[PATCH v6 00/10] block atomic writes

[Matthew Wilcox]

On Wed, Mar 27, 2024 at 01:37:41PM +0000, John Garry wrote:
> On 27/03/2024 03:50, Matthew Wilcox wrote:
> > On Tue, Mar 26, 2024 at 01:38:03PM +0000, John Garry wrote:
> > > The goal here is to provide an interface that allows applications use
> > > application-specific block sizes larger than logical block size
> > > reported by the storage device or larger than filesystem block size as
> > > reported by stat().
> > > 
> > > With this new interface, application blocks will never be torn or
> > > fractured when written. For a power fail, for each individual application
> > > block, all or none of the data to be written. A racing atomic write and
> > > read will mean that the read sees all the old data or all the new data,
> > > but never a mix of old and new.
> > > 
> > > Three new fields are added to struct statx - atomic_write_unit_min,
> > > atomic_write_unit_max, and atomic_write_segments_max. For each atomic
> > > individual write, the total length of a write must be a between
> > > atomic_write_unit_min and atomic_write_unit_max, inclusive, and a
> > > power-of-2. The write must also be at a natural offset in the file
> > > wrt the write length. For pwritev2, iovcnt is limited by
> > > atomic_write_segments_max.
> > > 
> > > There has been some discussion on supporting buffered IO and whether the
> > > API is suitable, like:
> > > https://lore.kernel.org/linux-nvme/ZeembVG-ygFal6Eb@casper.infradead.org/
> > > 
> > > Specifically the concern is that supporting a range of sizes of atomic IO
> > > in the pagecache is complex to support. For this, my idea is that FSes can
> > > fix atomic_write_unit_min and atomic_write_unit_max at the same size, the
> > > extent alignment size, which should be easier to support. We may need to
> > > implement O_ATOMIC to avoid mixing atomic and non-atomic IOs for this. I
> > > have no proposed solution for atomic write buffered IO for bdev file
> > > operations, but I know of no requirement for this.
> > 
> > The thing is that there's no requirement for an interface as complex as
> > the one you're proposing here.  I've talked to a few database people
> > and all they want is to increase the untorn write boundary from "one
> > disc block" to one database block, typically 8kB or 16kB.
> > 
> > So they would be quite happy with a much simpler interface where they
> > set the inode block size at inode creation time,
> 
> We want to support untorn writes for bdev file operations - how can we set
> the inode block size there? Currently it is based on logical block size.

ioctl(BLKBSZSET), I guess?  That currently limits to PAGE_SIZE, but I
think we can remove that limitation with the bs>PS patches.

> > and then all writes to
> > that inode were guaranteed to be untorn.  This would also be simpler to
> > implement for buffered writes.
> 
> We did consider that. Won't that lead to the possibility of breaking
> existing applications which want to do regular unaligned writes to these
> files? We do know that mysql/innodb does have some "compressed" mode of
> operation, which involves regular writes to the same file which wants untorn
> writes.

If you're talking about "regular unaligned buffered writes", then that
won't break.  If you cross a folio boundary, the result may be torn,
but if you're crossing a block boundary you expect that.

> Furthermore, untorn writes in HW are expensive - for SCSI anyway. Do we
> always want these for such a file?

Do untorn writes actually exist in SCSI?  I was under the impression
nobody had actually implemented them in SCSI hardware.

> We saw untorn writes as not being a property of the file or even the inode
> itself, but rather an attribute of the specific IO being issued from the
> userspace application.

The problem is that keeping track of that is expensive for buffered
writes.  It's a model that only works for direct IO.  Arguably we
could make it work for O_SYNC buffered IO, but that'll require some
surgery.

> > Who's asking for this more complex interface?
> 
> It's not a case of someone specifically asking for this interface. This is
> just a proposal to satisfy userspace requirement to do untorn writes in a
> generic way.
> 
> From a user point-of-view, untorn writes for a regular file can be enabled
> for up to a specific size* with FS_IOC_SETFLAGS API. Then they need to
> follow alignment and size rules for issuing untorn writes, but they would
> always need to do this. In addition, the user may still issue regular
> (tearable) writes to the file.
> 
> * I think that we could change this to only allow writes for that specific
> size, which was my proposal for buffered IO.
> 
> Thanks,
> John
>