RDMA: Fix dead loop when transfer large data (20KB)

[fengquyoumo]

Determine the status of the Client when attempting to read data. If state=CLIENT_COMPLETED_IO, no read attempt is made and I/O operations on the Client are rescheduled by the main thread.

And 20474 Byte = PROTO_IOBUF_LEN(16KB) + SDS_HDR_VAR(16, s)(4090 Byte)

Fixes #1385

[zuiderkwast]

@pizhenwei RDMA expert, please take a look.

@fengquyoumo Thanks for your contribution. You need to commit with git commit -s which adds the Signed-off-by header in the commit message. See CONTRIBUTING.md about why we need this. How to fix it, see the Details link at the "DCO" CI job.

[codecov]

Codecov Report

All modified and coverable lines are covered by tests ✅

Project coverage is 70.83%. Comparing base (3df609e) to head (8517cde).
Report is 7 commits behind head on unstable.

Additional details and impacted files

@@             Coverage Diff              @@
##           unstable    #1386      +/-   ##
============================================
+ Coverage     70.78%   70.83%   +0.05%     
============================================
  Files           118      118              
  Lines         63497    63549      +52     
============================================
+ Hits          44945    45014      +69     
+ Misses        18552    18535      -17     

Files with missing lines	Coverage Δ
src/rdma.c	100.00% <ø> (ø)

... and 16 files with indirect coverage changes

[pizhenwei]

Hi @fengquyoumo
I've got a cold, and will be back soon. please be a little more patient.

[fengquyoumo]

@zuiderkwast :) Thanks for your help!

[pizhenwei]

Hi @fengquyoumo
Thanks for your report and contribution!

There is a lack of multi-threads support from RDMA. In general, you are right on this issue, but it's not valkey connection style. I suggest:

• implement postpone_update_state and update_state for RDMA
• CLIENT_PENDING_IO is only used by networking.c and io_threads.c, don't break this. Use RDMA_CONN_FLAG_POSTPONE_UPDATE_STATE in rdma.c
• handle pending data in rdma.c correctly

Please see tls.c as example:

    .postpone_update_state = postPoneUpdateSSLState,
    .update_state = updateSSLState,

[fengquyoumo]

@pizhenwei Thanks for your help！
I implemented postpone_update_state and update_state for RDMA to solve this issue.
Also, the first implementation had serious performance issues, but this one worked just fine.

Please review the code again when you are free, Thanks!

[pizhenwei]

I guess

struct rdma_cm_id *cm_id;
int flags;
int last_errno;

would reduce 8 bytes.

Otherwise LGTM. I also test it locally, it works fine.

Hi @zuiderkwast , would you please take a look at this change?

[fengquyoumo]

:) Thanks a lot for the detailed feedback!

[pizhenwei]

LGTM.

[zuiderkwast]

Great. Thanks both of you for the fix and the review.