Re: Collapsing RT signals ...

Re: Collapsing RT signals ...

[Dan Kegel]

Davide Libenzi <davidel@xmailserver.org> wrote:
> I'm making some test with RT signals and looking at how they're implemented
> inside the kernel.
> After having experienced frequent queue overflow signals I looked at how
> signals are queued inside the task_struct.
> There's no signals optimization inside and this make the queue length depending
> on the request rate instead of the number of connections.
> It can happen that two ( or more ) POLL_IN signals are queued with a single
> read() that sweep the buffer leaving other signals to issue reads ( read this
> as user-mode / kernel-mode switch ) that will fail due lack of data.
> So for every "superfluous" signal we'll have two user-mode / kernel-mode
> switches, one for signal delivery and one for a failing read().
> I'm just thinking at a way to optimize the signal delivery that is ( draft ) :
> ...

I agree, the queue overflow case is a pain in the butt.

Before you get too far coding up your idea, have you read
http://marc.theaimsgroup.com/?l=linux-kernel&m=99023775430848&w=2
?  He's already implemented and benchmarked a variation on this
idea, maybe you could vet his code.  He has taken it a step
further than perhaps you were going to.

(See also http://www.kegel.com/c10k.html#nb.sigio )

- Dan

Re: Collapsing RT signals ...

[Davide Libenzi]

On 25-Jun-2001 Dan Kegel wrote:
> Davide Libenzi <davidel@xmailserver.org> wrote:
>> I'm making some test with RT signals and looking at how they're implemented
>> inside the kernel.
>> After having experienced frequent queue overflow signals I looked at how
>> signals are queued inside the task_struct.
>> There's no signals optimization inside and this make the queue length
>> depending
>> on the request rate instead of the number of connections.
>> It can happen that two ( or more ) POLL_IN signals are queued with a single
>> read() that sweep the buffer leaving other signals to issue reads ( read
>> this
>> as user-mode / kernel-mode switch ) that will fail due lack of data.
>> So for every "superfluous" signal we'll have two user-mode / kernel-mode
>> switches, one for signal delivery and one for a failing read().
>> I'm just thinking at a way to optimize the signal delivery that is ( draft )
>> :
>> ...
> 
> I agree, the queue overflow case is a pain in the butt.
> 
> Before you get too far coding up your idea, have you read
> http://marc.theaimsgroup.com/?l=linux-kernel&m=99023775430848&w=2
> ?  He's already implemented and benchmarked a variation on this
> idea, maybe you could vet his code.  He has taken it a step
> further than perhaps you were going to.

I'll do for sure, thank You.


> 
> (See also http://www.kegel.com/c10k.html#nb.sigio )

I already knew Your document, pretty cool.




- Davide

Re: Collapsing RT signals ...

[Davide Libenzi]

On 25-Jun-2001 Dan Kegel wrote:
> Davide Libenzi <davidel@xmailserver.org> wrote:
>> I'm making some test with RT signals and looking at how they're implemented
>> inside the kernel.
>> After having experienced frequent queue overflow signals I looked at how
>> signals are queued inside the task_struct.
>> There's no signals optimization inside and this make the queue length
>> depending
>> on the request rate instead of the number of connections.
>> It can happen that two ( or more ) POLL_IN signals are queued with a single
>> read() that sweep the buffer leaving other signals to issue reads ( read
>> this
>> as user-mode / kernel-mode switch ) that will fail due lack of data.
>> So for every "superfluous" signal we'll have two user-mode / kernel-mode
>> switches, one for signal delivery and one for a failing read().
>> I'm just thinking at a way to optimize the signal delivery that is ( draft )
>> :
>> ...
> 
> I agree, the queue overflow case is a pain in the butt.
> 
> Before you get too far coding up your idea, have you read
> http://marc.theaimsgroup.com/?l=linux-kernel&m=99023775430848&w=2

Double thank You Dan, they did exactly what I want to do :)




- Davide

Collapsing RT signals ...

[Davide Libenzi]

I'm making some test with RT signals and looking at how they're implemented
inside the kernel.
After having experienced frequent queue overflow signals I looked at how
signals are queued inside the task_struct.
There's no signals optimization inside and this make the queue length depending
on the request rate instead of the number of connections.
It can happen that two ( or more ) POLL_IN signals are queued with a single
read() that sweep the buffer leaving other signals to issue reads ( read this
as user-mode / kernel-mode switch ) that will fail due lack of data.
So for every "superfluous" signal we'll have two user-mode / kernel-mode
switches, one for signal delivery and one for a failing read().
I'm just thinking at a way to optimize the signal delivery that is ( draft ) :


struct sigqueue {
        struct sigqueue *next;
        struct sigqueue **fsig;
        siginfo_t info;
};


struct fown_struct {
        int pid;                /* pid or -pgrp where SIGIO should be sent */
        uid_t uid, euid;        /* uid/euid of process setting the owner */
        int signum;             /* posix.1b rt signal to be delivered on IO */
        struct sigqueue *sig_hint;
};



At the end we'll have ( where fsig = &file->fown.sig_hint ) :


static int send_signal_hint(int sig, struct siginfo *info,
        struct sigpending *signals, struct sigqueue **fsig) {

        if (*fsig != NULL)
                merge_info(&(*fsig)->info, info);
        else {
                *fsig = allog_sig_queue();

                (*fsig)->fsig = fsig;

                copy_info(&(*fsig)->info, info);

                ...
                add-to-queue();
                ...
        }

        ...

}


When the message is delivered we'll have :

void clean_sig_hint(struct sigqueue *qsig) {

        if (qsig->fsig != NULL) {
                *(qsig->fsig) = NULL;
                qsig->fsig = NULL;
        }

}


When the file is closed :

void clean_fown_sighint(struct fown_struct *fown) {

        if (fown->sig_hint != NULL) {
                fown->sig_hint->fsig = NULL;
                fown->sig_hint = NULL;
        }

}


All these ops must be done under  sigmask_lock  lock ( send_signal() and signal
dequeue already does ).
This will make multiple signal from the same file to be stocked inside the same
slot reducing the RT signal traffic.

Comments ?






- Davide