Re: [PATCH 3/5] vfs: add a zero-initialization mode to fallocate

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

 



On Tue, Sep 21, 2021 at 09:13:54PM -0700, Darrick J. Wong wrote:
> On Wed, Sep 22, 2021 at 01:59:07PM +1000, Dave Chinner wrote:
> > On Tue, Sep 21, 2021 at 07:38:01PM -0700, Darrick J. Wong wrote:
> > > On Tue, Sep 21, 2021 at 07:16:26PM -0700, Dan Williams wrote:
> > > > On Tue, Sep 21, 2021 at 1:32 AM Christoph Hellwig <hch@xxxxxxxxxxxxx> wrote:
> > > > >
> > > > > On Tue, Sep 21, 2021 at 10:44:31AM +1000, Dave Chinner wrote:
> > > > > > I think this wants to be a behavioural modifier for existing
> > > > > > operations rather than an operation unto itself. i.e. similar to how
> > > > > > KEEP_SIZE modifies ALLOC behaviour but doesn't fundamentally alter
> > > > > > the guarantees ALLOC provides userspace.
> > > > > >
> > > > > > In this case, the change of behaviour over ZERO_RANGE is that we
> > > > > > want physical zeros to be written instead of the filesystem
> > > > > > optimising away the physical zeros by manipulating the layout
> > > > > > of the file.
> > > > >
> > > > > Yes.
> > > > >
> > > > > > Then we have and API that looks like:
> > > > > >
> > > > > >       ALLOC           - allocate space efficiently
> > > > > >       ALLOC | INIT    - allocate space by writing zeros to it
> > > > > >       ZERO            - zero data and preallocate space efficiently
> > > > > >       ZERO | INIT     - zero range by writing zeros to it
> > > > > >
> > > > > > Which seems to cater for all the cases I know of where physically
> > > > > > writing zeros instead of allocating unwritten extents is the
> > > > > > preferred behaviour of fallocate()....
> > > > >
> > > > > Agreed.  I'm not sure INIT is really the right name, but I can't come
> > > > > up with a better idea offhand.
> > > > 
> > > > FUA? As in, this is a forced-unit-access zeroing all the way to media
> > > > bypassing any mechanisms to emulate zero-filled payloads on future
> > > > reads.
> > 
> > Yes, that's the semantic we want, but FUA already defines specific
> > data integrity behaviour in the storage stack w.r.t. volatile
> > caches.
> > 
> > Also, FUA is associated with devices - it's low level storage jargon
> > and so is not really appropriate to call a user interface operation
> > FUA where users have no idea what a "unit" or "access" actually
> > means.
> > 
> > Hence we should not overload this name with some other operation
> > that does not have (and should not have) explicit data integrity
> > requirements. That will just cause confusion for everyone.
> > 
> > > FALLOC_FL_ZERO_EXISTING, because you want to zero the storage that
> > > already exists at that file range?
> > 
> > IMO that doesn't work as a behavioural modifier for ALLOC because
> > the ALLOC semantics are explicitly "don't touch existing user
> > data"...
> 
> Well since you can't preallocate /and/ zerorange at the same time...
> 
> /* For FALLOC_FL_ZERO_RANGE, write zeroes to pre-existing mapped storage. */
> #define FALLOC_FL_ZERO_EXISTING		(0x80)

Except we also want the newly allocated regions (i.e. where holes
were) in that range being zeroed to have zeroes written to them as
well, yes? Otherwise we end up with a combination of unwritten
extents and physical zeroes, and you can't use
ZERORANGE|EXISTING as a replacement for PUNCH + ALLOC|INIT

/*
 * For preallocation and zeroing operations, force the filesystem to
 * write zeroes rather than use unwritten extents to indicate the
 * range contains zeroes.
 *
 * For filesystems that support unwritten extents, this trades off
 * slow fallocate performance for faster first write performance as
 * unwritten extent conversion on the first write to each block in
 * the range is not needed.
 *
 * Care is required when using FALLOC_FL_ALLOC_INIT_DATA as it will
 * be much slower overall for large ranges and/or slow storage
 * compared to using unwritten extents.
 */
#define FALLOC_FL_ALLOC_INIT_DATA	(1 << 7)

Cheers,

Dave.

-- 
Dave Chinner
david@xxxxxxxxxxxxx



[Index of Archives]     [XFS Filesystem Development (older mail)]     [Linux Filesystem Development]     [Linux Audio Users]     [Yosemite Trails]     [Linux Kernel]     [Linux RAID]     [Linux SCSI]


  Powered by Linux