Re: [PATCH v3 1/3] msi: free msi_desc entry only after we've released the kobject

[Veaceslav Falico <vfalico@redhat.com>]

On Tue, Oct 29, 2013 at 09:34:28AM -0700, Linus Torvalds wrote:
>On Tue, Oct 29, 2013 at 3:30 AM, Veaceslav Falico <vfalico@redhat.com> wrote:
>>                 /*
>>                  * Its possible that we get into this path
>>                  * When populate_msi_sysfs fails, which means the entries
>>                  * were not registered with sysfs.  In that case don't
>> -                * unregister them.
>> +                * unregister them, and just free. Otherwise the
>> +                * kobject->release will take care of freeing the entry via
>> +                * msi_kobj_release().
>>                  */
>>                 if (entry->kobj.parent) {
>>                         kobject_del(&entry->kobj);
>>                         kobject_put(&entry->kobj);
>> +               } else {
>> +                       kfree(entry);
>>                 }
>> -
>> -               list_del(&entry->list);
>> -               kfree(entry);
>
>So this code sequence still makes me very unhappy.
>
>Why does not just a simple unconditional
>
>        kobject_del(&entry->kobj);
>        kobject_put(&entry->kobj);
>
>work for the "not registered with sysfs" case? And if the sysfs code
>really gets confused, why not
>
>        if (entry->kobj.parent)
>                kobject_del(&entry->kobj);
>        kobject_put(&entry->kobj);

It was fixed this way in 424eb39 ("PCI: msi: fix imbalanced refcount of msi
irq sysfs objects"). kobject_put() still failed in case it wasn't
registered with sysfs earlier. populate_msi_sysfs() creates an kobject for
each entry in msi_list, and we have no idea (on fallback) up to which entry
was it already registered, on which it failed and which entries are still
not kobject_init_and_add()ed.

>
>(btw, looking at the sysfs code, this looks *very* suspicious in
>sysfs_remove_dir():
>
>        struct sysfs_dirent *sd = kobj->sd;
>
>        spin_lock(&sysfs_assoc_lock);
>        kobj->sd = NULL;
>        spin_unlock(&sysfs_assoc_lock);
>
>and I would suggest that "sd = kobj->sd" should be done under the
>lock, because otherwise the lock is kind of pointless..)
>
>Greg?
>
>                Linus