Re: [PATCH block/for-next v2 01/16] block: add a new helper to get inode from block_device

From: Yu Kuai
Date: Mon Nov 27 2023 - 08:07:35 EST


Hi,

在 2023/11/27 15:21, Christoph Hellwig 写道:
On Mon, Nov 27, 2023 at 02:21:01PM +0800, Yu Kuai wrote:
From: Yu Kuai <yukuai3@xxxxxxxxxx>

block_devcie is allocated from bdev_alloc() by bdev_alloc_inode(), and
currently block_device contains a pointer that point to the address of
inode, while such inode is allocated together:

This is going the wrong way. Nothing outside of core block layer code
should ever directly use the bdev inode. We've been rather sloppy
and added a lot of direct reference to it, but they really need to
go away and be replaced with well defined high level operation on
struct block_device. Once that is done we can remove the bd_inode
pointer, but replacing it with something that pokes even more deeply
into bdev internals is a bad idea.

Thanks for the advice, however, after collecting how other modules are
using bdev inode, I got two main questions:

1) Is't okay to add a new helper to pass in bdev for following apis?
If so, then almost all the fs and driver can avoid to access bd_inode
dirctly.

errseq_check(&bdev->bd_inode->i_mapping->wb_err, wb_err);
errseq_check_and_advance(&bdev->bd_inode->i_mapping->wb_err, &wb_err);
mapping_gfp_constraint(bdev->bd_inode->i_mapping, gfp);
i_size_read(bdev->bd_inode)
find_get_page(bdev->bd_inode->i_mapping, offset);
find_or_create_page(bdev->bd_inode->i_mapping, index, gfp);
read_cache_page_gfp(bdev->bd_inode->i_mapping, index, gfp);
invalidate_inode_pages2(bdev->bd_inode->i_mapping);
invalidate_inode_pages2_range(bdev->bd_inode->i_mapping, start, end);
read_mapping_folio(bdev->bd_inode->i_mapping, index, file);
read_mapping_page(bdev->bd_inode->i_mapping, index, file);
balance_dirty_pages_ratelimited(bdev->bd_inode->i_mapping)
file_ra_state_init(ra, bdev->bd_inode->i_mapping);
page_cache_sync_readahead(bdev->bd_inode->i_mapping, ra, file, index, req_count);
inode_to_bdi(bdev->bd_inode)

2) For the file fs/buffer.c, there are some special usage like
following that I don't think it's good to add a helper:

spin_lock(&bd_inode->i_mapping->private_lock);

Is't okay to move following apis from fs/buffer.c directly to
block/bdev.c?

__find_get_block
bdev_getblk

Thanks,
Kuai

.