Re: [PATCH] exec: fetch the alignment of Linux devdax pmem character device nodes

[Liu, Jingqi]

On 4/8/2020 5:42 PM, Joao Martins wrote:
> On 4/8/20 3:25 AM, Liu, Jingqi wrote:
> > On 4/8/2020 2:28 AM, Joao Martins wrote:
> > > On 4/7/20 5:55 PM, Dan Williams wrote:
> > > > On Tue, Apr 7, 2020 at 4:01 AM Joao Martins <address@hidden> wrote:
> > > > > Perhaps, you meant instead:
> > > > >
> > > > >           /sys/dev/char/%d:%d/align
> > > > Hmm, are you sure that's working?
> > > It is, except that I made the slight mistake of testing with a bunch of wip
> > > patches on top which one of them actually adds the 'align' to child dax device.
> > >
> > > Argh, my apologies - and thanks for noticing.
> > >
> > > > I expect the alignment to be found
> > > > in the region device:
> > > >
> > > > /sys/class/dax:
> > > > /sys/devices/LNXSYSTM:00/LNXSYBUS:00/ACPI0012:00/ndbus1/region1/dax1.1/dax1.0
> > > > $(readlink -f /sys/dev/char/253\:263)/../align
> > > > $(readlink -f /sys/dev/char/253\:263)/device/align
> > > >
> > > >
> > > > /sys/bus/dax:
> > > > /sys/devices/LNXSYSTM:00/LNXSYBUS:00/ACPI0012:00/ndbus1/region1/dax1.0/dax1.0
> > > > $(readlink -f /sys/dev/char/253\:265)/../align
> > > > $(readlink -f /sys/dev/char/253\:265)/device/align <-- No such file
> > > >
> > > > The use of the /sys/dev/char/%d:%d/device is only supported by the
> > > > deprecated /sys/class/dax.
> > Hi Dan,
> >
> > Thanks for your comments.
> >
> > Seems it is a mistake.
> >
> > It should be: $(readlink -f /sys/dev/char/253\:263)/../../align
> Hmm, perhaps you have an extra '../' in the path? This works for me:
>
> # ls $(readlink -f /sys/dev/char/252\:0/../align)
> /sys/devices/platform/e820_pmem/ndbus0/region0/dax0.0/dax0.0/../align
> # cat $(readlink -f /sys/dev/char/252\:0)/../align
> 2097152
> # cat /sys/dev/char/252\:0/../align
> 2097152

Hi Joao,

Hmm, I need to have an extra '../' in the path. The details are as follows:

# ll /dev/dax2.0
crw------- 1 root root 251, 5 Mar 20 13:35 /dev/dax2.0
# uname -r
5.6.0-rc1-00044-gb19e8c684703
# readlink -f /sys/dev/char/251\:5/
/sys/devices/LNXSYSTM:00/LNXSYBUS:00/ACPI0012:00/ndbus0/region2/dax2.1/dax/dax2.0
# ls $(readlink -f /sys/dev/char/251\:5)/../align
ls: cannot access 
'/sys/devices/LNXSYSTM:00/LNXSYBUS:00/ACPI0012:00/ndbus0/region2/dax2.1/dax/dax2.0/../align': 
No such file or directory
# ls $(readlink -f /sys/dev/char/251\:5)/../dax_region/align
ls: cannot access 
'/sys/devices/LNXSYSTM:00/LNXSYBUS:00/ACPI0012:00/ndbus0/region2/dax2.1/dax/dax2.0/../dax_region/align': 
No such file or directory
# ls $(readlink -f /sys/dev/char/251\:5)/../../align
/sys/devices/LNXSYSTM:00/LNXSYBUS:00/ACPI0012:00/ndbus0/region2/dax2.1/dax/dax2.0/../../align
# ls $(readlink -f /sys/dev/char/251\:5)/../../dax_region/align
/sys/devices/LNXSYSTM:00/LNXSYBUS:00/ACPI0012:00/ndbus0/region2/dax2.1/dax/dax2.0/../../dax_region/align
# lsmod|grep pmem
dax_pmem_compat        16384  0
device_dax             20480  1 dax_pmem_compat
dax_pmem_core          16384  1 dax_pmem_compat
# lsmod|grep dax
dax_pmem_compat        16384  0
device_dax             20480  1 dax_pmem_compat
dax_pmem_core          16384  1 dax_pmem_compat

Seems some configurations are different ?

Can you share your info as above ? Thanks.

> > > I don't have the deprecated dax class enabled as could you tell, so the second
> > > case is what I was testing. Except it wasn't a namespace/nvdimm but rather an
> > > hmem device-dax.
> > >
> > > '../align' though covers only one case? What about hmem which '../align' returns
> > > ENOENT; perhaps using '../dax_region/align' instead which is common to both?
> > > Albeit that wouldn't address the sub-division devices (that I mention above)
> > Seems that you mean to use $(readlink -f
> > /sys/dev/char/253\:263)/../../dax_region/align.
> >
> > Right ?
> An extra '../' ?
>
> # ls $(readlink -f /sys/dev/char/252\:0/../dax_region/align)
> /sys/devices/platform/e820_pmem/ndbus0/region0/dax0.0/dax0.0/../align
> # cat $(readlink -f /sys/dev/char/252\:0)/../dax_region/align
> 2097152
> # cat /sys/dev/char/252\:0/../dax_region/align
> 2097152
>
> For HMAT/hmem devdax, though, only 'dax_region/align' is available for now:
>
> # ls $(readlink -f /sys/dev/char/252:0)/../align
> ls: cannot access /sys/devices/platform/hmem.0/dax0.0/../align: No such file or
> directory
> # ls $(readlink -f /sys/dev/char/252:0)/../dax_region/align
> /sys/devices/platform/hmem.0/dax0.0/../dax_region/align
> # cat $(readlink -f /sys/dev/char/252:0)/../dax_region/align
> 2097152
>
> The 'dax_region/align' was just an idea mainly because it's common to both
> device-dax devices -- not sure how others feel about it.

Seems it's reasonable. I need to sync the above path with yours.

Thanks,

Jingqi

>         Joao
>
> > Thanks,
> >
> > Jingqi
> >
> > > > The current /sys/bus/dax device-model can
> > > > be a drop in replacement as long as software is not written to the
> > > > /sys/class sysfs layout, i.e. it uses ../ instead of device/ to walk
> > > > to the region properties.
> > > /nods