(e)poll and (rd)hup questions

(e)poll and (rd)hup questions

[Michael Noisternig]

Hello folks,

I hope this is a right place to ask this question. (Sorry if it isn't, 
please hint me where to post in that case.)

This has been bugging me for quite some while now. I'm using epoll in 
default level-triggered mode. I'm polling for input, output, rdhup, and 
forcibly for hup. I figured that when I get EPOLLRDHUP that doesn't mean 
there is no more data to read, presumably I must empty the kernel buffer 
by continously reading from the fd until I get zero as a result 
indicating rdhup. This means EPOLLRDHUP is delivered asynchronously and 
thus is pretty useless to me.

On the other hand, EPOLLHUP seems to be delivered synchronously. If it 
isn't then this means I get constantly notified about a closed fd until 
I have read all data from the kernel buffer and remove the fd from the 
epoll set (which makes being forced to hup notifications useless to me).

(And no, don't tell me to use edge-triggered mode, I have reasons to use 
level-triggered.)

Question: Is my observation correct that EPOLLRDHUP is delivered 
asynchronously in contrast to EPOLLHUP? Or is EPOLLHUP delivered 
asynchronously as well?

Thanks, anwers would help a lot!!!
Michael

Re: (e)poll and (rd)hup questions

[Davide Libenzi]

On Tue, 26 Aug 2008, Michael Noisternig wrote:

> Hello folks,
> 
> I hope this is a right place to ask this question. (Sorry if it isn't, please
> hint me where to post in that case.)
> 
> This has been bugging me for quite some while now. I'm using epoll in default
> level-triggered mode. I'm polling for input, output, rdhup, and forcibly for
> hup. I figured that when I get EPOLLRDHUP that doesn't mean there is no more
> data to read, presumably I must empty the kernel buffer by continously reading
> from the fd until I get zero as a result indicating rdhup. This means
> EPOLLRDHUP is delivered asynchronously and thus is pretty useless to me.
> 
> On the other hand, EPOLLHUP seems to be delivered synchronously. If it isn't
> then this means I get constantly notified about a closed fd until I have read
> all data from the kernel buffer and remove the fd from the epoll set (which
> makes being forced to hup notifications useless to me).
> 
> (And no, don't tell me to use edge-triggered mode, I have reasons to use
> level-triggered.)
> 
> Question: Is my observation correct that EPOLLRDHUP is delivered
> asynchronously in contrast to EPOLLHUP? Or is EPOLLHUP delivered
> asynchronously as well?

There's nothing synchronous in epoll WRT userspace. If you use epoll LT, 
you don't need RDHUP. RDHUP was introduced to deal with special 
connection states and epoll ET. Take a peek here for a detailed 
description about how/why RDHUP was introduced:

http://lkml.org/lkml/2003/7/12/116


- Davide

Re: (e)poll and (rd)hup questions

[Michael Noisternig]

David wrote:
> There's nothing synchronous in epoll WRT userspace. If you use epoll LT, 
> you don't need RDHUP. RDHUP was introduced to deal with special 
> connection states and epoll ET. Take a peek here for a detailed 
> description about how/why RDHUP was introduced:
> 
> http://lkml.org/lkml/2003/7/12/116
> 
> 
> - Davide

Thanks a lot for the explanation about RDHUP. Now that leaves me with 
only one question which I'm still not sure about: When I get HUP can 
there still be data in the kernel buffer to read out? (I guess so.)

Thanks again,
Michael

--Original message--
Michael Noisternig schrieb:
> Hello folks,
> 
> I hope this is a right place to ask this question. (Sorry if it isn't, 
> please hint me where to post in that case.)
> 
> This has been bugging me for quite some while now. I'm using epoll in 
> default level-triggered mode. I'm polling for input, output, rdhup, and 
> forcibly for hup. I figured that when I get EPOLLRDHUP that doesn't mean 
> there is no more data to read, presumably I must empty the kernel buffer 
> by continously reading from the fd until I get zero as a result 
> indicating rdhup. This means EPOLLRDHUP is delivered asynchronously and 
> thus is pretty useless to me.
> 
> On the other hand, EPOLLHUP seems to be delivered synchronously. If it 
> isn't then this means I get constantly notified about a closed fd until 
> I have read all data from the kernel buffer and remove the fd from the 
> epoll set (which makes being forced to hup notifications useless to me).
> 
> (And no, don't tell me to use edge-triggered mode, I have reasons to use 
> level-triggered.)
> 
> Question: Is my observation correct that EPOLLRDHUP is delivered 
> asynchronously in contrast to EPOLLHUP? Or is EPOLLHUP delivered 
> asynchronously as well?
> 
> Thanks, anwers would help a lot!!!
> Michael
>