Re: [PATCH v2 4/4] test_sysfs: demonstrate deadlock fix

[Luis Chamberlain]

On Sat, Jul 03, 2021 at 06:49:46AM +0200, Greg KH wrote:
> On Fri, Jul 02, 2021 at 05:46:32PM -0700, Luis Chamberlain wrote:
> > +#define MODULE_DEVICE_ATTR_FUNC_STORE(_name) \
> > +static ssize_t module_ ## _name ## _store(struct device *dev, \
> > +				   struct device_attribute *attr, \
> > +				   const char *buf, size_t len) \
> > +{ \
> > +	ssize_t __ret; \
> > +	if (!try_module_get(THIS_MODULE)) \
> > +		return -ENODEV; \
> > +	__ret = _name ## _store(dev, attr, buf, len); \
> > +	module_put(THIS_MODULE); \
> > +	return __ret; \
> > +}
> 
> As I have pointed out before, doing try_module_get(THIS_MODULE) is racy
> and should not be added back to the kernel tree.  We got rid of many
> instances of this "bad pattern" over the years, please do not encourage
> it to be added back as others will somehow think that it correct code.

It is noted this is used in lieu of any agreed upon solution to
*demonstrate* how this at least does fix it. In this case (and in the
generic solution I also had suggested for kernfs a while ago), if the
try fails, we give up. If it succeeds, we now know we can rely on the
device pointer. If the refcount succeeds, can the module still not
be present? Is try_module_get() racy in that way? In what way is it
racy and where is this documented? Do we have a selftest to prove the
race?

  Luis