[openib-general] Re: page allocation failure with kdapl/mthca

[Roland Dreier]

    Hal> Hi Roland, With the latest mthca changes (this test worked
    Hal> prior to these changes), I get the following running kdapl in
    Hal> "loopback":

You're running out of memory:

    Hal> Jun 21 08:56:09 localhost kernel: DT_Mdep_Thread_: page
    Hal> allocation failure. order:0, mode:0x20

An order 0 allocation failing means the kernel can't even find a
single free page.  However, the mode of 0x20 means this is effectively
a GFP_ATOMIC allocation.

I just checked in a change that converts from using
pci_alloc_consistent(), which is always GFP_ATOMIC, to
dma_alloc_coherent(..., GFP_KERNEL).  This should help with this
symptom.

I'm not sure what recent changes could have exposed this problem with
your test.

 - R.