[PATCH 4/6] UBI: Fastmap: Fix races in ubi_wl_get_peb()

[Richard Weinberger]

Tanya,

Am 05.12.2014 um 17:54 schrieb Tanya Brokhman:
> Hi Richard
> 
> On 12/5/2014 3:20 PM, Richard Weinberger wrote:
>> Tanya,
>>
>> Am 05.12.2014 um 14:09 schrieb Tanya Brokhman:
>>> On 11/24/2014 3:20 PM, Richard Weinberger wrote:
>>>> ubi_wl_get_peb() has two problems, it reads the pool
>>>> size and usage counters without any protection.
>>>> While reading one value would be perfectly fine it reads multiple
>>>> values and compares them. This is racy and can lead to incorrect
>>>> pool handling.
>>>> Furthermore ubi_update_fastmap() is called without wl_lock held,
>>>> before incrementing the used counter it needs to be checked again.
>>>
>>> I didn't see where you fixed the ubi_update_fastmap issue you just mentioned.
>>
>> This is exactly what you're questioning below.
>> We have to recheck as the pool counter could have changed.
>>
> 
> Oh, I understood the commit msg a bit differently, but now I see that it was my mistake. thanks!
> 
>>>> It could happen that another thread consumed all PEBs from the
>>>> pool and the counter goes beyond ->size.
>>>>
>>>> Signed-off-by: Richard Weinberger <richard at nod.at>
>>>> ---
>>>>    drivers/mtd/ubi/ubi.h |  3 ++-
>>>>    drivers/mtd/ubi/wl.c  | 34 +++++++++++++++++++++++-----------
>>>>    2 files changed, 25 insertions(+), 12 deletions(-)
>>>>
>>>> diff --git a/drivers/mtd/ubi/ubi.h b/drivers/mtd/ubi/ubi.h
>>>> index 04c4c05..d672412 100644
>>>> --- a/drivers/mtd/ubi/ubi.h
>>>> +++ b/drivers/mtd/ubi/ubi.h
>>>> @@ -439,7 +439,8 @@ struct ubi_debug_info {
>>>>     * @pq_head: protection queue head
>>>>     * @wl_lock: protects the @used, @free, @pq, @pq_head, @lookuptbl, @move_from,
>>>>     *         @move_to, @move_to_put @erase_pending, @wl_scheduled, @works,
>>>> - *         @erroneous, @erroneous_peb_count, and @fm_work_scheduled fields
>>>> + *         @erroneous, @erroneous_peb_count, @fm_work_scheduled, @fm_pool,
>>>> + *         and @fm_wl_pool fields
>>>>     * @move_mutex: serializes eraseblock moves
>>>>     * @work_sem: used to wait for all the scheduled works to finish and prevent
>>>>     * new works from being submitted
>>>> diff --git a/drivers/mtd/ubi/wl.c b/drivers/mtd/ubi/wl.c
>>>> index cb2e571..7730b97 100644
>>>> --- a/drivers/mtd/ubi/wl.c
>>>> +++ b/drivers/mtd/ubi/wl.c
>>>> @@ -629,24 +629,36 @@ void ubi_refill_pools(struct ubi_device *ubi)
>>>>     */
>>>>    int ubi_wl_get_peb(struct ubi_device *ubi)
>>>>    {
>>>> -    int ret;
>>>> +    int ret, retried = 0;
>>>>        struct ubi_fm_pool *pool = &ubi->fm_pool;
>>>>        struct ubi_fm_pool *wl_pool = &ubi->fm_wl_pool;
>>>>
>>>> -    if (!pool->size || !wl_pool->size || pool->used == pool->size ||
>>>> -        wl_pool->used == wl_pool->size)
>>>> +again:
>>>> +    spin_lock(&ubi->wl_lock);
>>>> +    /* We check here also for the WL pool because at this point we can
>>>> +     * refill the WL pool synchronous. */
>>>> +    if (pool->used == pool->size || wl_pool->used == wl_pool->size) {
>>>> +        spin_unlock(&ubi->wl_lock);
>>>>            ubi_update_fastmap(ubi);
>>>> -
>>>> -    /* we got not a single free PEB */
>>>> -    if (!pool->size)
>>>> -        ret = -ENOSPC;
>>>> -    else {
>>>>            spin_lock(&ubi->wl_lock);
>>>> -        ret = pool->pebs[pool->used++];
>>>> -        prot_queue_add(ubi, ubi->lookuptbl[ret]);
>>>> +    }
>>>> +
>>>> +    if (pool->used == pool->size) {
>>>
>>> Im confused about this "if" condition. You just tested pool->used == pool->size in the previous "if". If in the previous if pool->used != pool->size and wl_pool->used !=
>>> wl_pool->size, you didn't enter, the lock is still held so pool->used != pool->size still. If in the previos "if" wl_pool->used == wl_pool->size was true nd tou released the lock,
>>> ubi_update_fastmap(ubi) was called, which refills the pools. So again, if pools were refilled pool->used would be 0 here and pool->size > 0.
>>>
>>> So in both cases I don't see how at this point pool->used == pool->size could ever be true?
>>
>> If we enter the "if (pool->used == pool->size || wl_pool->used == wl_pool->size) {" branch we unlock wl_lock and call ubi_update_fastmap().
>> Another thread can enter ubi_wl_get_peb() and alter the pool counter. So we have to recheck the counter after taking wl_lock again.
> 
> hmmm... ok. Perhaps a comment could be added in the code to explain this case in a few words?
> 

Makes sense, I'll add a comment.

>>>>            spin_unlock(&ubi->wl_lock);
>>>> +        if (retried) {
>>>> +            ubi_err(ubi, "Unable to get a free PEB from user WL pool");
>>>> +            ret = -ENOSPC;
>>>> +            goto out;
>>>> +        }
>>>> +        retried = 1;
>>>
>>> Why did you decide to retry in this function? and why only 1 retry attempt? I'm not against it, trying to understand the logic.
>>
>> Because failing immediately with -ENOSPC is not nice.
> 
> Why not? this is what was done before....

The behavior from before was not good.
If we return here a -ENOSPC it is not because we ran out of free PEBs, it is because the pool contains
no free PEBs and needs refilling.
As between refilling the pool and requesting a fresh PEB from it another thread could "steal" all PEBs
we retry.

> I think what I really bothers me in this case is that you don't sleep, you branch immediately to retry again, so the chances that there will be context switch and free pebs appear
> aren't that high.
> I'm used to functions using some sort of "retry" logic to sleep before retrying. Of course sleeping isn't a good idea here. That's why the "retry" bugs me a bit.

You mean a cond_resched()?
This retry-logic is common pattern in UBI. For exmaple see ubi_wl_put_peb().

Thanks,
//richard