On 05/28, Christoph Hellwig wrote:

On Wed, May 27, 2026 at 03:59:35PM +0000, Jaegeuk Kim wrote:

quoted

F2FS merges bios before submit_bio, regardless of small or large folios,
since the block addresses are consecutive. So, I think IO subsystem was
working in full speed.

As does every other remotely modern file system.  But that merging is
surprisingly expensive, which is why using folios gets really major
performance improvements.

For one doing these checks to merge touch quite a few cache lines.
Second, devices are often a lot more efficient if they see fewer SGL
entries.  I.e. having a 1MB bio a single SGL tends to work better than
having 256 of them.
The same is true in the kernel code itself, both in the submission path
(dma mapping and co), and even more so in the page cache handling
both before submitting and in the completion path.

See Bart's patch about how long the walk of the bio_vecs in the f2fs
completion path can take.  We had similar issues in XFS even in the
workqueue completion path due to lack of rescheduling, and these simply
go away when you do the folio manipulation in larger chunks (LAZY_PREEMPT
would avoid the need to explicit rescheduling these days, but that just
papers over the symptoms in this case).

I see. That's also super helpful. Let me kick off the large folio support asap.
Thanks.