zfs hang in zio->io_cv) with dd read

[John Hay]

On Thu, Oct 07, 2010 at 02:35:31PM +0200, Ivan Voras wrote:
> On 10/07/10 14:15, John Hay wrote:
> >Hi,
> >
> >I got hold of a SunFire X4500 with 48 X 500G disks and thought to try
> >FreeBSD 8-stable with zfs on it.
> >
> >I have setup the two boot disks in a zfs mirror and then the rest in
> >a pool of 6 X raidz2 of 7 disks each.
> >
> >I have created a 10G file with dd in the second pool, but if I try to read
> >it with dd, dd will hang in "zio->io_cv)" according to ^T. This happens
> >everytime. The first time I saw messages about an interrupt storm, so I
> >have put "hw.intr_storm_threshold=10000" in /etc/sysctl.conf. According to
> >"systat -vm 1" there is atapci for 2-3 seconds and then it is quiet.
> 
> There are two things you could try: 1) use the AHCI driver 
> (ahci_load="YES" in /boot/loader.conf) and 2) disable superpages, they 
> don't get along on a few models of Opterons (vm.pmap.pg_ps_enabled=0 in 
> /boot/loader.conf).

ahci does not grab them. According to the ahci man page, it can handle
Marvell 88SX61xx, while these are MV88SX6081 according to pciconf -lcv:

atapci0 at pci0:1:1:0:     class=0x010000 card=0x11ab11ab chip=0x608111ab rev=0x09 hdr=0x00
    vendor     = 'Marvell Semiconductor (Was: Galileo Technology Ltd)'
    device     = 'MV88SX6081 8-port SATA II PCI-X Controller'
    class      = mass storage
    subclass   = SCSI
    cap 01[40] = powerspec 2  supports D0 D3  current D0
    cap 05[50] = MSI supports 1 message, 64 bit 
    cap 07[60] = PCI-X 64-bit supports 133MHz, 512 burst read, 4 split transactions

I have also set vm.pmap.pg_ps_enabled=0 in loader.conf, but that did not
make a difference either. :-(

Once dd hang in that "zio->io_cv)" state the rest of the machine is ok
and everything works as long as you stay away from the directory where
the file is that you dd from.

There are no messages in dmesg or /var/log/messages.

John
-- 
John Hay -- jhay at meraka.csir.co.za / jhay at FreeBSD.org