Re: [f2fs-dev] [PATCH v3] f2fs: change the current atomic write way

[Jaegeuk Kim <jaegeuk@kernel.org>]

On 05/24, Chao Yu wrote:
> On 2022/5/24 2:03, Jaegeuk Kim wrote:
> > On 05/22, Chao Yu wrote:
> > > On 2022/4/29 2:18, Daeho Jeong wrote:> +		*old_addr = dn.data_blkaddr;
> > > > +		f2fs_truncate_data_blocks_range(&dn, 1);
> > > > +		dec_valid_block_count(sbi, F2FS_I(inode)->cow_inode, count);
> > > > +		inc_valid_block_count(sbi, inode, &count);
> > > > +		f2fs_replace_block(sbi, &dn, dn.data_blkaddr, new_addr,
> > > > +					ni.version, true, false);
> > > 
> > > My concern is, if cow_inode's data was persisted into previous checkpoint,
> > > and then f2fs_replace_block() will update SSA from cow_inode to inode?
> > 
> > SSA for original file is intact, so we'll see the orignal file's block addresses
> > and SSA, if we flush cow_inode's SSA after committing the atomic writes?
> > It'd be good to flush any SSA for cow_inode, since we'll truncate
> > cow_inode after powercut by the ohphan recovery?
> 
> I think it's safe for recovery flow, but before that, fsck will report inconsistent
> status during checking orphan atomic_write inode.

That should be fine as well, since it'll just drop that cow_inode.

> 
> Thanks,
> 
> > 
> > > it will cause inconsistent status of last valid checkpoint? Or am I mssing
> > > something?
> > > 
> > > > -		f2fs_submit_merged_write_cond(sbi, inode, NULL, 0, DATA);
> > > > +			new = f2fs_kmem_cache_alloc(revoke_entry_slab, GFP_NOFS,
> > > > +							true, NULL);
> > > > +			if (!new) {
> > > > +				f2fs_put_dnode(&dn);
> > > > +				ret = -ENOMEM;
> > > > +				goto out;
> > > 
> > > It doesn't need to handle failure of f2fs_kmem_cache_alloc()
> > > due to nofail parameter is true.

Let's get this by another patch. Could you please send one?

> > > 
> > > Thanks,
> > > 
> > > 
> > > _______________________________________________
> > > Linux-f2fs-devel mailing list
> > > Linux-f2fs-devel@lists.sourceforge.net
> > > https://lists.sourceforge.net/lists/listinfo/linux-f2fs-devel