Re: [PATCH] xen/scsiback: don't call scsiback_free_translation_entry() under lock

[Oleksandr Tyshchenko]

On 29.03.23 09:20, Juergen Gross wrote:

Hello Juergen


> On 28.03.23 17:47, Oleksandr Tyshchenko wrote:
>>
>>
>> On 28.03.23 11:46, Juergen Gross wrote:
>>
>> Hello Juergen
>>
>>> scsiback_free_translation_entry() shouldn't be called under spinlock,
>>> as it can sleep.
>>>
>>> This requires to split removing a translation entry from the v2p list
>>> from actually calling kref_put() for the entry.
>>>
>>> Reported-by: Dan Carpenter <error27@xxxxxxxxx>
>>> Link: 
>>> https://urldefense.com/v3/__https://lore.kernel.org/lkml/Y*JUIl64UDmdkboh@kadam/__;Kw!!GF_29dbcQIUBPA!23IKdVhamoFq8ptUnprd_TubDMObj-0QAalsGiffBHCeEdOuwrq7z4ohg92Sj0olgl0nh73oXvSr-i1zqXhY$ [lore[.]kernel[.]org]
>>> Signed-off-by: Juergen Gross <jgross@xxxxxxxx>
>>> ---
>>>    drivers/xen/xen-scsiback.c | 27 ++++++++++++++-------------
>>>    1 file changed, 14 insertions(+), 13 deletions(-)
>>>
>>> diff --git a/drivers/xen/xen-scsiback.c b/drivers/xen/xen-scsiback.c
>>> index 954188b0b858..294f29cdc7aa 100644
>>> --- a/drivers/xen/xen-scsiback.c
>>> +++ b/drivers/xen/xen-scsiback.c
>>> @@ -1010,12 +1010,6 @@ static int 
>>> scsiback_add_translation_entry(struct vscsibk_info *info,
>>>        return err;
>>>    }
>>> -static void __scsiback_del_translation_entry(struct v2p_entry *entry)
>>> -{
>>> -    list_del(&entry->l);
>>> -    kref_put(&entry->kref, scsiback_free_translation_entry);
>>> -}
>>> -
>>>    /*
>>>      Delete the translation entry specified
>>>    */
>>> @@ -1024,18 +1018,20 @@ static int 
>>> scsiback_del_translation_entry(struct vscsibk_info *info,
>>>    {
>>>        struct v2p_entry *entry;
>>>        unsigned long flags;
>>> -    int ret = 0;
>>>        spin_lock_irqsave(&info->v2p_lock, flags);
>>>        /* Find out the translation entry specified */
>>>        entry = scsiback_chk_translation_entry(info, v);
>>>        if (entry)
>>> -        __scsiback_del_translation_entry(entry);
>>> -    else
>>> -        ret = -ENOENT;
>>> +        list_del(&entry->l);
>>>        spin_unlock_irqrestore(&info->v2p_lock, flags);
>>> -    return ret;
>>> +
>>> +    if (!entry)
>>> +        return -ENOENT;
>>> +
>>> +    kref_put(&entry->kref, scsiback_free_translation_entry);
>>> +    return 0;
>>>    }
>>>    static void scsiback_do_add_lun(struct vscsibk_info *info, const 
>>> char *state,
>>> @@ -1239,14 +1235,19 @@ static void 
>>> scsiback_release_translation_entry(struct vscsibk_info *info)
>>>    {
>>>        struct v2p_entry *entry, *tmp;
>>>        struct list_head *head = &(info->v2p_entry_lists);
>>> +    struct list_head tmp_list;
>>
>>
>> I would use LIST_HEAD(tmp_list);
> 
> There is no need to initialize it, so I think I will keep it as is.
> 
>>
>>>        unsigned long flags;
>>>        spin_lock_irqsave(&info->v2p_lock, flags);
>>> -    list_for_each_entry_safe(entry, tmp, head, l)
>>> -        __scsiback_del_translation_entry(entry);
>>> +    list_cut_before(&tmp_list, head, head);
>>
>> so we just move all entries from head to tmp_list here to be processed...
> 
> Correct.
> 
>>
>>>        spin_unlock_irqrestore(&info->v2p_lock, flags);
>>
>> ... when the lock is not held, ok
>>
>> Patch LGTM, but one (maybe stupid) question to clarify.
>>
>> Why do we need to use a lock here in the first place? The
>> scsiback_release_translation_entry() gets called when the driver
>> instance is about to be removed and *after* the disconnection from
>> otherend (so no requests are expected), so what else might cause this
>> list to be accessed concurrently?
> 
> Maybe nothing, but I think it is good practice to keep the lock in order
> to avoid future code changes to cause problems.


Thanks for the explanation, it sounds reasonable to me.

Reviewed-by: Oleksandr Tyshchenko <oleksandr_tyshchenko@xxxxxxxx>

> 
> 
> Juergen
>