Re: [PATCH 2/2] f2fs: support {d,id,did,x}node checksum

From: Chao Yu
Date: Fri Apr 13 2018 - 04:40:43 EST


Ping again..

Do you have time to discuss this?

On 2018/2/27 22:16, Chao Yu wrote:
> Ping,
>
> On 2018/2/13 15:34, Chao Yu wrote:
>> Hi Jaegeuk,
>>
>> On 2018/2/10 10:52, Chao Yu wrote:
>>> On 2018/2/10 9:41, Jaegeuk Kim wrote:
>>>> On 02/01, Chao Yu wrote:
>>>>>
>>>>>
>>>>> On 2018/2/1 6:15, Jaegeuk Kim wrote:
>>>>>> On 01/31, Chao Yu wrote:
>>>>>>> On 2018/1/31 10:02, Jaegeuk Kim wrote:
>>>>>>>> What if we want to add more entries in addition to node_checksum? Do we have
>>>>>>>> to add a new feature flag at every time? How about adding a layout value instead
>>>>>>>
>>>>>>> Hmm.. for previous implementation, IMO, we'd better add a new feature flag at
>>>>>>> every time, otherwise, w/ extra_nsize only, in current image, we can know a
>>>>>>> valid range of extended area in node block, but we don't know which
>>>>>>> fields/features are valid/enabled or not.
>>>>>>>
>>>>>>> One more thing is that if we can add one feature flag for each field, we got one
>>>>>>> more chance to disable it dynamically.
>>>>>>>
>>>>>>>> of extra_nsize? For example, layout #1 means node_checksum with extra_nsize=X?
>>>>>>>>
>>>>>>>>
>>>>>>>> What does 1017 mean? We need to make this structure more flexibly for new
>>>>>>>
>>>>>>> Yes, using raw 1017 is not appropriate here.
>>>>>>>
>>>>>>>> entries. Like this?
>>>>>>>> union {
>>>>>>>> struct node_v1;
>>>>>>>> struct node_v2;
>>>>>>>> struct node_v3;
>>>>>>>> ...
>>>>>>>> struct direct_node dn;
>>>>>>>> struct indirect_node in;
>>>>>>>> };
>>>>>>>> };
>>>>>>>>
>>>>>>>> struct node_v1 {
>>>>>>>> __le32 data[DEF_ADDRS_PER_BLOCK - V1_NSIZE=1];
>>>>>>>> __le32 node_checksum;
>>>>>>>> }
>>>>>>>>
>>>>>>>> struct node_v2 {
>>>>>>>> __le32 data[DEF_ADDRS_PER_BLOCK - V2_NSIZE=500];
>>>>>>>
>>>>>>> Hmm.. If we only need to add one more 4 bytes field in struct node_v2, but
>>>>>>> V2_NSIZE is defined as fixed 500, there must be 492 bytes wasted.
>>>>>>>
>>>>>>> Or we can define V2_NSIZE as 8, but if there comes more and more extended
>>>>>>> fields, node version count can be a large number, it results in complicated
>>>>>>> version recognization and handling.
>>>>>>>
>>>>>>> One more question is how can we control which fields are valid or not in
>>>>>>> comp[Vx_NSIZE]?
>>>>>>>
>>>>>>>
>>>>>>> Anyway, what I'm thinking is maybe we can restructure layout of node block like
>>>>>>> the one used by f2fs_inode:
>>>>>>>
>>>>>>> struct f2fs_node {
>>>>>>> union {
>>>>>>> struct f2fs_inode i;
>>>>>>> union {
>>>>>>> struct {
>>>>>>> __le32 node_checksum;
>>>>>>> __le32 feature_field_1;
>>>>>>> __le32 feature_field_2;
>>>>>>> ....
>>>>>>> __le32 addr[];
>>>>>>>
>>>>>>> };
>>>>>>> struct direct_node dn;
>>>>>>> struct indirect_node in;
>>>>>>> };
>>>>>>> };
>>>>>>> struct node_footer footer;
>>>>>>> } __packed;
>>>>>>>
>>>>>>> Moving all extended fields to the head of f2fs_node, so we don't have to use
>>>>>>> macro to indicate actual size of addr.
>>>>>>
>>>>>> Thinking what'd be the best way. My concern is, once getting more entries, we
>>>>>
>>>>> OK, I think we need more discussion.. ;)
>>>>>
>>>>>> can't set each of features individually. Like the second entry should have
>>>>>
>>>>> Oh, that will be hard. If we have to avoid that, we have to tag in somewhere
>>>>> e.g. f2fs_inode::i_flags2 to indicate which new field in f2fs_node is valid, for
>>>>> example:
>>>>>
>>>>> #define F2FS_NODE_CHECKSUM 0x0001
>>>>> #define F2FS_NODE_FIELD1 0x0002
>>>>> #define F2FS_NODE_FIELD2 0x0004
>>>>>
>>>>> union {
>>>>> struct {
>>>>> __le32 node_checksum;
>>>>> __le32 field_1;
>>>>> __le32 field_2;
>>>>> ....
>>>>> __le32 addr[];
>>>>> };
>>>>> struct direct_node dn;
>>>>> struct indirect_node in;
>>>>> };
>>>>>
>>>>> f2fs_inode::i_flags2 = F2FS_NODE_CHECKSUM | F2FS_NODE_FIELD1
>>>>> indicates that f2fs_node::node_checksum and f2fs_node::field_1 are valid;
>>>>>
>>>>> f2fs_inode::i_flags2 = F2FS_NODE_FIELD1 | F2FS_NODE_FIELD2
>>>>> indicates that f2fs_node::field_1 and f2fs_node::field_2 are valid.
>>>>
>>>> So, that's why I thought we may need a sort of each formats.
>>>
>>> Hmm.. if we have two new added fields, there are (2 << 2) combinations
>>> of all formats, as:
>>>
>>> struct original {
>>> __le32 data[DEF_ADDRS_PER_BLOCK];
>>> }
>>>
>>> struct node_v1 {
>>> __le32 data[DEF_ADDRS_PER_BLOCK - V1_NSIZE=1];
>>> __le32 field_1;
>>> }
>>>
>>> struct node_v2 {
>>> __le32 data[DEF_ADDRS_PER_BLOCK - V2_NSIZE=1];
>>> __le32 field_2;
>>> }
>>>
>>> struct node_v2 {
>>> __le32 data[DEF_ADDRS_PER_BLOCK - V3_NSIZE=2];
>>> __le32 field_1;
>>> __le32 field_2;
>>> }
>>>
>>> If we add more new fields, the node version will increase sharply due
>>> to there is (n << 2) combination with n fields. Right? Any thoughts to
>>> reduce maintaining overhead on those node versions structures?
>>
>> Do you have time to explain more about the design of multiple version structure
>> for node block, I'm still be confused about two things:
>> 1. what will we do if we want to add one new field in node structure.
>> 2. how can we recognize which fields are valid and which ones are invalid.
>>
>> Thanks,
>>
>>>
>>> Thanks,
>>>
>>>>
>>>>>
>>>>> Any thoughts?
>>>>>
>>>>> Thanks,
>>>>>
>>>>>> enabled node_checksum, which we may not want to do.
>>>>>>
>>>>>>>
>>>>>>> Thanks,
>>>>>>>
>>>>>>>> __le32 comp[V2_NSIZE];
>>>>>>>> }
>>>>>>>> ...
>>>>>>>>
>>>>>>>>> + };
>>>>>>>>> + struct direct_node dn;
>>>>>>>>> + struct indirect_node in;
>>>>>>>>> + };
>>>>>>>>> };
>>>>>>>>> struct node_footer footer;
>>>>>>>>> } __packed;
>>>>>>>>> --
>>>>>>>>> 2.15.0.55.gc2ece9dc4de6
>>>>>>>>
>>>>>>>> .
>>>>>>>>
>>>
>>> .
>>>
>>
>
> .
>