Re: Dirty bits and sync writes

[Matthew Wilcox <willy@infradead.org>]

On Mon, Aug 09, 2021 at 03:48:56PM +0100, Christoph Hellwig wrote:
> On Tue, Aug 03, 2021 at 04:28:14PM +0100, Matthew Wilcox wrote:
> > Solution 1: Add an array of dirty bits to the iomap_page
> > data structure.  This patch already exists; would need
> > to be adjusted slightly to apply to the current tree.
> > https://lore.kernel.org/linux-xfs/7fb4bb5a-adc7-5914-3aae-179dd8f3adb1@huawei.com/
> 
> > Solution 2a: Replace the array of uptodate bits with an array of
> > dirty bits.  It is not often useful to know which parts of the page are
> > uptodate; usually the entire page is uptodate.  We can actually use the
> > dirty bits for the same purpose as uptodate bits; if a block is dirty, it
> > is definitely uptodate.  If a block is !dirty, and the page is !uptodate,
> > the block may or may not be uptodate, but it can be safely re-read from
> > storage without losing any data.
> 
> 1 or 2a seems like something we should do once we have lage folio
> support.
> 
> 
> > Solution 2b: Lose the concept of partially uptodate pages.  If we're
> > going to write to a partial page, just bring the entire page uptodate
> > first, then write to it.  It's not clear to me that partially-uptodate
> > pages are really useful.  I don't know of any network filesystems that
> > support partially-uptodate pages, for example.  It seems to have been
> > something we did for buffer_head based filesystems "because we could"
> > rather than finding a workload that actually cares.
> 
> The uptodate bit is important for the use case of a smaller than page
> size buffered write into a page that hasn't been read in already, which
> is fairly common for things like log writes.  So I'd hate to lose this
> optimization.
> 
> > (it occurs to me that solution 3 actually allows us to do IOs at storage
> > block size instead of filesystem block size, potentially reducing write
> > amplification even more, although we will need to be a bit careful if
> > we're doing a CoW.)
> 
> number 3 might be nice optimization.  The even better version would
> be a disk format change to just log those updates in the log and
> otherwise use the normal dirty mechanism.  I once had a crude prototype
> for that.

That's a bit beyond my scope at this point.  I'm currently working on
write-through.  Once I have that working, I think the next step is:

 - Replace the ->uptodate array with a ->dirty array
 - If the entire page is Uptodate, drop the iomap_page.  That means that
   writebacks will write back the entire folio, not just the dirty
   pieces.
 - If doing a partial page write
   - If the write is block-aligned (offset & length), leave the page
     !Uptodate and mark the dirty blocks
   - Otherwise bring the entire page Uptodate first, then mark it dirty

To take an example of a 512-byte block size file accepting a 520 byte
write at offset 500, we currently submit two reads, one for bytes 0-511
and the second for 1024-1535.  We're better off submitting a read for
bytes 0-4095 and then overwriting the entire thing.

But it's still better to do no reads at all if someone submits a write
for bytes 512-1023, or 512-N where N is past EOF.  And I'd preserve
that behaviour.