Re: [PATCH 0/5] Enable per-file/directory DAX operations

[Boaz Harrosh]

On 22/10/2019 14:21, Boaz Harrosh wrote:
> On 20/10/2019 18:59, ira.weiny@xxxxxxxxx wrote:
<>
>> At LSF/MM we discussed the difficulties of switching the mode of a file with
>> active mappings / page cache. Rather than solve those races the decision was to
>> just limit mode flips to 0-length files.
>>
> 
> What I understand above is that only "writers" before writing any bytes may
> turn the flag on, which then persists. But as a very long time user of DAX, usually
> it is the writers that are least interesting. With lots of DAX technologies and
> emulations the write is slower and needs slow "flushing".
> 
> The more interesting and performance gains comes from DAX READs actually.
> specially cross the VM guest. (IE. All VMs share host memory or pmem)
> 
> This fixture as I understand it, that I need to know before I write if I will
> want DAX or not and then the write is DAX as well as reads after that, looks
> not very interesting for me as a user.
> 
> Just my $0.17
> Boaz
> 

I want to say one more thing about this patchset please. I was sleeping on it
and I think it is completely wrong approach with a completely wrong API.

The DAX or not DAX is a matter of transport. and is nothing to do with data
content and persistency. It's like connecting a disk behind, say, a scsi bus and then
take the same DB and putting it behind a multy-path or an NFS server. It is always
the same data.
(Same mistake we did with ndctl which is cry for generations)

For me the DAX or NO-DAX is an application thing and not a data thing.

The right API is perhaps an open O_FLAG where if you are not the first opener
the open returns EBUSY and then the app can decide if to open without the
flag or complain and fail. (Or apply open locks)

If you are a second opener when the file is already opened O_DAX you are
silently running DAX. If you are 2nd open(O_DAX) when already oppened
O_DAX then off course you succeed.
(Also the directory inheritance thing is all mute too)

Please explain the use case behind your model?

Thanks
Boaz