Improve UBIFS nospc_retries
hujianyang
hujianyang at huawei.com
Sun Nov 9 17:32:30 PST 2014
On 2014/11/7 19:41, Artem Bityutskiy wrote:
> On Fri, 2014-11-07 at 18:27 +0800, hujianyang wrote:
>> Hi Artem,
>>
>> I'm puzzling with *nospc_retries* in make_reservation() nowadays.
>> My colleagues in testing department use a less than 20M flash
>> and run lots of processes on it. These processes will read, write,
>> delete from flash and the flash is always in a nearly-full state.
>> The board only has one core.
>
> If it switches to R/O mode, this means that we did not reserve enough
> space for the operation. Probably we need to reserve more LEBs for
> deletions.
>
>> So, the *nospc_retries* in make_reservation(), line 341 in journal.c
>> will easily reach 2 as we set and turn the filesystem to RO mode.
>>
>> I know we can't perform an infinite loop here. Can we improve it?
>> Not only just turn current *2* to some larger number but also add
>> some useful mechanism to avoid filesystem turning to RO mode.
>
> Did you try to increase it, does it help?
>
Yes, I've increased it to 5, and see it takes a longer time to
switch to RO. I add some log messages when nospc_retries increase
to 2 and filesystem will work OK when nospc_retries hits 2. But it
will turn to RO when nospc_retries equals to 5.
> We did not stress-test it for small flash size, because we assumed that
> JFFS2 would take care of those.
>
> And yes, there may be issues in the "little space left" handling. We
> tested that, but no too extensively. The I/O becomes extremely slow when
> there is little space, so we preferred to make sure the FS does not get
> too full at all by reserving more space for 'root' (UBIFS feature, which
> makes sure that 'root' can always write even if users consumed all the
> space).
I think you are right. Runing UBIFS on a small flash device will
waste high percents of space.
>
> You could also experiment by reserving more LEBs for deletions, and see
> if it helps. Just increase 'UBIFS_MIN_MAIN_LEBS' in ubifs.h and run your
> test.
>
Thanks. I think it will help~! I will try to increase this.
Hu
>
>
> .
>
More information about the linux-mtd
mailing list