Re: [PATCH v2 3/4] cxl/pci: Fix sanitize notifier setup

[Dan Williams <dan.j.williams@intel.com>]

Davidlohr Bueso wrote:
> On Tue, 03 Oct 2023, Dan Williams wrote:
> 
> >Davidlohr Bueso wrote:
> >> On Fri, 29 Sep 2023, Dan Williams wrote:
> >>
> >> >+static int cxl_sanitize_setup_notifier(struct cxl_memdev *cxlmd)
> >> >+{
> >> >+	struct cxl_dev_state *cxlds = cxlmd->cxlds;
> >> >+	struct cxl_memdev_state *mds = to_cxl_memdev_state(cxlds);
> >> >+	struct device *dev = cxlds->dev;
> >>
> >> This wants to be dev = &cxlmd->dev;
> >
> >No, the notifier needs to be torn down in the cxl_pci teardown path. If
> >cxlmd->dev was the devm operations host then this notifier would need to
> >be setup from the cxl_mem driver, not cxl_pci. The teardown order of
> >cxl_pci ends with @cxlds->dev as the devm host ends up as:
> >
> >cxl_sanitize_teardown_notifier()
> >cxl_memdev_unregister()
> >
> >...otherwise if the @cxlmd->dev is used then the devm callback may not
> >fire until device_release() time since it is possible that the cxl_mem
> >driver never attaches to trigger the typical devm action around
> >->remove() time.
> 
> As is, the first sysfs_get_dirent() fails, fyi.

Yup, thanks for testing, I am using the wrong device to lookup the sysfs
entry.

What would it take to get your test scenario into cxl_test... or at
least a recipe for it in QEMU that can be checked into the cxl-cli
tests?