really strange ext4 behavior

Ric Wheeler rwheeler at redhat.com
Tue Feb 15 17:13:28 UTC 2011


On 02/15/2011 09:05 AM, Gerd v. Egidy wrote:
>>>> (except for ext3, where, for data integrity with a volatile writeback
>>>> cache, defaults + barriers=1, since that safe default was never accepted
>>>> upstream)
>>> Why isn't it the Fedora default?
>> Excellent question - we probably should flip it over in fedora to the safe
>> default.
>>
>> Historically, there were times that applications got fsync happy (like
>> firefox) and basically stopped working when barriers were enabled (only
>> because they were waiting for the fsync to actually do something).
>>
>> I believe that we have exposed most of this kind of thing by now after
>> years of default ext4 use so this should be mostly resolved,
> weren't barriers kicked out with kernel 2.6.37?
>
> see
> http://kernelnewbies.org/Linux_2_6_37#head-9d59a04087857adeda842dd4ed97f00abd824e23
> and
> http://lwn.net/Articles/400541/
>
> Kind regards,
>
> Gerd

Short answer is no.

How barriers work has been redone, but the need for them is the same.

What Jon tried to convey was that the assumptions about who needed to drain 
queues (block or file system, etc) was clarified and some subsystems have been 
updated to avoid doing un-needed queue drains.

You absolutely, 100% for sure, will loose data if you have a volatile write 
cache storage device and run without barriers enabled when the power drops :)

ric



More information about the devel mailing list