Re: [PATCH v2] f2fs: get rid of buffer_head use

From: Chao Yu
Date: Fri Jun 07 2024 - 10:11:04 EST


On 2024/6/7 21:33, Matthew Wilcox wrote:
On Fri, Jun 07, 2024 at 06:18:29PM +0800, Chao Yu wrote:
@@ -1990,6 +1989,12 @@ static inline struct f2fs_super_block *F2FS_RAW_SUPER(struct f2fs_sb_info *sbi)
return (struct f2fs_super_block *)(sbi->raw_super);
}
+static inline struct f2fs_super_block *F2FS_SUPER_BLOCK(struct folio *folio)
+{
+ return (struct f2fs_super_block *)(page_address(folio_page(folio, 0)) +
+ F2FS_SUPER_OFFSET);
+}

This assumes that the superblock is in the first page of the folio.
That's not necessarily guaranteed; let's say you have a 64KiB folio
that covers the start of the bdev.

Oh, I missed to add large folio support in this version.

For the case: page size is 4KiB, and folio size is 64KiB,
read_mapping_folio(mapping, 0, NULL) and read_mapping_folio(mapping,
1, NULL) will return the same #0 folio, right?


I don't quite know how to write this because f2fs defines its block size
in terms of PAGE_SIZE, which just seems like nonsense to me. If you
format a filesystem on a 16KiB PAGE_SIZE machine and then try to mount
it on a machine with a 4KiB PAGE_SIZE, it's going to go horribly wrong.

f2fs will check on-disk block size w/ PAGE_SIZE in sanity_check_raw_super()
as below:

/* only support block_size equals to PAGE_SIZE */
if (le32_to_cpu(raw_super->log_blocksize) != F2FS_BLKSIZE_BITS) {
f2fs_info(sbi, "Invalid log_blocksize (%u), supports only %u",
le32_to_cpu(raw_super->log_blocksize),
F2FS_BLKSIZE_BITS);
return -EFSCORRUPTED;
}


You'd need to pass in something that indicates whether you're trying to
access the first or second superblock; there's no way to tell from the
folio which one it is.

+static int __f2fs_commit_super(struct f2fs_sb_info *sbi, struct folio *folio,
+ bool update)
{
- lock_buffer(bh);
- if (super)
- memcpy(bh->b_data + F2FS_SUPER_OFFSET, super, sizeof(*super));
- set_buffer_dirty(bh);
- unlock_buffer(bh);
-
+ struct bio *bio;
/* it's rare case, we can do fua all the time */
- return __sync_dirty_buffer(bh, REQ_SYNC | REQ_PREFLUSH | REQ_FUA);
+ blk_opf_t opf = REQ_OP_WRITE | REQ_SYNC | REQ_PREFLUSH | REQ_FUA;
+ int ret;
+
+ folio_lock(folio);
+ folio_wait_writeback(folio);
+ if (update)
+ memcpy(F2FS_SUPER_BLOCK(folio), F2FS_RAW_SUPER(sbi),
+ sizeof(struct f2fs_super_block));
+ folio_mark_dirty(folio);
+ folio_clear_dirty_for_io(folio);
+ folio_start_writeback(folio);
+ folio_unlock(folio);
+
+ bio = bio_alloc(sbi->sb->s_bdev, 1, opf, GFP_NOFS);
+
+ /* it doesn't need to set crypto context for superblock update */
+ bio->bi_iter.bi_sector = SECTOR_FROM_BLOCK(folio_index(folio));
+
+ if (!bio_add_folio(bio, folio, PAGE_SIZE, 0))
+ f2fs_bug_on(sbi, 1);

Better make that folio_size(folio) to support bs>PS.

Looks better, will update.

Thanks,