On 2023/7/31 14:09, Christoph Hellwig wrote:

On Tue, Jul 25, 2023 at 09:00:59PM +0800, chengming.zhou@linux.dev wrote:

quoted

From: Chengming Zhou <redacted>

The cmd_flags in blk_kick_flush() should inherit the original request's
cmd_flags, but the current code looks buggy to me:

Should it?  I know the code is kinda trying to do it, but does it really
make sense?  Adding Hannes who originally added this inheritance and
discussing the details below:

I'm not sure, actually I don't get what the current code is doing...
Hope Hannes could provide some details.

blk_flush_complete_seq(rq) -> blk_kick_flush(rq->cmd_flags)

flush_rq will use the cmd_flags of request which just complete a sequence,
there are three cases:

1. blk_insert_flush(rq):	rq is pending, wait for flush
2. flush_end_io(flush_rq):	rq flush seq done
3. mq_flush_data_end_io(rq):	rq data seq done

Only in the 1st case, the rq is the pending request that wait for flush_rq.
In the 2nd and 3rd cases, the rq has nothing to do with the next flush_rq?

So it's more reasonable for flush_rq to use its pending first_rq's cmd_flags?

quoted

 	flush_rq->cmd_flags = REQ_OP_FLUSH | REQ_PREFLUSH;
-	flush_rq->cmd_flags |= (flags & REQ_DRV) | (flags & REQ_FAILFAST_MASK);
+	flush_rq->cmd_flags |= (first_rq->cmd_flags & REQ_DRV) |
+			       (first_rq->cmd_flags & REQ_FAILFAST_MASK);

Two cases here:

 1) REQ_FAILFAST_MASK:  I don't think this is actually set on flush request
    currently, and even if it was applying it to the flush that serves more
    than a single originating command seems wrong to me.
 2) REQ_DRV is only set by drivers that have seen a bio.  For dm this
    is used as REQ_DM_POLL_LIST which should never be set for a flush/fua
    request.  For nvme-mpath it is REQ_NVME_MPATH, which is set in the
    bio based driver and used for decision making in the I/O completion
    handler.  So I guess this one actually does need to get passed
    through.

The commit 84fca1b0c461 ("block: pass failfast and driver-specific flags to
flush requests") says:
If flush requests are being sent to the device we need to inherit the
failfast and driver-specific flags, too, otherwise I/O will fail.

1) REQ_FAILFAST_MASK: agree, shouldn't set to the flush_rq I think?
2) REQ_DRV: I don't get why this flag not set would cause I/O fail?

Thanks!