poor performance when building the Solaris kernel

[peter.dennis@oracle.com]

Host Operating System: OpenSolaris - build snv_134a Guest: OpenSolaris - build snv_134a - with the guest additions installed Physical h/w: Sun Fire X4600 M2 (8 quad core AMD CPU), 65Gb memory

Issue: When building the Solaris kernel (ON) the performance of the Virtual Machine drops and the build ultimately fails. Upon initial investigation it would appear some I/O issue occurs.

The guest reports the following errors:

Jun 16 13:31:22 on-osol1005-bld-x scsi: [ID 107833 kern.warning] WARNING: /pci@0,0/pci-ide@1,1/ide@1 (ata1): Jun 16 13:31:22 on-osol1005-bld-x timeout: abort request, target=0 lun=0 Jun 16 13:31:22 on-osol1005-bld-x scsi: [ID 107833 kern.warning] WARNING: /pci@0,0/pci-ide@1,1/ide@1 (ata1): Jun 16 13:31:22 on-osol1005-bld-x timeout: abort device, target=0 lun=0 Jun 16 13:31:22 on-osol1005-bld-x scsi: [ID 107833 kern.warning] WARNING: /pci@0,0/pci-ide@1,1/ide@1 (ata1): Jun 16 13:31:22 on-osol1005-bld-x timeout: reset target, target=0 lun=0 Jun 16 13:31:22 on-osol1005-bld-x scsi: [ID 107833 kern.warning] WARNING: /pci@0,0/pci-ide@1,1/ide@1 (ata1): Jun 16 13:31:22 on-osol1005-bld-x timeout: reset bus, target=0 lun=0

and the VBox.log showed:

02:08:33.194 PIIX3 ATA: Ctl#0: RESET, DevSel=0 AIOIf=0 CmdIf0=0xc8 (-1 usec ago) CmdIf1=0x00 (-1 usec ago) 02:08:33.227 PIIX3 ATA: Ctl#0: finished processing RESET 02:08:49.731 PIIX3 ATA: Ctl#0: RESET, DevSel=0 AIOIf=0 CmdIf0=0xe7 (-1 usec ago) CmdIf1=0x00 (-1 usec ago) 02:08:49.761 PIIX3 ATA: Ctl#0: finished processing RESET 02:15:54.858 PIIX3 ATA: Ctl#0: RESET, DevSel=0 AIOIf=0 CmdIf0=0xc8 (-1 usec ago) CmdIf1=0x00 (-1 usec ago) 02:15:54.892 PIIX3 ATA: Ctl#0: finished processing RESET 02:18:16.518 PIIX3 ATA: Ctl#0: RESET, DevSel=0 AIOIf=0 CmdIf0=0xca (-1 usec ago) CmdIf1=0x00 (-1 usec ago) 02:18:16.550 PIIX3 ATA: Ctl#0: finished processing RESET 02:30:28.437 PIIX3 ATA: Ctl#0: RESET, DevSel=0 AIOIf=0 CmdIf0=0xc8 (-1 usec ago) CmdIf1=0x00 (-1 usec ago) 02:30:28.478 PIIX3 ATA: Ctl#0: finished processing RESET 02:35:41.089 PIIX3 ATA: Ctl#0: RESET, DevSel=0 AIOIf=0 CmdIf0=0xc8 (-1 usec ago) CmdIf1=0x00 (-1 usec ago) 02:35:41.119 PIIX3 ATA: Ctl#0: finished processing RESET 02:49:37.093 PIIX3 ATA: LUN#2: performing device RESET

This issue has not been isolated to just this particular system - another build system - using the same software mix had the same issues.

The host was not busy at all.

I have marked this as major as this is essentially preventing the use of VB as tool to host build systems.

I have attached the following information:

o VBox.log o showvminfo o showhdinfo

Any information on what could be the problem or any diagnostics that can help track it down appreciated.

[peter.dennis@oracle.com]

VBoxManage showhdinfo

[peter.dennis@oracle.com]

VBoxManage showvminfo

[peter.dennis@oracle.com]

Log file with the resets in it

[Juergen Keil]

Hmm, your vbox solaris guest is configured to use 8GB of memory, and is configured to use the P-ATA disk controller.

AFAIK, DMA on the legacy P-ATA ide controller is limited to < 4GB physical memory addresses, so I suspect that you may see a slow down because the system often has to wait for free memory at addresses < 4GB to set up DMA bounce buffers. Maybe with a massive slow down waiting for DMA memory the system might even timeout some ata commands, which leads to the controller resets.

What happens when you configure the solaris vbox guest to use less than 4GB of memory (but still use the P-ATA controller) ? Is there still a slow down with such a setup?

The other test would be a vbox guest with 8GB of memory, but using the AHCI S-ATA ontroller instead of P-ATA. I think AHCI is able to access memory >= 4GB.

[peter.dennis@oracle.com]

I initially started this machine with 4Gb of memory and only increased it to 8Gb to see if that helped.

I'll try out the AHCI S-ATA controller and report back.

thanks

[Juergen Keil]

There is a message on the opensolaris forums that describes similar problems wuth the ATA controller in the opensolaris vbox guest when building opensolaris from source.

​http://www.opensolaris.org/jive/thread.jspa?messageID=495588#495588

That user reports that the problem exists when using multiple virtual cpus. With one cpu there apparently are no ATA timeout errors.

[Frank Mehnert]

Still relevant with VBox 4.0.6?

[peter.dennis@oracle.com]

Do not know as the approach to using VB for this purpose has been stopped for the time being.

Suggest closing this one out until the issue can be revisited using the later versions of the software.