OSDN Git Service

Revert "RDS: IB: split the mr registration and invalidation path"
authorGerd Rausch <gerd.rausch@oracle.com>
Sat, 29 Jun 2019 00:31:19 +0000 (17:31 -0700)
committerSantosh Shilimkar <santosh.shilimkar@oracle.com>
Wed, 10 Jul 2019 04:45:41 +0000 (21:45 -0700)
commita55207884708bf1d5b8c87a3c504502de77a5416
tree65b06062d9c6f97fc55c68d15abfc63bc18106a8
parent616d37a070bb33ea387d0e93343acd8336a30886
Revert "RDS: IB: split the mr registration and invalidation path"

This reverts commit 56012459310a1dbcc55c2dbf5500a9f7571402cb.

RDS kept spinning inside function "rds_ib_post_reg_frmr", waiting for
"i_fastreg_wrs" to become incremented:
         while (atomic_dec_return(&ibmr->ic->i_fastreg_wrs) <= 0) {
                 atomic_inc(&ibmr->ic->i_fastreg_wrs);
                 cpu_relax();
         }

Looking at the original commit:

commit 56012459310a ("RDS: IB: split the mr registration and
invalidation path")

In there, the "rds_ib_mr_cqe_handler" was changed in the following
way:

 void rds_ib_mr_cqe_handler(struct
 rds_ib_connection *ic,
 struct ib_wc *wc)
        if (frmr->fr_inv) {
                  frmr->fr_state = FRMR_IS_FREE;
                  frmr->fr_inv = false;
                atomic_inc(&ic->i_fastreg_wrs);
        } else {
                atomic_inc(&ic->i_fastunreg_wrs);
        }

It looks like it's got it exactly backwards:

Function "rds_ib_post_reg_frmr" keeps track of the outstanding
requests via "i_fastreg_wrs".

Function "rds_ib_post_inv" keeps track of the outstanding requests
via "i_fastunreg_wrs" (post original commit). It also sets:
         frmr->fr_inv = true;

However the completion handler "rds_ib_mr_cqe_handler" adjusts
"i_fastreg_wrs" when "fr_inv" had been true, and adjusts
"i_fastunreg_wrs" otherwise.

The original commit was done in the name of performance:
to remove the performance bottleneck

No performance benefit could be observed with a fixed-up version
of the original commit measured between two Oracle X7 servers,
both equipped with Mellanox Connect-X5 HCAs.

The prudent course of action is to revert this commit.

Signed-off-by: Gerd Rausch <gerd.rausch@oracle.com>
Signed-off-by: Santosh Shilimkar <santosh.shilimkar@oracle.com>
net/rds/ib.h
net/rds/ib_cm.c
net/rds/ib_frmr.c