavior, my second question is, what 
should I do to prevent the user program from waking up? I hope by not 
waking up the user program I could better use the available resources.

Sincerely,
Farbod

Re: question about BPF sk_skb_stream_verdict redirect and poll/epoll events

[Ferenc Fejes]

Hi!

On Fri, 2023-09-22 at 19:32 +0200, Farbod Shahinfar wrote:
> Hello,
> 
> I am doing a simple experiment in which I send a message to a TCP
> server 
> and the server echoes the message. I am attaching a BPF sk_skb 
> stream_verdict program to the server socket to redirect the message
> back 
> to the client (redirects the SKB on the same socket but to the TX 
> queue). In my test, I noticed that the user-space server, which is
> using 
> the poll system call, is woken up, and when it reads the socket, 
> receives zero as the number of bytes read.

Do you poll for POLLIN events?

> 
> I first want to ask if this is the intended behavior.
> In case this is the intended behavior, my second question is, what 
> should I do to prevent the user program from waking up? I hope by not
> waking up the user program I could better use the available
> resources.

AFAIK sockets in sockmap are propagates every events to poll (fixme),
but with pollfd::events you can specify the ones makes sense to you
e.g.: POLLERR, POLLHUP, POLLRDHUP, etc.
> 
> Sincerely,
> Farbod

Ferenc

Re: question about BPF sk_skb_stream_verdict redirect and poll/epoll events

[John Fastabend]

Ferenc Fejes wrote:
> Hi!
> 
> On Fri, 2023-09-22 at 19:32 +0200, Farbod Shahinfar wrote:
> > Hello,
> > 
> > I am doing a simple experiment in which I send a message to a TCP
> > server 
> > and the server echoes the message. I am attaching a BPF sk_skb 
> > stream_verdict program to the server socket to redirect the message
> > back 
> > to the client (redirects the SKB on the same socket but to the TX 
> > queue). In my test, I noticed that the user-space server, which is
> > using 
> > the poll system call, is woken up, and when it reads the socket, 
> > receives zero as the number of bytes read.
> 
> Do you poll for POLLIN events?
> 
> > 
> > I first want to ask if this is the intended behavior.
> > In case this is the intended behavior, my second question is, what 
> > should I do to prevent the user program from waking up? I hope by not
> > waking up the user program I could better use the available
> > resources.
> 
> AFAIK sockets in sockmap are propagates every events to poll (fixme),
> but with pollfd::events you can specify the ones makes sense to you
> e.g.: POLLERR, POLLHUP, POLLRDHUP, etc.

Agree we should not wake up the application in this case. I currently
fixing the wake up for cases where SK_PASS is passed. I can add this
to the todo list if no one beats me to it. For reference with parser
program we have a series of similar issue I'll fix next.

> > 
> > Sincerely,
> > Farbod
> 
> Ferenc
>