Re: [PATCH] libnvdimm: Clarify nd_pfn_init() flow

[Dan Williams]

On Mon, Jan 21, 2019 at 12:57 PM Wei Yang <richard.weiyang@xxxxxxxxx> wrote:
>
> On Mon, Jan 21, 2019 at 10:04:40AM -0800, Dan Williams wrote:
> >On Sun, Jan 20, 2019 at 11:51 PM Wei Yang <richardw.yang@xxxxxxxxxxxxxxx> wrote:
> >> On Fri, Jan 18, 2019 at 04:47:23PM -0800, Dan Williams wrote:
> >[..]
> >> Also, I have one confusion about your saying: two probes.
> >>
> >> If the two probes are:
> >>
> >>     * for dax%d.%d: 1. nd_dax_probe 2. dax_pmem_probe
> >>     * for pfn%d.%d: 1. nd_pfn_probe 2. nd_pmem_probe
> >>
> >> Then, if the first probe fails, the device itself would be destroyed. How the
> >> second probe do its job?
> >>
> >> >       rc = nd_pfn_validate(nd_pfn, sig);
> >> >       if (rc != -ENODEV)
> >> >               return rc;
> >
> >Here is an example path for a device-dax instance:
> >
> >    /sys/devices/platform/e820_pmem/ndbus0/region0/dax0.1/dax0.0
> >
> >In this case the order of events is:
> >
> >1/ region0 discovers it contains a pmem namespace and registers namespace0.0
> >2/ The pmem namespace driver calls nd_dax_probe() to check for the
> >presence of a device-dax configuration
> >3/ If present, nd_pfn_validate() returns 0 and nd_dax_probe()
> >registers the dax0.1 device (this is a libnvdimm 'personality device).
> >4/ When nd_pmem_probe() sees nd_dax_probe() return 0 it in turn fails
> >the probe of namespace0.0 with -ENXIO. All devm allocations during the
> >probe of namespace0.0 are released.
>
> I may have another opinion here.
>
> The probe return error means the device will not attach to this driver.
> But the device itself it not released.
>
> We allocate devm on one device and those memory will be released when
> the device is destroyed. If I am correct.
>
> This means at this point, pfn_sb's memory still exists in the system.
> Even finally we will release it, when namespace0.0 is destroyed.

No, that's not the way devm works. Memory allocated by devm is
released at ->probe() failure, or after ->remove()

See the devres_release_all() call in drivers/base/dd.c::really_probe()