[Xenomai-core] Re:rare brokenness in latency output - FPU bug!

[Xenomai-core] Re:rare brokenness in latency output - FPU bug!

[gilles.chanteperdrix]

---------- Debut du message initial -----------

De     : jan.kiszka@domain.hid
A      : "xenomai-core" xenomai@xenomai.org
Chanteperdrix" gilles.chanteperdrix@xenomai.org
Copies :
Date   : Mon, 18 Sep 2006 11:30:45 +0200
Objet  : rare brokenness in latency output - FPU bug!

> Gilles, what info do you need? It's all the latest on x86 here:
> 2.6.17.13 (.config attached), ipipe-1.4, SVN #1646, gcc4.1.0.

It would be interesting if you could:
- try reverting the last FPU changes
- try to narrow down the issue by reducing the number of
threads created
by switchtest (for example, try keeping only the threads whose
switch
fails)
- try dumping the FPU context before the first fp_regs_set,
chances are that the FPU context is already clobbered before
the first fp_regs_set;
when the FPU context is already clobbered, running fp_regs_set
set the
value of the already set registers to qnan, and fp_regs_check
returns
INT_MAX when the register is set to qnan. I will send tomorrow
the
header I use for converting FPU contexts into human readable form.

--
                 Gilles Chanteperdrix

Accédez au courrier électronique de La Poste
sur www.laposte.net ou sur 3615 LAPOSTENET (0,34€ TTC /mn)
1 Giga de stockage gratuit – Antispam et antivirus intégrés

[Xenomai-core] Re: rare brokenness in latency output - FPU bug!

[Jan Kiszka]

[-- Attachment #1: Type: text/plain, Size: 1465 bytes --]

gilles.chanteperdrix wrote:
> ---------- Debut du message initial -----------
> 
> De     : jan.kiszka@domain.hid
> A      : "xenomai-core" xenomai@xenomai.org
> Chanteperdrix" gilles.chanteperdrix@xenomai.org
> Copies : 
> Date   : Mon, 18 Sep 2006 11:30:45 +0200
> Objet  : rare brokenness in latency output - FPU bug!
> 
>> Gilles, what info do you need? It's all the latest on x86 here:
>> 2.6.17.13 (.config attached), ipipe-1.4, SVN #1646, gcc4.1.0.
> 
> It would be interesting if you could:
> - try reverting the last FPU changes

Meanwhile identified: 1595 for trunk, likely 1597 for 2.2.x (so 2.2.3 is
affected as well).

> - try to narrow down the issue by reducing the number of
> threads created
> by switchtest (for example, try keeping only the threads whose
> switch
> fails)

I tried both with "rtup0 rtup0 rtup_ufpp0 rtup_ufpp0" and only "rtup0
rtup_ufpp0", but that seems to work fine.

> - try dumping the FPU context before the first fp_regs_set,
> chances are that the FPU context is already clobbered before
> the first fp_regs_set; 

I guess in rtup() in the switchtest driver, right? Will try.

> when the FPU context is already clobbered, running fp_regs_set
> set the
> value of the already set registers to qnan, and fp_regs_check
> returns
> INT_MAX when the register is set to qnan. I will send tomorrow
> the 
> header I use for converting FPU contexts into human readable form.
> 

TIA.


[-- Attachment #2: OpenPGP digital signature --]
[-- Type: application/pgp-signature, Size: 250 bytes --]