From: Mike Marciniszyn <mike.marciniszyn@xxxxxxxxx>
In some cases, the cost of user memory deregistration is more
important than the data path benefit of percpu reference counts.
Add a (default off) module parameter to disarm percpu for user memory
regions.
Reviewed-by: Dennis Dalessandro <dennis.dalessandro@xxxxxxxxx>
Signed-off-by: Mike Marciniszyn <mike.marciniszyn@xxxxxxxxx>
Signed-off-by: Dennis Dalessandro <dennis.dalessandro@xxxxxxxxx>
---
drivers/infiniband/hw/hfi1/verbs.c | 7 +++++++
drivers/infiniband/hw/qib/qib_verbs.c | 7 +++++++
drivers/infiniband/sw/rdmavt/mr.c | 6 +++++-
include/rdma/rdma_vt.h | 1 +
4 files changed, 20 insertions(+), 1 deletions(-)
diff --git a/drivers/infiniband/hw/hfi1/verbs.c b/drivers/infiniband/hw/hfi1/verbs.c
index 5e7e577..552b26d 100644
--- a/drivers/infiniband/hw/hfi1/verbs.c
+++ b/drivers/infiniband/hw/hfi1/verbs.c
@@ -68,6 +68,12 @@
MODULE_PARM_DESC(lkey_table_size,
"LKEY table size in bits (2^n, 1 <= n <= 23)");
+static unsigned int hfi1_no_user_mr_percpu;
+module_param_named(no_user_mr_percpu, hfi1_no_user_mr_percpu, uint,
+ S_IRUGO);
+MODULE_PARM_DESC(no_user_mr_percpu,
+ "Avoid percpu refcount for user MRs (default 0)");
+
static unsigned int hfi1_max_pds = 0xFFFF;
module_param_named(max_pds, hfi1_max_pds, uint, S_IRUGO);
MODULE_PARM_DESC(max_pds,
@@ -1841,6 +1847,7 @@ int hfi1_register_ib_device(struct hfi1_devdata *dd)
/* misc settings */
dd->verbs_dev.rdi.flags = 0; /* Let rdmavt handle it all */
dd->verbs_dev.rdi.dparms.lkey_table_size = hfi1_lkey_table_size;
+ dd->verbs_dev.rdi.dparms.no_user_mr_percpu = hfi1_no_user_mr_percpu;
dd->verbs_dev.rdi.dparms.nports = dd->num_pports;
dd->verbs_dev.rdi.dparms.npkeys = hfi1_get_npkeys(dd);
diff --git a/drivers/infiniband/hw/qib/qib_verbs.c b/drivers/infiniband/hw/qib/qib_verbs.c
index e120efe..6c718cd 100644
--- a/drivers/infiniband/hw/qib/qib_verbs.c
+++ b/drivers/infiniband/hw/qib/qib_verbs.c
@@ -56,6 +56,12 @@
MODULE_PARM_DESC(lkey_table_size,
"LKEY table size in bits (2^n, 1 <= n <= 23)");
+static unsigned int qib_no_user_mr_percpu;
+module_param_named(no_user_mr_percpu, qib_no_user_mr_percpu, uint,
+ S_IRUGO);
+MODULE_PARM_DESC(no_user_mr_percpu,
+ "Avoid percpu refcount for user MRs (default 0)");
+
static unsigned int ib_qib_max_pds = 0xFFFF;
module_param_named(max_pds, ib_qib_max_pds, uint, S_IRUGO);
MODULE_PARM_DESC(max_pds,
@@ -1606,6 +1612,7 @@ int qib_register_ib_device(struct qib_devdata *dd)
dd->verbs_dev.rdi.dparms.max_rdma_atomic = QIB_MAX_RDMA_ATOMIC;
dd->verbs_dev.rdi.driver_f.get_guid_be = qib_get_guid_be;
dd->verbs_dev.rdi.dparms.lkey_table_size = qib_lkey_table_size;
+ dd->verbs_dev.rdi.dparms.no_user_mr_percpu = qib_no_user_mr_percpu;
dd->verbs_dev.rdi.dparms.qp_table_size = ib_qib_qp_table_size;
dd->verbs_dev.rdi.dparms.qpn_start = 1;
dd->verbs_dev.rdi.dparms.qpn_res_start = QIB_KD_QP;
diff --git a/drivers/infiniband/sw/rdmavt/mr.c b/drivers/infiniband/sw/rdmavt/mr.c
index 7c86955..bbcc31f 100644
--- a/drivers/infiniband/sw/rdmavt/mr.c
+++ b/drivers/infiniband/sw/rdmavt/mr.c
@@ -280,6 +280,7 @@ static void rvt_free_lkey(struct rvt_mregion *mr)
struct rvt_mr *mr;
int rval = -ENOMEM;
int m;
+ struct rvt_dev_info *dev = ib_to_rvt(pd->device);
/* Allocate struct plus pointers to first level page tables. */
m = (count + RVT_SEGSZ - 1) / RVT_SEGSZ;
@@ -287,7 +288,10 @@ static void rvt_free_lkey(struct rvt_mregion *mr)
if (!mr)
goto bail;
- rval = rvt_init_mregion(&mr->mr, pd, count, 0);
+ rval = rvt_init_mregion(&mr->mr, pd, count,
+ ibpd_to_rvtpd(pd)->user &&
+ dev->dparms.no_user_mr_percpu ?
+ PERCPU_REF_INIT_ATOMIC : 0);
if (rval)
goto bail;
/*
diff --git a/include/rdma/rdma_vt.h b/include/rdma/rdma_vt.h
index 8fc1ca7..d60a41e 100644
--- a/include/rdma/rdma_vt.h
+++ b/include/rdma/rdma_vt.h
@@ -142,6 +142,7 @@ struct rvt_driver_params {
* For instance special module parameters. Goes here.
*/
unsigned int lkey_table_size;
+ unsigned int no_user_mr_percpu;
unsigned int qp_table_size;
int qpn_start;
int qpn_inc;
--
To unsubscribe from this list: send the line "unsubscribe linux-rdma" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
