Some time you said that arima board connect all PCI-X slots of amd 8131 irq to amd 8111 ioapic intead of 8131.
So even disable ioapic of 8131, device on PCI-X still works properly.
After enable ioapic of 8131, devices that are not using prefmem works well.
YH.
-----邮件原件----- 发件人: ebiederman@lnxi.com [mailto:ebiederman@lnxi.com] 发送时间: 2004年3月1日 10:48 收件人: ron minnich 抄送: YhLu; ebiederman@lnxi.com; LinuxBIOS 主题: Re: Prefmem of bus 3
ron minnich rminnich@lanl.gov writes:
On Fri, 27 Feb 2004, YhLu wrote:
Have you ever tested myrinet card on K8 MB under LinuxBIOS?
we have a 1408-node K8 machine with myrinet working under linuxbios, and a
256-node K8 machine working with linuxbios. Arima HDAMA mainboards.
I wonder what's going on here, I have not had time to look in detail.
So the problem appears to be that the prefmem region on the upper bus is to large.
I wonder if why this work on LANL clusters and not on others is the result of code skew.
Lightning nodes with myrinet only seem to have one prefmem region.
Yet another reminder that I really need to finish syncing up the code bases.
Eric
Ron,
I have verified the myerinet card with s2882 in LinuxBIOS. It works well.
Eric,
I wonder if round function in pci_device.c cause the bus prefmem region calculation problem.
Also according to suggestion from quadrics and amd, it seems that BIOS and kernel need to be patched to allow the prefmem region ( after mmio_basek) to be put the mtrr as write-back to get the better perfmormance. i.e. 2.5us->1. 8us 0-byte ping pong. So may need to change the write_linuxbios_table and update the struct mem_reagion.
I have attached the boot log with QM card. Also found the display mmio region of 0:18:1 is shown ad 0:18.0, but the real reg content is right. But the prefmem region register of 1:2:0 is not the same as 0xe0000000 to 0xffffffff, and it (0x24) is "01 e0 f1 ff"
Can some one send me the boot log of quadrics on other K8 board?
Regards
YH.
-----邮件原件----- 发件人: YhLu 发送时间: 2004年3月1日 13:15 收件人: ebiederman@lnxi.com; ron minnich 抄送: LinuxBIOS 主题: 答复: Prefmem of bus 3
Some time you said that arima board connect all PCI-X slots of amd 8131 irq to amd 8111 ioapic intead of 8131.
So even disable ioapic of 8131, device on PCI-X still works properly.
After enable ioapic of 8131, devices that are not using prefmem works well.
YH.
-----邮件原件----- 发件人: ebiederman@lnxi.com [mailto:ebiederman@lnxi.com] 发送时间: 2004年3月1日 10:48 收件人: ron minnich 抄送: YhLu; ebiederman@lnxi.com; LinuxBIOS 主题: Re: Prefmem of bus 3
ron minnich rminnich@lanl.gov writes:
On Fri, 27 Feb 2004, YhLu wrote:
Have you ever tested myrinet card on K8 MB under LinuxBIOS?
we have a 1408-node K8 machine with myrinet working under linuxbios, and a
256-node K8 machine working with linuxbios. Arima HDAMA mainboards.
I wonder what's going on here, I have not had time to look in detail.
So the problem appears to be that the prefmem region on the upper bus is to large.
I wonder if why this work on LANL clusters and not on others is the result of code skew.
Lightning nodes with myrinet only seem to have one prefmem region.
Yet another reminder that I really need to finish syncing up the code bases.
Eric _______________________________________________ Linuxbios mailing list Linuxbios@clustermatic.org http://www.clustermatic.org/mailman/listinfo/linuxbios
Ron,
I have verified the myerinet card with s2882 in LinuxBIOS. It works well.
Eric,
I wonder if round function in pci_device.c cause the bus prefmem region calculation problem.
Also according to suggestion from quadrics and amd, it seems that BIOS and kernel need to be patched to allow the prefmem region ( after mmio_basek) to be put the mtrr as write-back to get the better perfmormance. i.e. 2.5us->1. 8us 0-byte ping pong. So may need to change the write_linuxbios_table and update the struct mem_reagion.
I have attached the boot log with QM card. Also found the display mmio region of 0:18:1 is shown ad 0:18.0, but the real reg content is right. But the prefmem region register of 1:2:0 is not the same as 0xe0000000 to 0xffffffff, and it (0x24) is "01 e0 f1 ff"
Can some one send me the boot log of quadrics on other K8 board?
Regards
YH.
-----邮件原件----- 发件人: YhLu 发送时间: 2004年3月1日 13:15 收件人: ebiederman@lnxi.com; ron minnich 抄送: LinuxBIOS 主题: 答复: Prefmem of bus 3
Some time you said that arima board connect all PCI-X slots of amd 8131 irq to amd 8111 ioapic intead of 8131.
So even disable ioapic of 8131, device on PCI-X still works properly.
After enable ioapic of 8131, devices that are not using prefmem works well.
YH.
-----邮件原件----- 发件人: ebiederman@lnxi.com [mailto:ebiederman@lnxi.com] 发送时间: 2004年3月1日 10:48 收件人: ron minnich 抄送: YhLu; ebiederman@lnxi.com; LinuxBIOS 主题: Re: Prefmem of bus 3
ron minnich rminnich@lanl.gov writes:
On Fri, 27 Feb 2004, YhLu wrote:
Have you ever tested myrinet card on K8 MB under LinuxBIOS?
we have a 1408-node K8 machine with myrinet working under linuxbios, and a
256-node K8 machine working with linuxbios. Arima HDAMA mainboards.
I wonder what's going on here, I have not had time to look in detail.
So the problem appears to be that the prefmem region on the upper bus is to large.
I wonder if why this work on LANL clusters and not on others is the result of code skew.
Lightning nodes with myrinet only seem to have one prefmem region.
Yet another reminder that I really need to finish syncing up the code bases.
Eric _______________________________________________ Linuxbios mailing list Linuxbios@clustermatic.org http://www.clustermatic.org/mailman/listinfo/linuxbios
-
YhLu -
Yinghai Lu