Re: [PATCH v22 10/16] RFC xfs: Skip flip flags for delayed attrs

[Allison Henderson <allison.henderson@oracle.com>]

On 7/28/21 12:18 PM, Darrick J. Wong wrote:
> On Mon, Jul 26, 2021 at 11:20:47PM -0700, Allison Henderson wrote:
>> This is a clean up patch that skips the flip flag logic for delayed attr
>> renames.  Since the log replay keeps the inode locked, we do not need to
>> worry about race windows with attr lookups.  So we can skip over
>> flipping the flag and the extra transaction roll for it
>>
>> RFC: In the last review, folks asked for some performance analysis, so I
>> did a few perf captures with and with out this patch.  What I found was
>> that there wasnt very much difference at all between having the patch or
>> not having it.  Of the time we do spend in the affected code, the
>> percentage is small.  Most of the time we spend about %0.03 of the time
>> in this function, with or with out the patch.  Occasionally we get a
>> 0.02%, though not often.  So I think this starts to challenge needing
>> this patch at all. This patch was requested some number of reviews ago,
>> be perhaps in light of the findings, it may no longer be of interest.
>>
>>       0.03%     0.00%  fsstress  [xfs]               [k] xfs_attr_set_iter
>>
>> Keep it or drop it?
>>
>> Signed-off-by: Allison Henderson <allison.henderson@oracle.com>
>> Reviewed-by: Darrick J. Wong <djwong@kernel.org>
> 
> /me hates it when he notices things after review... :/
> 
>> ---
>>   fs/xfs/libxfs/xfs_attr.c      | 51 +++++++++++++++++++++++++------------------
>>   fs/xfs/libxfs/xfs_attr_leaf.c |  3 ++-
>>   2 files changed, 32 insertions(+), 22 deletions(-)
>>
>> diff --git a/fs/xfs/libxfs/xfs_attr.c b/fs/xfs/libxfs/xfs_attr.c
>> index 11d8081..eee219c6 100644
>> --- a/fs/xfs/libxfs/xfs_attr.c
>> +++ b/fs/xfs/libxfs/xfs_attr.c
>> @@ -355,6 +355,7 @@ xfs_attr_set_iter(
>>   	struct xfs_inode		*dp = args->dp;
>>   	struct xfs_buf			*bp = NULL;
>>   	int				forkoff, error = 0;
>> +	struct xfs_mount		*mp = args->dp->i_mount;
>>   
>>   	/* State machine switch */
>>   	switch (dac->dela_state) {
>> @@ -476,16 +477,21 @@ xfs_attr_set_iter(
>>   		 * In a separate transaction, set the incomplete flag on the
>>   		 * "old" attr and clear the incomplete flag on the "new" attr.
>>   		 */
>> -		error = xfs_attr3_leaf_flipflags(args);
>> -		if (error)
>> -			return error;
>> -		/*
>> -		 * Commit the flag value change and start the next trans in
>> -		 * series.
>> -		 */
>> -		dac->dela_state = XFS_DAS_FLIP_LFLAG;
>> -		trace_xfs_attr_set_iter_return(dac->dela_state, args->dp);
>> -		return -EAGAIN;
>> +		if (!xfs_hasdelattr(mp)) {
> 
> More nitpicking: this should be gated directly on the log incompat
> feature check, not the LARP knob...
I think they're equivelent functionally right now.  Looking forward, if 
we assume that the knob will one day disappear, it might be a good idea 
to leave it wrapping any code that would disappear along with it.  Just 
as a sort of reminder to clean it out.

So for example, if we ever decide that delayed attrs are just alway on, 
this whole chunk could just go away with it.  OTOTH, if there is an 
internal need to run attrs without the logging, it would make sense to 
check the feature bit here, as this code would still be needed even when 
the knob is gone.

So i guess this is a question of what we think we will need in the future?

> 
> 		if (!xfs_sb_version_haslogxattrs(&mp->m_sb)) {
> 
>> +			error = xfs_attr3_leaf_flipflags(args);
>> +			if (error)
>> +				return error;
>> +			/*
>> +			 * Commit the flag value change and start the next trans
>> +			 * in series.
>> +			 */
>> +			dac->dela_state = XFS_DAS_FLIP_LFLAG;
>> +			trace_xfs_attr_set_iter_return(dac->dela_state,
>> +						       args->dp);
>> +			return -EAGAIN;
>> +		}
>> +
>> +		/* fallthrough */
>>   	case XFS_DAS_FLIP_LFLAG:
>>   		/*
>>   		 * Dismantle the "old" attribute/value pair by removing a
>> @@ -587,17 +593,21 @@ xfs_attr_set_iter(
>>   		 * In a separate transaction, set the incomplete flag on the
>>   		 * "old" attr and clear the incomplete flag on the "new" attr.
>>   		 */
>> -		error = xfs_attr3_leaf_flipflags(args);
>> -		if (error)
>> -			goto out;
>> -		/*
>> -		 * Commit the flag value change and start the next trans in
>> -		 * series
>> -		 */
>> -		dac->dela_state = XFS_DAS_FLIP_NFLAG;
>> -		trace_xfs_attr_set_iter_return(dac->dela_state, args->dp);
>> -		return -EAGAIN;
>> +		if (!xfs_hasdelattr(mp)) {
> 
> ...and here...
> 
>> +			error = xfs_attr3_leaf_flipflags(args);
>> +			if (error)
>> +				goto out;
>> +			/*
>> +			 * Commit the flag value change and start the next trans
>> +			 * in series
>> +			 */
>> +			dac->dela_state = XFS_DAS_FLIP_NFLAG;
>> +			trace_xfs_attr_set_iter_return(dac->dela_state,
>> +						       args->dp);
>> +			return -EAGAIN;
>> +		}
>>   
>> +		/* fallthrough */
>>   	case XFS_DAS_FLIP_NFLAG:
>>   		/*
>>   		 * Dismantle the "old" attribute/value pair by removing a
>> @@ -1241,7 +1251,6 @@ xfs_attr_node_addname_clear_incomplete(
>>   	 * Re-find the "old" attribute entry after any split ops. The INCOMPLETE
>>   	 * flag means that we will find the "old" attr, not the "new" one.
>>   	 */
>> -	args->attr_filter |= XFS_ATTR_INCOMPLETE;
> 
> Why is this removed from the query arguments?  If we're going to skip
> the INCOMPLETE flag dance, I would have thought that you'd change the
> XFS_DAS_CLR_FLAG case to omit xfs_attr_node_addname_clear_incomplete if
> the logged xattr feature is set?
> 
>>   	state = xfs_da_state_alloc(args);
>>   	state->inleaf = 0;
>>   	error = xfs_da3_node_lookup_int(state, &retval);
>> diff --git a/fs/xfs/libxfs/xfs_attr_leaf.c b/fs/xfs/libxfs/xfs_attr_leaf.c
>> index b910bd2..a9116ee 100644
>> --- a/fs/xfs/libxfs/xfs_attr_leaf.c
>> +++ b/fs/xfs/libxfs/xfs_attr_leaf.c
>> @@ -1482,7 +1482,8 @@ xfs_attr3_leaf_add_work(
>>   	if (tmp)
>>   		entry->flags |= XFS_ATTR_LOCAL;
>>   	if (args->op_flags & XFS_DA_OP_RENAME) {
>> -		entry->flags |= XFS_ATTR_INCOMPLETE;
>> +		if (!xfs_hasdelattr(mp))
> 
> Same change here as above.
> 
> --D
> 
>> +			entry->flags |= XFS_ATTR_INCOMPLETE;
>>   		if ((args->blkno2 == args->blkno) &&
>>   		    (args->index2 <= args->index)) {
>>   			args->index2++;
>> -- 
>> 2.7.4
>>