on 9/19/2023 9:12 PM, Bernd Schubert wrote:
>
>
> On 9/19/23 08:11, Kemeng Shi wrote:
>>
>>
>> on 9/16/2023 7:06 PM, Bernd Schubert wrote:
>>>
>>>
>>> On 9/14/23 17:45, Kemeng Shi wrote:
>>>> Commit 670d21c6e17f6 ("fuse: remove reliance on bdi congestion") change how
>>>> congestion_threshold is used and lock in
>>>> fuse_conn_congestion_threshold_write is not needed anymore.
>>>> 1. Access to supe_block is removed along with removing of bdi congestion.
>>>> Then down_read(&fc->killsb) which protecting access to super_block is no
>>>> needed.
>>>> 2. Compare num_background and congestion_threshold without holding
>>>> bg_lock. Then there is no need to hold bg_lock to update
>>>> congestion_threshold.
>>>>
>>>> Signed-off-by: Kemeng Shi <shikemeng@xxxxxxxxxxxxxxx>
>>>> ---
>>>> fs/fuse/control.c | 4 ----
>>>> 1 file changed, 4 deletions(-)
>>>>
>>>> diff --git a/fs/fuse/control.c b/fs/fuse/control.c
>>>> index 247ef4f76761..c5d7bf80efed 100644
>>>> --- a/fs/fuse/control.c
>>>> +++ b/fs/fuse/control.c
>>>> @@ -174,11 +174,7 @@ static ssize_t fuse_conn_congestion_threshold_write(struct file *file,
>>>> if (!fc)
>>>> goto out;
>>>> - down_read(&fc->killsb);
>>>> - spin_lock(&fc->bg_lock);
>>>> fc->congestion_threshold = val;
>>>> - spin_unlock(&fc->bg_lock);
>>>> - up_read(&fc->killsb);
>>>> fuse_conn_put(fc);
>>>> out:
>>>> return ret;
>>>
>>> Yeah, I don't see readers holding any of these locks.
>>> I just wonder if it wouldn't be better to use WRITE_ONCE to ensure a single atomic operation to store the value.
>> Sure, WRITE_ONCE looks better. I wonder if we should use READ_ONCE from reader.
>> Would like to get any advice. Thanks!
>
Sorry for the dealy - it toke me some time to go through the barrier documents.
> I'm not entirely sure either, but I _think_ the compiler is free to store a 32 bit value with multiple operations (like 2 x 16 bit). In that case a competing reading thread might read garbage...
> Although I don't see this documented here
> https://www.kernel.org/doc/Documentation/memory-barriers.txt
I found this is documented in section
"(*) For aligned memory locations whose size allows them to be accessed..."
Then WRITE_ONCE is absolutely needed now as you menthioned before.
> Though documented there is that the compile is free to optimize out the storage at all, see
> "(*) Similarly, the compiler is within its rights to omit a store entirely"
>
>
> Regarding READ_ONCE, I don't have a strong opinion, if the compiler makes some optimizations and the value would be wrong for a few cycles, would that matter for that variable? Unless the compiler would be really creative and the variable would get never updated... For sure READ_ONCE would be safer, but I don't know if it is needed
> SSee section
> "The compiler is within its rights to omit a load entirely if it know"
> in the document above.
I go through all examples of optimizations in document and congestion_threshold
has no same trouble descripted in document.
For specifc case you menthioned that "The compiler is within its rights to omit
a load entirely if it know". The compiler will keep the first load and only omit
successive loads from same variable in loop. As there is no repeat loading from
congestion_threshold in loop, congestion_threshold is out of this trouble.
Anyway, congestion_threshold is more like a hint and the worst case is that it is
stale for a few cycles. I prefer to keep reading congestion_threshold without
READ_ONCE and will do it in next version if it's fine to you. Thanks!
>
> Thanks,
> Bernd
>
>
>
>
