Re: [RFC PATCH 2/8] mm: Add PG_atomic

[Dave Chinner <david@fromorbit.com>]

On Tue, Nov 18, 2025 at 09:47:42PM +0530, Ritesh Harjani wrote:
> Matthew Wilcox <willy@infradead.org> writes:
> 
> > On Fri, Nov 14, 2025 at 10:30:09AM +0530, Ritesh Harjani wrote:
> >> Matthew Wilcox <willy@infradead.org> writes:
> >> 
> >> > On Wed, Nov 12, 2025 at 04:36:05PM +0530, Ojaswin Mujoo wrote:
> >> >> From: John Garry <john.g.garry@oracle.com>
> >> >> 
> >> >> Add page flag PG_atomic, meaning that a folio needs to be written back
> >> >> atomically. This will be used by for handling RWF_ATOMIC buffered IO
> >> >> in upcoming patches.
> >> >
> >> > Page flags are a precious resource.  I'm not thrilled about allocating one
> >> > to this rather niche usecase.  Wouldn't this be more aptly a flag on the
> >> > address_space rather than the folio?  ie if we're doing this kind of write
> >> > to a file, aren't most/all of the writes to the file going to be atomic?
> >> 
> >> As of today the atomic writes functionality works on the per-write
> >> basis (given it's a per-write characteristic). 
> >> 
> >> So, we can have two types of dirty folios sitting in the page cache of
> >> an inode. Ones which were done using atomic buffered I/O flag
> >> (RWF_ATOMIC) and the other ones which were non-atomic writes. Hence a
> >> need of a folio flag to distinguish between the two writes.
> >
> > I know, but is this useful?  AFAIK, the files where Postgres wants to
> > use this functionality are the log files, and all writes to the log
> > files will want to use the atomic functionality.  What's the usecase
> > for "I want to mix atomic and non-atomic buffered writes to this file"?
> 
> Actually this goes back to the design of how we added support of atomic
> writes during DIO. So during the initial design phase we decided that
> this need not be a per-inode attribute or an open flag, but this is a
> per write I/O characteristic.
> 
> So as per the current design, we don't have any open flag or a
> persistent inode attribute which says kernel should permit _only_ atomic
> writes I/O to this file. Instead what we support today is DIO atomic
> writes using RWF_ATOMIC flag in pwritev2 syscall.

Which, if we can't do with REQ_ATOMIC IO, we fall back to the
filesystem COW IO path to provide RWF_ATOMIC semantics without
needing to involve the page cache.

IOWs, DIO REQ_ATOMIC writes are simply a fast path for the atomic
COW IO path inherent in COW-capable filesystems.

This is no different for buffered RWF_ATOMIC writes. We need to
ingest the data into the page cache as a COW operation, then at
writeback time we optimise away the COW operations if REQ_ATOMIC IO
can be performed instead.

Using COW for buffered RWF_ATOMIC writes means don't need to involve
the page caceh at all - this can all be implemented at the
filesystem extent mapping and iomap layers....

> Having said that there can be several policy decision that could still be
> discussed e.g. make sure any previous dirty data is flushed to disk when a
> buffered atomic write request is made to an inode. 

We don't need to care about mixed dirty non-atomic/atomic data on the
same file if REQ_ATOMIC is used as an optimisation for COW-based
atomic IO.  Filesystems like XFS naturally separate COW and non-COW
extents. If we combine non-atomic and atomic data into a single
atomic update at writeback(be it COW or REQ_ATOMIC IO), then we
have still honoured the requested atomic semantics required to
persist the data. It just doesn't matter.

IMO, trying to hack atomic physical IO semantics through the page
cache creates all sorts of issues that simply don't exist when we
use the atomic overwrite paths present in modern COW capable
filesystems....

-Dave.
-- 
Dave Chinner
david@fromorbit.com