5:30 a.m.
On Wed, Mar 29, 2017 at 1:19 PM, Jeff Moyer <jmoyer(a)redhat.com> wrote:
Dan Williams <dan.j.williams(a)intel.com> writes:
> On Wed, Mar 29, 2017 at 1:02 PM, Jeff Moyer <jmoyer(a)redhat.com> wrote:
>> Dan Williams <dan.j.williams(a)intel.com> writes:
>>
>>> +check_min_kver()
>>> +{
>>> + local ver="$1"
>>> + : "${KVER:=$(uname -r)}"
>>> +
>>> + [ -n "$ver" ] || return 1
>>> + [[ "$ver" == "$(echo -e "$ver\n$KVER" | sort
-V | head -1)" ]]
>>> +}
>>> +
>>> +check_min_kver "4.11" || { echo "kernel $KVER may lack latest
device-dax fixes"; exit $rc; }
>>
>> Can we stop with this kernel version checking, please? Test to see if
>> you can create a device dax instance. If not, skip the test. If so,
>> and if you have a kernel that isn't fixed, so be it, you'll get
>> failures.
>
> I'd rather not. It helps me keep track of what went in where. If you
> want to run all the tests on a random kernel just do:
>
> KVER="4.11.0" make check
This, of course, breaks completely with distro kernels.
Why does this break distro kernels? The KVER variable overrides "uname -r"
You don't see
this kind of checking in xfstests, for example. git helps you keep
track of what changes went in where (see git describe --contains). It's
weird to track that via your test harness. So, I would definitely
prefer to move to a model where we check for features instead of kernel
versions.
I see this as a deficiency of xfstests. We have had to go through and
qualify and track each xfstest and why it may fail with random
combinations of kernel, xfsprogs, or e2fsprogs versions. I'd much
prefer that upstream xfstests track the minimum versions of components
to make a given test pass so we can stop doing it out of tree.
>>> +
>>> +set -e
>>> +trap 'err $LINENO' ERR
>>> +
>>> +if ! fio --enghelp | grep -q "dev-dax"; then
>>> + echo "fio lacks dev-dax engine"
>>> + exit 77
>>> +fi
>>> +
>>> +dev=$(./dax-dev)
>>> +for align in 4k 2m 1g
>>> +do
>>> + json=$($NDCTL create-namespace -m dax -a $align -f -e $dev)
>>> + chardev=$(echo $json | jq -r ". | select(.mode ==
\"dax\") | .daxregion.devices[0].chardev")
>>> + if [ align = "1g" ]; then
>>> + bs="1g"
>>> + else
>>> + bs="2m"
>>> + fi
>>
>> I'm not sure the blocksize even matters.
>>
>
> iirc it affects the alignment of the mmap() request. So for example
> bs=4k should fail when the alignment is larger.
I think iomem_align is a better hint for that, no?
Certainly sounds better, I'll check it out.