Re: [PATCH 3/6] direct-io: add support for write stream IDs

From: Jens Axboe
Date: Fri Apr 17 2015 - 22:01:06 EST


On 04/17/2015 05:51 PM, Dave Chinner wrote:
On Fri, Apr 17, 2015 at 05:11:40PM -0600, Jens Axboe wrote:
On 04/17/2015 05:06 PM, Dave Chinner wrote:
On Thu, Apr 16, 2015 at 11:20:45PM -0700, Ming Lin wrote:
On Sat, Apr 11, 2015 at 4:59 AM, Dave Chinner <david@xxxxxxxxxxxxx> wrote:
On Fri, Apr 10, 2015 at 04:50:05PM -0700, Ming Lin wrote:
On Wed, Mar 25, 2015 at 7:26 AM, Jens Axboe <axboe@xxxxxxxxx> wrote:
If iocb->ki_filp->f_streamid is not set, then it should fall back to
whatever is set on the inode->i_streamid.

Why should do the fall back?

Because then you have a method of using streams with applications
that aren't aware of streams.

Or perhaps you have a file you know has different access patterns to
the rest of the files in a directory, and you don't want to have to
set the stream on every process that opens and uses that file. e.g.
database writeahead log files (sequential write, never read) vs
database index/table files (random read/write).....

Good point, agree. Will make that change.

That change causes problem for direct IO, for example

process 1:
fd = open("/dev/nvme0n1", O_DIRECT...);
//set stream_id 1
fadvise(fd, 1, 0, POSIX_FADV_STREAMID);
pwrite(fd, ....);

process 2:
fd = open("/dev/nvme0n1", O_DIRECT...);
//should be legacy stream_id 0
pwrite(fd, ....);

But now process 2 also see stream_id 1, which is wrong.

It's not wrong, your behaviour model is just different You have
defined a process/fd based stream model and not considered
considered that admins and applications might want to use a file
based stream model instead, so applications don't need to even be
aware that write streams are in use...

The stream must be opened, otherwise device will return error if application
write to a not-opened stream.

That's an extremely device specific *implementation* of a write
stream. The *concept* of a write stream being passed from userspace to
the block layer doesn't have such constraints, and I get realy
concerned when implementations of a generic concept are so tightly
focussed around one type of hardware implementation of the
concept...

Indeed, which is why the implementation posted cares ONLY about the
stream ID itself, and passing that through.

But the point about fallback is valid, however, for some use cases
that will not be what you want. But we have to make some sort of
decision, and falling back to the inode set value (if one is set) is
probably the right thing to do in most use cases.

Right, the question is then whether fadvise should set the value on
the inode at all, because then the effect of setting it on a fd also
changes the fallback. Perhaps we need to a distinction between
"setting the stream for this fd" which lasts as long as the fd is
active, and "setting the default inode stream" which is potentially
a persistent operation if the filesystem stores it on disk...

Yes, that might be a good compromise. The easiest would be to define a second fadvise advice, where the stronger advice would be file + inode. Another option would be changing the file approach to use fcntl(), and keeping the fadvise for the inode. I'll be happy to take input on what people would prefer here.

Device has limited number of streams, for example, 16 streams.
There are 2 APIs to open/close the stream.

What's to stop me writing something for DM-thinp that understands
write streams in bios and uses it to separate out the write streams
into different regions of the thinp device to improve locality of
it's data placement and hence reduce fragmentation?

Absolutely nothing, in fact that's one of the use cases that I had
in mind. Or for for caching software.

*nod*. We are on the same page, then :)

Yes completely, basically just wanted to clarify that.

--
Jens Axboe

--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/