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>
#define RDS_IB_DEFAULT_RECV_WR 1024
#define RDS_IB_DEFAULT_SEND_WR 256
-#define RDS_IB_DEFAULT_FR_WR 256
-#define RDS_IB_DEFAULT_FR_INV_WR 256
+#define RDS_IB_DEFAULT_FR_WR 512
#define RDS_IB_DEFAULT_RETRY_COUNT 1
/* To control the number of wrs from fastreg */
atomic_t i_fastreg_wrs;
- atomic_t i_fastunreg_wrs;
/* interrupt handling */
struct tasklet_struct i_send_tasklet;
* completion queue and send queue. This extra space is used for FRMR
* registration and invalidation work requests
*/
- fr_queue_space = rds_ibdev->use_fastreg ?
- (RDS_IB_DEFAULT_FR_WR + 1) +
- (RDS_IB_DEFAULT_FR_INV_WR + 1)
- : 0;
+ fr_queue_space = (rds_ibdev->use_fastreg ? RDS_IB_DEFAULT_FR_WR : 0);
/* add the conn now so that connection establishment has the dev */
rds_ib_add_conn(rds_ibdev, conn);
attr.send_cq = ic->i_send_cq;
attr.recv_cq = ic->i_recv_cq;
atomic_set(&ic->i_fastreg_wrs, RDS_IB_DEFAULT_FR_WR);
- atomic_set(&ic->i_fastunreg_wrs, RDS_IB_DEFAULT_FR_INV_WR);
/*
* XXX this can fail if max_*_wr is too large? Are we supposed
wait_event(rds_ib_ring_empty_wait,
rds_ib_ring_empty(&ic->i_recv_ring) &&
(atomic_read(&ic->i_signaled_sends) == 0) &&
- (atomic_read(&ic->i_fastreg_wrs) == RDS_IB_DEFAULT_FR_WR) &&
- (atomic_read(&ic->i_fastunreg_wrs) == RDS_IB_DEFAULT_FR_INV_WR));
+ (atomic_read(&ic->i_fastreg_wrs) == RDS_IB_DEFAULT_FR_WR));
tasklet_kill(&ic->i_send_tasklet);
tasklet_kill(&ic->i_recv_tasklet);
if (frmr->fr_state != FRMR_IS_INUSE)
goto out;
- while (atomic_dec_return(&ibmr->ic->i_fastunreg_wrs) <= 0) {
- atomic_inc(&ibmr->ic->i_fastunreg_wrs);
+ while (atomic_dec_return(&ibmr->ic->i_fastreg_wrs) <= 0) {
+ atomic_inc(&ibmr->ic->i_fastreg_wrs);
cpu_relax();
}
if (unlikely(ret)) {
frmr->fr_state = FRMR_IS_STALE;
frmr->fr_inv = false;
- atomic_inc(&ibmr->ic->i_fastunreg_wrs);
+ atomic_inc(&ibmr->ic->i_fastreg_wrs);
pr_err("RDS/IB: %s returned error(%d)\n", __func__, ret);
goto out;
}
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);
}
+
+ atomic_inc(&ic->i_fastreg_wrs);
}
void rds_ib_unreg_frmr(struct list_head *list, unsigned int *nfreed,