[PATCH 5/5] nvme: enable logical block size > PAGE_SIZE

[Luis Chamberlain]

On Sat, May 11, 2024 at 04:51:32PM -0700, Luis Chamberlain wrote:
> On Sun, May 12, 2024 at 12:30:40AM +0100, Matthew Wilcox wrote:
> > On Sat, May 11, 2024 at 04:09:48PM -0700, Luis Chamberlain wrote:
> > > We can't just do this, we need to consider the actual nvme cap (test it,
> > > and if it crashes and below what the page cache supports, then we have
> > > to go below) and so to make the enablment easier. So we could just move
> > > this to helper [0]. Then when the bdev cache patch goes through the
> > > check for CONFIG_BUFFER_HEAD can be removed, if this goes first.
> > > 
> > > We crash if we go above 1 MiB today, we should be able to go up to 2
> > > MiB but that requires some review to see what stupid thing is getting
> > > in the way.
> > > 
> > > [0] https://git.kernel.org/pub/scm/linux/kernel/git/mcgrof/linux.git/commit/?h=20240408-lbs-scsi-kludge&id=1f7f4dce548cc11872e977939a872b107c68ad53
> > 
> > This is overengineered garbage.  What's the crash?
> 
> I had only tested it with iomap, I had not tested it with buffer-heads,
> and so it would require re-testing. It's Saturday 5pm, I should be doing
> something else other than being on my computer.

It crashes because we forgot two things in this series below, so the
change below its us to enable to *at least* boot up to 64k LBA format on
NVMe.

One can reproduce this with kdevops with:

make defconfig-lbs-xfs-bdev-nvme
make bringup
make linux

I've added another defconfig which bumps the LBA format up to 512 KiB to
see if bootup blows up, that has another defconfig:

make lbs-xfs-bdev-large-nvme
make bringup
make linux

That at least booted. Note that the above defconfigs use this thread's
message ID, so it applies this series on top of the min order branch.
The patch below is just needed.

I'll try next going above 512 KiB.
 
diff --git a/fs/buffer.c b/fs/buffer.c 
index 4f73d23c2c46..fa88e300a946 100644
--- a/fs/buffer.c
+++ b/fs/buffer.c
@@ -2360,8 +2360,6 @@ int block_read_full_folio(struct folio *folio, get_block_t *get_block)
 	if (IS_ENABLED(CONFIG_FS_VERITY) && IS_VERITY(inode))
 		limit = inode->i_sb->s_maxbytes;
 
-	VM_BUG_ON_FOLIO(folio_test_large(folio), folio);
-
 	head = folio_create_buffers(folio, inode, 0);
 	blocksize = head->b_size;
 
diff --git a/fs/mpage.c b/fs/mpage.c
index e3732686e65f..e124c924b2e7 100644
--- a/fs/mpage.c
+++ b/fs/mpage.c
@@ -178,7 +178,6 @@ static struct bio *do_mpage_readpage(struct mpage_readpage_args *args)
 	gfp_t gfp = mapping_gfp_constraint(folio->mapping, GFP_KERNEL);
 
 	/* MAX_BUF_PER_PAGE, for example */
-	VM_BUG_ON_FOLIO(folio_test_large(folio), folio);
 
 	if (args->is_readahead) {
 		opf |= REQ_RAHEAD;