Re: [RFC PATCH 10/17] btrfs-progs: add checksum type to checksumming functions

[Johannes Thumshirn <jthumshirn@suse.de>]

On 12/08/2019 12:21, Nikolay Borisov wrote:
> 
> 
> On 25.07.19 г. 12:33 ч., Johannes Thumshirn wrote:
>> Add the checksum type to csum_tree_block_size(), __csum_tree_block_size()
>> and verify_tree_block_csum_silent().
>>
>> Signed-off-by: Johannes Thumshirn <jthumshirn@suse.de>
>> ---
>>  btrfs-corrupt-block.c       |  3 ++-
>>  cmds/rescue-chunk-recover.c |  3 ++-
>>  convert/common.c            |  3 ++-
>>  convert/main.c              |  3 ++-
>>  disk-io.c                   | 21 ++++++++++++---------
>>  disk-io.h                   |  5 +++--
>>  mkfs/common.c               | 21 ++++++++++++++-------
>>  7 files changed, 37 insertions(+), 22 deletions(-)
>>
>> diff --git a/btrfs-corrupt-block.c b/btrfs-corrupt-block.c
>> index bbef0c02e5d1..1dde9594bdcc 100644
>> --- a/btrfs-corrupt-block.c
>> +++ b/btrfs-corrupt-block.c
>> @@ -158,7 +158,8 @@ static void corrupt_keys(struct btrfs_trans_handle *trans,
>>  	if (!trans) {
>>  		u16 csum_size =
>>  			btrfs_super_csum_size(fs_info->super_copy);
>> -		csum_tree_block_size(eb, csum_size, 0);
>> +		u16 csum_type = btrfs_super_csum_type(fs_info->super_copy);
>> +		csum_tree_block_size(eb, csum_size, 0, csum_type);
> 
> I'm not a big fan of open-coding this csum manipulation. Generally if we
> ensure that we always pass a well-formed eb to csum_tree_block_size and
> by well-formed I mean one which has its ->fs_info, which in turn is
> well-formed then we can encapsulate this interrogation of csum type into
> csum_tree_block.
> 
>>  		write_extent_to_disk(eb);
>>  	}
>>  }
>> diff --git a/cmds/rescue-chunk-recover.c b/cmds/rescue-chunk-recover.c
>> index 308731ea5ea6..1a368310d895 100644
>> --- a/cmds/rescue-chunk-recover.c
>> +++ b/cmds/rescue-chunk-recover.c
>> @@ -768,7 +768,8 @@ static int scan_one_device(void *dev_scan_struct)
>>  			continue;
>>  		}
>>  
>> -		if (verify_tree_block_csum_silent(buf, rc->csum_size)) {
>> +		if (verify_tree_block_csum_silent(buf, rc->csum_size,
>> +						  rc->csum_type)) {
> 
> But this is not a well-formed eb :(
> 
>>  			bytenr += rc->sectorsize;
>>  			continue;
>>  		}
>> diff --git a/convert/common.c b/convert/common.c
>> index dea5f5b20d50..f8bbb23cba89 100644
>> --- a/convert/common.c
>> +++ b/convert/common.c
>> @@ -223,7 +223,8 @@ static inline int write_temp_extent_buffer(int fd, struct extent_buffer *buf,
>>  {
>>  	int ret;
>>  
>> -	csum_tree_block_size(buf, btrfs_csum_sizes[BTRFS_CSUM_TYPE_CRC32], 0);
>> +	csum_tree_block_size(buf, btrfs_csum_sizes[cfg->csum_type], 0,
>> +			     cfg->csum_type);
> 
> Neither is this ...
>>  
>>  	/* Temporary extent buffer is always mapped 1:1 on disk */
>>  	ret = pwrite(fd, buf->data, buf->len, bytenr);
>> diff --git a/convert/main.c b/convert/main.c
>> index 9711874bd137..5e6b12431f59 100644
>> --- a/convert/main.c
>> +++ b/convert/main.c
>> @@ -1058,7 +1058,8 @@ static int migrate_super_block(int fd, u64 old_bytenr)
>>  	BUG_ON(btrfs_super_bytenr(super) != old_bytenr);
>>  	btrfs_set_super_bytenr(super, BTRFS_SUPER_INFO_OFFSET);
>>  
>> -	csum_tree_block_size(buf, btrfs_csum_sizes[BTRFS_CSUM_TYPE_CRC32], 0);
>> +	csum_tree_block_size(buf, btrfs_csum_sizes[BTRFS_CSUM_TYPE_CRC32], 0,
>> +			     btrfs_super_csum_type(super));
>>  	ret = pwrite(fd, buf->data, BTRFS_SUPER_INFO_SIZE,
>>  		BTRFS_SUPER_INFO_OFFSET);
>>  	if (ret != BTRFS_SUPER_INFO_SIZE)
>> diff --git a/disk-io.c b/disk-io.c
>> index 01314504a50a..a4995a628210 100644
>> --- a/disk-io.c
>> +++ b/disk-io.c
>> @@ -149,7 +149,7 @@ void btrfs_csum_final(u32 crc, u8 *result)
>>  }
>>  
>>  static int __csum_tree_block_size(struct extent_buffer *buf, u16 csum_size,
>> -				  int verify, int silent)
>> +				  int verify, int silent, u16 csum_type)
>>  {
>>  	u8 result[BTRFS_CSUM_SIZE];
>>  	u32 len;
>> @@ -174,24 +174,27 @@ static int __csum_tree_block_size(struct extent_buffer *buf, u16 csum_size,
>>  	return 0;
>>  }
>>  
>> -int csum_tree_block_size(struct extent_buffer *buf, u16 csum_size, int verify)
>> +int csum_tree_block_size(struct extent_buffer *buf, u16 csum_size, int verify,
>> +			 u16 csum_type)
>>  {
>> -	return __csum_tree_block_size(buf, csum_size, verify, 0);
>> +	return __csum_tree_block_size(buf, csum_size, verify, 0, csum_type);
>>  }
>>  
>> -int verify_tree_block_csum_silent(struct extent_buffer *buf, u16 csum_size)
>> +int verify_tree_block_csum_silent(struct extent_buffer *buf, u16 csum_size,
>> +				  u16 csum_type)
>>  {
>> -	return __csum_tree_block_size(buf, csum_size, 1, 1);
>> +	return __csum_tree_block_size(buf, csum_size, 1, 1, csum_type);
>>  }
> 
> Now that those function take the csum_type it renders passing the
> csum_size redundant. Refactor them so that only csum_type is passed and
> make __csum_tree_block_size (perhahps it should be renamed now?) get the
> size based on the type.
> 
> I think the csum interface can be simplified further:
> 
> 1. Leave csum_tree_block for callers that pass well-formed eb's (as is
> currently)
> 
> 2. Rename __csum_tree_block_size to __csum_tree_block which takes just
> the buffer, boolean flag verify and u16 (or better enum btrfs_csum_type)
> csum_type arguments and the nitty-gritty details
> 
> 3. Remove csum_tree_block_size and directly call the new
> __csum_tree_block. The former just wraps the 'silent' parameter...
> 
> In the end we will be left with just 2 functions:
> 
> (__)csum_tree_blocks .
> 
> Additionally verify_tree_block_csum_silent has really one caller because
> in csum_tree_block it's enough to call __csum_tree_block_size just once
> and pass verify and fs_info->suppress_check_block_errors directly to it.
> Then you can also directly call __csum_tree_block in scan_one_device in
> rescue-chunk-recover.c . And this would put some sanity into the current
> csum interfaces in progs.
> 
> I guess this could land as a follow-up to your patches...
> 

I'd really like to have this rework as a follow up as it introduces a
non-negligible regression potential and can "easily" be done after this
series has landed.


-- 
Johannes Thumshirn                            SUSE Labs Filesystems
jthumshirn@suse.de                                +49 911 74053 689
SUSE Software Solutions Germany GmbH
Maxfeldstr. 5
90409 Nürnberg
Germany
(HRB 247165, AG München)
Key fingerprint = EC38 9CAB C2C4 F25D 8600 D0D0 0393 969D 2D76 0850