Opening of blocking FIFOs not reliable?

Opening of blocking FIFOs not reliable?

[Teemu Koponen]

Hi,

Opening a FIFO for WR_ONLY should release a previously blocked RD_ONLY 
open. I suspect this is not guaranteed on a heavily loaded Linux box.

Consider fs/fifo.c (of 2.6.13-rc6-git4) and its wait_for_partner() and 
wake_up_partner() functions. I think wait_for_partner()'s while loop 
never exits, *iff* the writer, after opening the FIFO, closes it before 
the reader gets scheduled. Shouldn't the writer's open() block until 
the reader's open() is done?

Teemu

--

Re: Opening of blocking FIFOs not reliable?

[Alan Cox]

On Gwe, 2005-08-12 at 20:18 -0700, Teemu Koponen wrote: 
> the reader gets scheduled. Shouldn't the writer's open() block until 
> the reader's open() is done?

Not neccessarily - but you are correct that it should lead to the
readers open completing and then EOFs being returned. Do you have a test
case for the problem?

Alan

Re: Opening of blocking FIFOs not reliable?

[Manfred Spraul]

Teemu wrote:

>Opening a FIFO for WR_ONLY should release a previously blocked RD_ONLY 
>open. I suspect this is not guaranteed on a heavily loaded Linux box.
>  
>
Do you have a test case?
IIRC we had that bug, and it was fixed by adding PIPE_WCOUNTER:
PIPE_WRITERS counts the number of writers. This one is decreased during 
close(). PIPE_WCOUNTER counts how often a writer was seen. It's never 
decreased. Readers that wait for a writer wait until PIPE_WCOUNTER 
changes, they do not look at PIPE_WRITERS.

--
    Manfred

Re: Opening of blocking FIFOs not reliable?

[Teemu Koponen]

On Aug 14, 2005, at 3:15, Manfred Spraul wrote:

>> Opening a FIFO for WR_ONLY should release a previously blocked 
>> RD_ONLY open. I suspect this is not guaranteed on a heavily loaded 
>> Linux box.
>>
> Do you have a test case?
> IIRC we had that bug, and it was fixed by adding PIPE_WCOUNTER:
> PIPE_WRITERS counts the number of writers. This one is decreased 
> during close(). PIPE_WCOUNTER counts how often a writer was seen. It's 
> never decreased. Readers that wait for a writer wait until 
> PIPE_WCOUNTER changes, they do not look at PIPE_WRITERS.

Ah, I missed the semantics of WCOUNTER.  True, writer's open should 
always release the reader' s open (as Alan pointed out).

I do have a test application, which experienced dead-locked readers' 
opens before I artificially delayed writers' closes. The application 
runs on PlanetLab, which uses VServer, though. I'll try to construct 
next a minimal test case based on the application and reproduce the 
behavior on a vanilla box...

Teemu

--