* [net-next][PATCH 01/13] RDS: Drop stale iWARP RDMA transport
[not found] ` <1455967804-27644-1-git-send-email-santosh.shilimkar-QHcLZuEGTsvQT0dZR+AlfA@public.gmane.org>
@ 2016-02-20 11:29 ` Santosh Shilimkar
2016-02-20 11:29 ` [net-next][PATCH 04/13] RDS: IB: Remove the RDS_IB_SEND_OP dependency Santosh Shilimkar
` (7 subsequent siblings)
8 siblings, 0 replies; 18+ messages in thread
From: Santosh Shilimkar @ 2016-02-20 11:29 UTC (permalink / raw)
To: netdev-u79uwXL29TY76Z2rM5mHXA, davem-fT/PcQaiUtIeIZ0/mPfg9Q
Cc: linux-rdma-u79uwXL29TY76Z2rM5mHXA,
linux-kernel-u79uwXL29TY76Z2rM5mHXA, Santosh Shilimkar
RDS iWarp support code has become stale and non testable. As
indicated earlier, am dropping the support for it.
If new iWarp user(s) shows up in future, we can adapat the RDS IB
transprt for the special RDMA READ sink case. iWarp needs an MR
for the RDMA READ sink.
Signed-off-by: Santosh Shilimkar <ssantosh-DgEjT+Ai2ygdnm+yROfE0A@public.gmane.org>
Signed-off-by: Santosh Shilimkar <santosh.shilimkar-QHcLZuEGTsvQT0dZR+AlfA@public.gmane.org>
---
Documentation/networking/rds.txt | 4 +-
net/rds/Kconfig | 7 +-
net/rds/Makefile | 4 +-
net/rds/iw.c | 312 -------------
net/rds/iw.h | 398 ----------------
net/rds/iw_cm.c | 769 ------------------------------
net/rds/iw_rdma.c | 837 ---------------------------------
net/rds/iw_recv.c | 904 ------------------------------------
net/rds/iw_ring.c | 169 -------
net/rds/iw_send.c | 981 ---------------------------------------
net/rds/iw_stats.c | 95 ----
net/rds/iw_sysctl.c | 123 -----
net/rds/rdma_transport.c | 13 +-
net/rds/rdma_transport.h | 5 -
14 files changed, 7 insertions(+), 4614 deletions(-)
delete mode 100644 net/rds/iw.c
delete mode 100644 net/rds/iw.h
delete mode 100644 net/rds/iw_cm.c
delete mode 100644 net/rds/iw_rdma.c
delete mode 100644 net/rds/iw_recv.c
delete mode 100644 net/rds/iw_ring.c
delete mode 100644 net/rds/iw_send.c
delete mode 100644 net/rds/iw_stats.c
delete mode 100644 net/rds/iw_sysctl.c
diff --git a/Documentation/networking/rds.txt b/Documentation/networking/rds.txt
index e1a3d59..9d219d8 100644
--- a/Documentation/networking/rds.txt
+++ b/Documentation/networking/rds.txt
@@ -19,9 +19,7 @@ to N*N if you use a connection-oriented socket transport like TCP.
RDS is not Infiniband-specific; it was designed to support different
transports. The current implementation used to support RDS over TCP as well
-as IB. Work is in progress to support RDS over iWARP, and using DCE to
-guarantee no dropped packets on Ethernet, it may be possible to use RDS over
-UDP in the future.
+as IB.
The high-level semantics of RDS from the application's point of view are
diff --git a/net/rds/Kconfig b/net/rds/Kconfig
index f2c670b..bffde4b 100644
--- a/net/rds/Kconfig
+++ b/net/rds/Kconfig
@@ -4,14 +4,13 @@ config RDS
depends on INET
---help---
The RDS (Reliable Datagram Sockets) protocol provides reliable,
- sequenced delivery of datagrams over Infiniband, iWARP,
- or TCP.
+ sequenced delivery of datagrams over Infiniband or TCP.
config RDS_RDMA
- tristate "RDS over Infiniband and iWARP"
+ tristate "RDS over Infiniband"
depends on RDS && INFINIBAND && INFINIBAND_ADDR_TRANS
---help---
- Allow RDS to use Infiniband and iWARP as a transport.
+ Allow RDS to use Infiniband as a transport.
This transport supports RDMA operations.
config RDS_TCP
diff --git a/net/rds/Makefile b/net/rds/Makefile
index 56d3f60..19e5485 100644
--- a/net/rds/Makefile
+++ b/net/rds/Makefile
@@ -6,9 +6,7 @@ rds-y := af_rds.o bind.o cong.o connection.o info.o message.o \
obj-$(CONFIG_RDS_RDMA) += rds_rdma.o
rds_rdma-y := rdma_transport.o \
ib.o ib_cm.o ib_recv.o ib_ring.o ib_send.o ib_stats.o \
- ib_sysctl.o ib_rdma.o \
- iw.o iw_cm.o iw_recv.o iw_ring.o iw_send.o iw_stats.o \
- iw_sysctl.o iw_rdma.o
+ ib_sysctl.o ib_rdma.o
obj-$(CONFIG_RDS_TCP) += rds_tcp.o
diff --git a/net/rds/iw.c b/net/rds/iw.c
deleted file mode 100644
index f4a9fff..0000000
diff --git a/net/rds/iw.h b/net/rds/iw.h
deleted file mode 100644
index 5af01d1..0000000
diff --git a/net/rds/iw_cm.c b/net/rds/iw_cm.c
deleted file mode 100644
index aea4c91..0000000
diff --git a/net/rds/iw_rdma.c b/net/rds/iw_rdma.c
deleted file mode 100644
index b09a40c..0000000
diff --git a/net/rds/iw_recv.c b/net/rds/iw_recv.c
deleted file mode 100644
index a66d179..0000000
diff --git a/net/rds/iw_ring.c b/net/rds/iw_ring.c
deleted file mode 100644
index da8e3b6..0000000
diff --git a/net/rds/iw_send.c b/net/rds/iw_send.c
deleted file mode 100644
index e20bd50..0000000
diff --git a/net/rds/iw_stats.c b/net/rds/iw_stats.c
deleted file mode 100644
index 5fe67f6..0000000
diff --git a/net/rds/iw_sysctl.c b/net/rds/iw_sysctl.c
deleted file mode 100644
index 139239d..0000000
diff --git a/net/rds/rdma_transport.c b/net/rds/rdma_transport.c
index 9c1fed8..4f4b3d8 100644
--- a/net/rds/rdma_transport.c
+++ b/net/rds/rdma_transport.c
@@ -49,9 +49,7 @@ int rds_rdma_cm_event_handler(struct rdma_cm_id *cm_id,
rdsdebug("conn %p id %p handling event %u (%s)\n", conn, cm_id,
event->event, rdma_event_msg(event->event));
- if (cm_id->device->node_type == RDMA_NODE_RNIC)
- trans = &rds_iw_transport;
- else
+ if (cm_id->device->node_type == RDMA_NODE_IB_CA)
trans = &rds_ib_transport;
/* Prevent shutdown from tearing down the connection
@@ -200,10 +198,6 @@ static int rds_rdma_init(void)
if (ret)
goto out;
- ret = rds_iw_init();
- if (ret)
- goto err_iw_init;
-
ret = rds_ib_init();
if (ret)
goto err_ib_init;
@@ -211,8 +205,6 @@ static int rds_rdma_init(void)
goto out;
err_ib_init:
- rds_iw_exit();
-err_iw_init:
rds_rdma_listen_stop();
out:
return ret;
@@ -224,11 +216,10 @@ static void rds_rdma_exit(void)
/* stop listening first to ensure no new connections are attempted */
rds_rdma_listen_stop();
rds_ib_exit();
- rds_iw_exit();
}
module_exit(rds_rdma_exit);
MODULE_AUTHOR("Oracle Corporation <rds-devel-N0ozoZBvEnrZJqsBc5GL+g@public.gmane.org>");
-MODULE_DESCRIPTION("RDS: IB/iWARP transport");
+MODULE_DESCRIPTION("RDS: IB transport");
MODULE_LICENSE("Dual BSD/GPL");
diff --git a/net/rds/rdma_transport.h b/net/rds/rdma_transport.h
index faba4e3..ff2010e 100644
--- a/net/rds/rdma_transport.h
+++ b/net/rds/rdma_transport.h
@@ -16,9 +16,4 @@ extern struct rds_transport rds_ib_transport;
int rds_ib_init(void);
void rds_ib_exit(void);
-/* from iw.c */
-extern struct rds_transport rds_iw_transport;
-int rds_iw_init(void);
-void rds_iw_exit(void);
-
#endif
--
1.9.1
--
To unsubscribe from this list: send the line "unsubscribe linux-rdma" in
the body of a message to majordomo-u79uwXL29TY76Z2rM5mHXA@public.gmane.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
^ permalink raw reply related [flat|nested] 18+ messages in thread* [net-next][PATCH 04/13] RDS: IB: Remove the RDS_IB_SEND_OP dependency
[not found] ` <1455967804-27644-1-git-send-email-santosh.shilimkar-QHcLZuEGTsvQT0dZR+AlfA@public.gmane.org>
2016-02-20 11:29 ` [net-next][PATCH 01/13] RDS: Drop stale iWARP RDMA transport Santosh Shilimkar
@ 2016-02-20 11:29 ` Santosh Shilimkar
2016-02-20 11:29 ` [net-next][PATCH 05/13] RDS: IB: Re-organise ibmr code Santosh Shilimkar
` (6 subsequent siblings)
8 siblings, 0 replies; 18+ messages in thread
From: Santosh Shilimkar @ 2016-02-20 11:29 UTC (permalink / raw)
To: netdev-u79uwXL29TY76Z2rM5mHXA, davem-fT/PcQaiUtIeIZ0/mPfg9Q
Cc: linux-rdma-u79uwXL29TY76Z2rM5mHXA,
linux-kernel-u79uwXL29TY76Z2rM5mHXA, Santosh Shilimkar
This helps to combine asynchronous fastreg MR completion handler
with send completion handler.
No functional change.
Signed-off-by: Santosh Shilimkar <ssantosh-DgEjT+Ai2ygdnm+yROfE0A@public.gmane.org>
Signed-off-by: Santosh Shilimkar <santosh.shilimkar-QHcLZuEGTsvQT0dZR+AlfA@public.gmane.org>
---
net/rds/ib.h | 1 -
net/rds/ib_cm.c | 42 +++++++++++++++++++++++++++---------------
net/rds/ib_send.c | 6 ++----
3 files changed, 29 insertions(+), 20 deletions(-)
diff --git a/net/rds/ib.h b/net/rds/ib.h
index b3fdebb..09cd8e3 100644
--- a/net/rds/ib.h
+++ b/net/rds/ib.h
@@ -28,7 +28,6 @@
#define RDS_IB_RECYCLE_BATCH_COUNT 32
#define RDS_IB_WC_MAX 32
-#define RDS_IB_SEND_OP BIT_ULL(63)
extern struct rw_semaphore rds_ib_devices_lock;
extern struct list_head rds_ib_devices;
diff --git a/net/rds/ib_cm.c b/net/rds/ib_cm.c
index da5a7fb..7f68abc 100644
--- a/net/rds/ib_cm.c
+++ b/net/rds/ib_cm.c
@@ -236,12 +236,10 @@ static void rds_ib_cq_comp_handler_recv(struct ib_cq *cq, void *context)
tasklet_schedule(&ic->i_recv_tasklet);
}
-static void poll_cq(struct rds_ib_connection *ic, struct ib_cq *cq,
- struct ib_wc *wcs,
- struct rds_ib_ack_state *ack_state)
+static void poll_scq(struct rds_ib_connection *ic, struct ib_cq *cq,
+ struct ib_wc *wcs)
{
- int nr;
- int i;
+ int nr, i;
struct ib_wc *wc;
while ((nr = ib_poll_cq(cq, RDS_IB_WC_MAX, wcs)) > 0) {
@@ -251,10 +249,7 @@ static void poll_cq(struct rds_ib_connection *ic, struct ib_cq *cq,
(unsigned long long)wc->wr_id, wc->status,
wc->byte_len, be32_to_cpu(wc->ex.imm_data));
- if (wc->wr_id & RDS_IB_SEND_OP)
- rds_ib_send_cqe_handler(ic, wc);
- else
- rds_ib_recv_cqe_handler(ic, wc, ack_state);
+ rds_ib_send_cqe_handler(ic, wc);
}
}
}
@@ -263,14 +258,12 @@ static void rds_ib_tasklet_fn_send(unsigned long data)
{
struct rds_ib_connection *ic = (struct rds_ib_connection *)data;
struct rds_connection *conn = ic->conn;
- struct rds_ib_ack_state state;
rds_ib_stats_inc(s_ib_tasklet_call);
- memset(&state, 0, sizeof(state));
- poll_cq(ic, ic->i_send_cq, ic->i_send_wc, &state);
+ poll_scq(ic, ic->i_send_cq, ic->i_send_wc);
ib_req_notify_cq(ic->i_send_cq, IB_CQ_NEXT_COMP);
- poll_cq(ic, ic->i_send_cq, ic->i_send_wc, &state);
+ poll_scq(ic, ic->i_send_cq, ic->i_send_wc);
if (rds_conn_up(conn) &&
(!test_bit(RDS_LL_SEND_FULL, &conn->c_flags) ||
@@ -278,6 +271,25 @@ static void rds_ib_tasklet_fn_send(unsigned long data)
rds_send_xmit(ic->conn);
}
+static void poll_rcq(struct rds_ib_connection *ic, struct ib_cq *cq,
+ struct ib_wc *wcs,
+ struct rds_ib_ack_state *ack_state)
+{
+ int nr, i;
+ struct ib_wc *wc;
+
+ while ((nr = ib_poll_cq(cq, RDS_IB_WC_MAX, wcs)) > 0) {
+ for (i = 0; i < nr; i++) {
+ wc = wcs + i;
+ rdsdebug("wc wr_id 0x%llx status %u byte_len %u imm_data %u\n",
+ (unsigned long long)wc->wr_id, wc->status,
+ wc->byte_len, be32_to_cpu(wc->ex.imm_data));
+
+ rds_ib_recv_cqe_handler(ic, wc, ack_state);
+ }
+ }
+}
+
static void rds_ib_tasklet_fn_recv(unsigned long data)
{
struct rds_ib_connection *ic = (struct rds_ib_connection *)data;
@@ -291,9 +303,9 @@ static void rds_ib_tasklet_fn_recv(unsigned long data)
rds_ib_stats_inc(s_ib_tasklet_call);
memset(&state, 0, sizeof(state));
- poll_cq(ic, ic->i_recv_cq, ic->i_recv_wc, &state);
+ poll_rcq(ic, ic->i_recv_cq, ic->i_recv_wc, &state);
ib_req_notify_cq(ic->i_recv_cq, IB_CQ_SOLICITED);
- poll_cq(ic, ic->i_recv_cq, ic->i_recv_wc, &state);
+ poll_rcq(ic, ic->i_recv_cq, ic->i_recv_wc, &state);
if (state.ack_next_valid)
rds_ib_set_ack(ic, state.ack_next, state.ack_required);
diff --git a/net/rds/ib_send.c b/net/rds/ib_send.c
index eac30bf..f27d2c8 100644
--- a/net/rds/ib_send.c
+++ b/net/rds/ib_send.c
@@ -195,7 +195,7 @@ void rds_ib_send_init_ring(struct rds_ib_connection *ic)
send->s_op = NULL;
- send->s_wr.wr_id = i | RDS_IB_SEND_OP;
+ send->s_wr.wr_id = i;
send->s_wr.sg_list = send->s_sge;
send->s_wr.ex.imm_data = 0;
@@ -263,9 +263,7 @@ void rds_ib_send_cqe_handler(struct rds_ib_connection *ic, struct ib_wc *wc)
oldest = rds_ib_ring_oldest(&ic->i_send_ring);
- completed = rds_ib_ring_completed(&ic->i_send_ring,
- (wc->wr_id & ~RDS_IB_SEND_OP),
- oldest);
+ completed = rds_ib_ring_completed(&ic->i_send_ring, wc->wr_id, oldest);
for (i = 0; i < completed; i++) {
send = &ic->i_sends[oldest];
--
1.9.1
--
To unsubscribe from this list: send the line "unsubscribe linux-rdma" in
the body of a message to majordomo-u79uwXL29TY76Z2rM5mHXA@public.gmane.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
^ permalink raw reply related [flat|nested] 18+ messages in thread* [net-next][PATCH 05/13] RDS: IB: Re-organise ibmr code
[not found] ` <1455967804-27644-1-git-send-email-santosh.shilimkar-QHcLZuEGTsvQT0dZR+AlfA@public.gmane.org>
2016-02-20 11:29 ` [net-next][PATCH 01/13] RDS: Drop stale iWARP RDMA transport Santosh Shilimkar
2016-02-20 11:29 ` [net-next][PATCH 04/13] RDS: IB: Remove the RDS_IB_SEND_OP dependency Santosh Shilimkar
@ 2016-02-20 11:29 ` Santosh Shilimkar
2016-02-20 11:29 ` [net-next][PATCH 06/13] RDS: IB: create struct rds_ib_fmr Santosh Shilimkar
` (5 subsequent siblings)
8 siblings, 0 replies; 18+ messages in thread
From: Santosh Shilimkar @ 2016-02-20 11:29 UTC (permalink / raw)
To: netdev-u79uwXL29TY76Z2rM5mHXA, davem-fT/PcQaiUtIeIZ0/mPfg9Q
Cc: linux-rdma-u79uwXL29TY76Z2rM5mHXA,
linux-kernel-u79uwXL29TY76Z2rM5mHXA, Santosh Shilimkar
No functional changes. This is in preperation towards adding
fastreg memory resgitration support.
Signed-off-by: Santosh Shilimkar <ssantosh-DgEjT+Ai2ygdnm+yROfE0A@public.gmane.org>
Signed-off-by: Santosh Shilimkar <santosh.shilimkar-QHcLZuEGTsvQT0dZR+AlfA@public.gmane.org>
---
net/rds/Makefile | 2 +-
net/rds/ib.c | 37 +++---
net/rds/ib.h | 25 +---
net/rds/ib_fmr.c | 217 +++++++++++++++++++++++++++++++
net/rds/ib_mr.h | 109 ++++++++++++++++
net/rds/ib_rdma.c | 379 +++++++++++-------------------------------------------
6 files changed, 422 insertions(+), 347 deletions(-)
create mode 100644 net/rds/ib_fmr.c
create mode 100644 net/rds/ib_mr.h
diff --git a/net/rds/Makefile b/net/rds/Makefile
index 19e5485..bcf5591 100644
--- a/net/rds/Makefile
+++ b/net/rds/Makefile
@@ -6,7 +6,7 @@ rds-y := af_rds.o bind.o cong.o connection.o info.o message.o \
obj-$(CONFIG_RDS_RDMA) += rds_rdma.o
rds_rdma-y := rdma_transport.o \
ib.o ib_cm.o ib_recv.o ib_ring.o ib_send.o ib_stats.o \
- ib_sysctl.o ib_rdma.o
+ ib_sysctl.o ib_rdma.o ib_fmr.o
obj-$(CONFIG_RDS_TCP) += rds_tcp.o
diff --git a/net/rds/ib.c b/net/rds/ib.c
index 9481d55..bb32cb9 100644
--- a/net/rds/ib.c
+++ b/net/rds/ib.c
@@ -42,15 +42,16 @@
#include "rds.h"
#include "ib.h"
+#include "ib_mr.h"
-unsigned int rds_ib_fmr_1m_pool_size = RDS_FMR_1M_POOL_SIZE;
-unsigned int rds_ib_fmr_8k_pool_size = RDS_FMR_8K_POOL_SIZE;
+unsigned int rds_ib_mr_1m_pool_size = RDS_MR_1M_POOL_SIZE;
+unsigned int rds_ib_mr_8k_pool_size = RDS_MR_8K_POOL_SIZE;
unsigned int rds_ib_retry_count = RDS_IB_DEFAULT_RETRY_COUNT;
-module_param(rds_ib_fmr_1m_pool_size, int, 0444);
-MODULE_PARM_DESC(rds_ib_fmr_1m_pool_size, " Max number of 1M fmr per HCA");
-module_param(rds_ib_fmr_8k_pool_size, int, 0444);
-MODULE_PARM_DESC(rds_ib_fmr_8k_pool_size, " Max number of 8K fmr per HCA");
+module_param(rds_ib_mr_1m_pool_size, int, 0444);
+MODULE_PARM_DESC(rds_ib_mr_1m_pool_size, " Max number of 1M mr per HCA");
+module_param(rds_ib_mr_8k_pool_size, int, 0444);
+MODULE_PARM_DESC(rds_ib_mr_8k_pool_size, " Max number of 8K mr per HCA");
module_param(rds_ib_retry_count, int, 0444);
MODULE_PARM_DESC(rds_ib_retry_count, " Number of hw retries before reporting an error");
@@ -140,13 +141,13 @@ static void rds_ib_add_one(struct ib_device *device)
rds_ibdev->max_sge = min(device->attrs.max_sge, RDS_IB_MAX_SGE);
rds_ibdev->fmr_max_remaps = device->attrs.max_map_per_fmr?: 32;
- rds_ibdev->max_1m_fmrs = device->attrs.max_mr ?
+ rds_ibdev->max_1m_mrs = device->attrs.max_mr ?
min_t(unsigned int, (device->attrs.max_mr / 2),
- rds_ib_fmr_1m_pool_size) : rds_ib_fmr_1m_pool_size;
+ rds_ib_mr_1m_pool_size) : rds_ib_mr_1m_pool_size;
- rds_ibdev->max_8k_fmrs = device->attrs.max_mr ?
+ rds_ibdev->max_8k_mrs = device->attrs.max_mr ?
min_t(unsigned int, ((device->attrs.max_mr / 2) * RDS_MR_8K_SCALE),
- rds_ib_fmr_8k_pool_size) : rds_ib_fmr_8k_pool_size;
+ rds_ib_mr_8k_pool_size) : rds_ib_mr_8k_pool_size;
rds_ibdev->max_initiator_depth = device->attrs.max_qp_init_rd_atom;
rds_ibdev->max_responder_resources = device->attrs.max_qp_rd_atom;
@@ -172,10 +173,10 @@ static void rds_ib_add_one(struct ib_device *device)
goto put_dev;
}
- rdsdebug("RDS/IB: max_mr = %d, max_wrs = %d, max_sge = %d, fmr_max_remaps = %d, max_1m_fmrs = %d, max_8k_fmrs = %d\n",
+ rdsdebug("RDS/IB: max_mr = %d, max_wrs = %d, max_sge = %d, fmr_max_remaps = %d, max_1m_mrs = %d, max_8k_mrs = %d\n",
device->attrs.max_fmr, rds_ibdev->max_wrs, rds_ibdev->max_sge,
- rds_ibdev->fmr_max_remaps, rds_ibdev->max_1m_fmrs,
- rds_ibdev->max_8k_fmrs);
+ rds_ibdev->fmr_max_remaps, rds_ibdev->max_1m_mrs,
+ rds_ibdev->max_8k_mrs);
INIT_LIST_HEAD(&rds_ibdev->ipaddr_list);
INIT_LIST_HEAD(&rds_ibdev->conn_list);
@@ -364,7 +365,7 @@ void rds_ib_exit(void)
rds_ib_sysctl_exit();
rds_ib_recv_exit();
rds_trans_unregister(&rds_ib_transport);
- rds_ib_fmr_exit();
+ rds_ib_mr_exit();
}
struct rds_transport rds_ib_transport = {
@@ -400,13 +401,13 @@ int rds_ib_init(void)
INIT_LIST_HEAD(&rds_ib_devices);
- ret = rds_ib_fmr_init();
+ ret = rds_ib_mr_init();
if (ret)
goto out;
ret = ib_register_client(&rds_ib_client);
if (ret)
- goto out_fmr_exit;
+ goto out_mr_exit;
ret = rds_ib_sysctl_init();
if (ret)
@@ -430,8 +431,8 @@ out_sysctl:
rds_ib_sysctl_exit();
out_ibreg:
rds_ib_unregister_client();
-out_fmr_exit:
- rds_ib_fmr_exit();
+out_mr_exit:
+ rds_ib_mr_exit();
out:
return ret;
}
diff --git a/net/rds/ib.h b/net/rds/ib.h
index 09cd8e3..c88cb22 100644
--- a/net/rds/ib.h
+++ b/net/rds/ib.h
@@ -9,12 +9,6 @@
#include "rds.h"
#include "rdma_transport.h"
-#define RDS_FMR_1M_POOL_SIZE (8192 / 2)
-#define RDS_FMR_1M_MSG_SIZE 256
-#define RDS_FMR_8K_MSG_SIZE 2
-#define RDS_MR_8K_SCALE (256 / (RDS_FMR_8K_MSG_SIZE + 1))
-#define RDS_FMR_8K_POOL_SIZE (RDS_MR_8K_SCALE * (8192 / 2))
-
#define RDS_IB_MAX_SGE 8
#define RDS_IB_RECV_SGE 2
@@ -206,12 +200,12 @@ struct rds_ib_device {
struct list_head conn_list;
struct ib_device *dev;
struct ib_pd *pd;
- unsigned int max_fmrs;
+ unsigned int max_mrs;
struct rds_ib_mr_pool *mr_1m_pool;
struct rds_ib_mr_pool *mr_8k_pool;
unsigned int fmr_max_remaps;
- unsigned int max_8k_fmrs;
- unsigned int max_1m_fmrs;
+ unsigned int max_8k_mrs;
+ unsigned int max_1m_mrs;
int max_sge;
unsigned int max_wrs;
unsigned int max_initiator_depth;
@@ -316,8 +310,6 @@ struct rds_ib_device *rds_ib_get_client_data(struct ib_device *device);
void rds_ib_dev_put(struct rds_ib_device *rds_ibdev);
extern struct ib_client rds_ib_client;
-extern unsigned int rds_ib_fmr_1m_pool_size;
-extern unsigned int rds_ib_fmr_8k_pool_size;
extern unsigned int rds_ib_retry_count;
extern spinlock_t ib_nodev_conns_lock;
@@ -347,17 +339,6 @@ int rds_ib_update_ipaddr(struct rds_ib_device *rds_ibdev, __be32 ipaddr);
void rds_ib_add_conn(struct rds_ib_device *rds_ibdev, struct rds_connection *conn);
void rds_ib_remove_conn(struct rds_ib_device *rds_ibdev, struct rds_connection *conn);
void rds_ib_destroy_nodev_conns(void);
-struct rds_ib_mr_pool *rds_ib_create_mr_pool(struct rds_ib_device *rds_dev,
- int npages);
-void rds_ib_get_mr_info(struct rds_ib_device *rds_ibdev, struct rds_info_rdma_connection *iinfo);
-void rds_ib_destroy_mr_pool(struct rds_ib_mr_pool *);
-void *rds_ib_get_mr(struct scatterlist *sg, unsigned long nents,
- struct rds_sock *rs, u32 *key_ret);
-void rds_ib_sync_mr(void *trans_private, int dir);
-void rds_ib_free_mr(void *trans_private, int invalidate);
-void rds_ib_flush_mrs(void);
-int rds_ib_fmr_init(void);
-void rds_ib_fmr_exit(void);
/* ib_recv.c */
int rds_ib_recv_init(void);
diff --git a/net/rds/ib_fmr.c b/net/rds/ib_fmr.c
new file mode 100644
index 0000000..d4f200d
--- /dev/null
+++ b/net/rds/ib_fmr.c
@@ -0,0 +1,217 @@
+/*
+ * Copyright (c) 2016 Oracle. All rights reserved.
+ *
+ * This software is available to you under a choice of one of two
+ * licenses. You may choose to be licensed under the terms of the GNU
+ * General Public License (GPL) Version 2, available from the file
+ * COPYING in the main directory of this source tree, or the
+ * OpenIB.org BSD license below:
+ *
+ * Redistribution and use in source and binary forms, with or
+ * without modification, are permitted provided that the following
+ * conditions are met:
+ *
+ * - Redistributions of source code must retain the above
+ * copyright notice, this list of conditions and the following
+ * disclaimer.
+ *
+ * - Redistributions in binary form must reproduce the above
+ * copyright notice, this list of conditions and the following
+ * disclaimer in the documentation and/or other materials
+ * provided with the distribution.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
+ * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
+ * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
+ * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
+ * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
+ * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
+ * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+
+#include "ib_mr.h"
+
+struct rds_ib_mr *rds_ib_alloc_fmr(struct rds_ib_device *rds_ibdev, int npages)
+{
+ struct rds_ib_mr_pool *pool;
+ struct rds_ib_mr *ibmr = NULL;
+ int err = 0, iter = 0;
+
+ if (npages <= RDS_MR_8K_MSG_SIZE)
+ pool = rds_ibdev->mr_8k_pool;
+ else
+ pool = rds_ibdev->mr_1m_pool;
+
+ if (atomic_read(&pool->dirty_count) >= pool->max_items / 10)
+ queue_delayed_work(rds_ib_mr_wq, &pool->flush_worker, 10);
+
+ /* Switch pools if one of the pool is reaching upper limit */
+ if (atomic_read(&pool->dirty_count) >= pool->max_items * 9 / 10) {
+ if (pool->pool_type == RDS_IB_MR_8K_POOL)
+ pool = rds_ibdev->mr_1m_pool;
+ else
+ pool = rds_ibdev->mr_8k_pool;
+ }
+
+ while (1) {
+ ibmr = rds_ib_reuse_mr(pool);
+ if (ibmr)
+ return ibmr;
+
+ /* No clean MRs - now we have the choice of either
+ * allocating a fresh MR up to the limit imposed by the
+ * driver, or flush any dirty unused MRs.
+ * We try to avoid stalling in the send path if possible,
+ * so we allocate as long as we're allowed to.
+ *
+ * We're fussy with enforcing the FMR limit, though. If the
+ * driver tells us we can't use more than N fmrs, we shouldn't
+ * start arguing with it
+ */
+ if (atomic_inc_return(&pool->item_count) <= pool->max_items)
+ break;
+
+ atomic_dec(&pool->item_count);
+
+ if (++iter > 2) {
+ if (pool->pool_type == RDS_IB_MR_8K_POOL)
+ rds_ib_stats_inc(s_ib_rdma_mr_8k_pool_depleted);
+ else
+ rds_ib_stats_inc(s_ib_rdma_mr_1m_pool_depleted);
+ return ERR_PTR(-EAGAIN);
+ }
+
+ /* We do have some empty MRs. Flush them out. */
+ if (pool->pool_type == RDS_IB_MR_8K_POOL)
+ rds_ib_stats_inc(s_ib_rdma_mr_8k_pool_wait);
+ else
+ rds_ib_stats_inc(s_ib_rdma_mr_1m_pool_wait);
+ rds_ib_flush_mr_pool(pool, 0, &ibmr);
+ if (ibmr)
+ return ibmr;
+ }
+
+ ibmr = kzalloc_node(sizeof(*ibmr), GFP_KERNEL,
+ rdsibdev_to_node(rds_ibdev));
+ if (!ibmr) {
+ err = -ENOMEM;
+ goto out_no_cigar;
+ }
+
+ ibmr->fmr = ib_alloc_fmr(rds_ibdev->pd,
+ (IB_ACCESS_LOCAL_WRITE |
+ IB_ACCESS_REMOTE_READ |
+ IB_ACCESS_REMOTE_WRITE |
+ IB_ACCESS_REMOTE_ATOMIC),
+ &pool->fmr_attr);
+ if (IS_ERR(ibmr->fmr)) {
+ err = PTR_ERR(ibmr->fmr);
+ ibmr->fmr = NULL;
+ pr_warn("RDS/IB: %s failed (err=%d)\n", __func__, err);
+ goto out_no_cigar;
+ }
+
+ ibmr->pool = pool;
+ if (pool->pool_type == RDS_IB_MR_8K_POOL)
+ rds_ib_stats_inc(s_ib_rdma_mr_8k_alloc);
+ else
+ rds_ib_stats_inc(s_ib_rdma_mr_1m_alloc);
+
+ return ibmr;
+
+out_no_cigar:
+ if (ibmr) {
+ if (ibmr->fmr)
+ ib_dealloc_fmr(ibmr->fmr);
+ kfree(ibmr);
+ }
+ atomic_dec(&pool->item_count);
+ return ERR_PTR(err);
+}
+
+int rds_ib_map_fmr(struct rds_ib_device *rds_ibdev, struct rds_ib_mr *ibmr,
+ struct scatterlist *sg, unsigned int nents)
+{
+ struct ib_device *dev = rds_ibdev->dev;
+ struct scatterlist *scat = sg;
+ u64 io_addr = 0;
+ u64 *dma_pages;
+ u32 len;
+ int page_cnt, sg_dma_len;
+ int i, j;
+ int ret;
+
+ sg_dma_len = ib_dma_map_sg(dev, sg, nents, DMA_BIDIRECTIONAL);
+ if (unlikely(!sg_dma_len)) {
+ pr_warn("RDS/IB: %s failed!\n", __func__);
+ return -EBUSY;
+ }
+
+ len = 0;
+ page_cnt = 0;
+
+ for (i = 0; i < sg_dma_len; ++i) {
+ unsigned int dma_len = ib_sg_dma_len(dev, &scat[i]);
+ u64 dma_addr = ib_sg_dma_address(dev, &scat[i]);
+
+ if (dma_addr & ~PAGE_MASK) {
+ if (i > 0)
+ return -EINVAL;
+ else
+ ++page_cnt;
+ }
+ if ((dma_addr + dma_len) & ~PAGE_MASK) {
+ if (i < sg_dma_len - 1)
+ return -EINVAL;
+ else
+ ++page_cnt;
+ }
+
+ len += dma_len;
+ }
+
+ page_cnt += len >> PAGE_SHIFT;
+ if (page_cnt > ibmr->pool->fmr_attr.max_pages)
+ return -EINVAL;
+
+ dma_pages = kmalloc_node(sizeof(u64) * page_cnt, GFP_ATOMIC,
+ rdsibdev_to_node(rds_ibdev));
+ if (!dma_pages)
+ return -ENOMEM;
+
+ page_cnt = 0;
+ for (i = 0; i < sg_dma_len; ++i) {
+ unsigned int dma_len = ib_sg_dma_len(dev, &scat[i]);
+ u64 dma_addr = ib_sg_dma_address(dev, &scat[i]);
+
+ for (j = 0; j < dma_len; j += PAGE_SIZE)
+ dma_pages[page_cnt++] =
+ (dma_addr & PAGE_MASK) + j;
+ }
+
+ ret = ib_map_phys_fmr(ibmr->fmr, dma_pages, page_cnt, io_addr);
+ if (ret)
+ goto out;
+
+ /* Success - we successfully remapped the MR, so we can
+ * safely tear down the old mapping.
+ */
+ rds_ib_teardown_mr(ibmr);
+
+ ibmr->sg = scat;
+ ibmr->sg_len = nents;
+ ibmr->sg_dma_len = sg_dma_len;
+ ibmr->remap_count++;
+
+ if (ibmr->pool->pool_type == RDS_IB_MR_8K_POOL)
+ rds_ib_stats_inc(s_ib_rdma_mr_8k_used);
+ else
+ rds_ib_stats_inc(s_ib_rdma_mr_1m_used);
+ ret = 0;
+
+out:
+ kfree(dma_pages);
+
+ return ret;
+}
diff --git a/net/rds/ib_mr.h b/net/rds/ib_mr.h
new file mode 100644
index 0000000..d88724f
--- /dev/null
+++ b/net/rds/ib_mr.h
@@ -0,0 +1,109 @@
+/*
+ * Copyright (c) 2016 Oracle. All rights reserved.
+ *
+ * This software is available to you under a choice of one of two
+ * licenses. You may choose to be licensed under the terms of the GNU
+ * General Public License (GPL) Version 2, available from the file
+ * COPYING in the main directory of this source tree, or the
+ * OpenIB.org BSD license below:
+ *
+ * Redistribution and use in source and binary forms, with or
+ * without modification, are permitted provided that the following
+ * conditions are met:
+ *
+ * - Redistributions of source code must retain the above
+ * copyright notice, this list of conditions and the following
+ * disclaimer.
+ *
+ * - Redistributions in binary form must reproduce the above
+ * copyright notice, this list of conditions and the following
+ * disclaimer in the documentation and/or other materials
+ * provided with the distribution.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
+ * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
+ * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
+ * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
+ * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
+ * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
+ * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+#ifndef _RDS_IB_MR_H
+#define _RDS_IB_MR_H
+
+#include <linux/kernel.h>
+
+#include "rds.h"
+#include "ib.h"
+
+#define RDS_MR_1M_POOL_SIZE (8192 / 2)
+#define RDS_MR_1M_MSG_SIZE 256
+#define RDS_MR_8K_MSG_SIZE 2
+#define RDS_MR_8K_SCALE (256 / (RDS_MR_8K_MSG_SIZE + 1))
+#define RDS_MR_8K_POOL_SIZE (RDS_MR_8K_SCALE * (8192 / 2))
+
+/* This is stored as mr->r_trans_private. */
+struct rds_ib_mr {
+ struct rds_ib_device *device;
+ struct rds_ib_mr_pool *pool;
+ struct ib_fmr *fmr;
+
+ struct llist_node llnode;
+
+ /* unmap_list is for freeing */
+ struct list_head unmap_list;
+ unsigned int remap_count;
+
+ struct scatterlist *sg;
+ unsigned int sg_len;
+ u64 *dma;
+ int sg_dma_len;
+};
+
+/* Our own little MR pool */
+struct rds_ib_mr_pool {
+ unsigned int pool_type;
+ struct mutex flush_lock; /* serialize fmr invalidate */
+ struct delayed_work flush_worker; /* flush worker */
+
+ atomic_t item_count; /* total # of MRs */
+ atomic_t dirty_count; /* # dirty of MRs */
+
+ struct llist_head drop_list; /* MRs not reached max_maps */
+ struct llist_head free_list; /* unused MRs */
+ struct llist_head clean_list; /* unused & unmapped MRs */
+ wait_queue_head_t flush_wait;
+
+ atomic_t free_pinned; /* memory pinned by free MRs */
+ unsigned long max_items;
+ unsigned long max_items_soft;
+ unsigned long max_free_pinned;
+ struct ib_fmr_attr fmr_attr;
+};
+
+extern struct workqueue_struct *rds_ib_mr_wq;
+extern unsigned int rds_ib_mr_1m_pool_size;
+extern unsigned int rds_ib_mr_8k_pool_size;
+
+struct rds_ib_mr_pool *rds_ib_create_mr_pool(struct rds_ib_device *rds_dev,
+ int npages);
+void rds_ib_get_mr_info(struct rds_ib_device *rds_ibdev,
+ struct rds_info_rdma_connection *iinfo);
+void rds_ib_destroy_mr_pool(struct rds_ib_mr_pool *);
+void *rds_ib_get_mr(struct scatterlist *sg, unsigned long nents,
+ struct rds_sock *rs, u32 *key_ret);
+void rds_ib_sync_mr(void *trans_private, int dir);
+void rds_ib_free_mr(void *trans_private, int invalidate);
+void rds_ib_flush_mrs(void);
+int rds_ib_mr_init(void);
+void rds_ib_mr_exit(void);
+
+void __rds_ib_teardown_mr(struct rds_ib_mr *);
+void rds_ib_teardown_mr(struct rds_ib_mr *);
+struct rds_ib_mr *rds_ib_alloc_fmr(struct rds_ib_device *, int);
+int rds_ib_map_fmr(struct rds_ib_device *, struct rds_ib_mr *,
+ struct scatterlist *, unsigned int);
+struct rds_ib_mr *rds_ib_reuse_mr(struct rds_ib_mr_pool *);
+int rds_ib_flush_mr_pool(struct rds_ib_mr_pool *, int, struct rds_ib_mr **);
+#endif
diff --git a/net/rds/ib_rdma.c b/net/rds/ib_rdma.c
index a234074..c594519 100644
--- a/net/rds/ib_rdma.c
+++ b/net/rds/ib_rdma.c
@@ -35,78 +35,13 @@
#include <linux/rculist.h>
#include <linux/llist.h>
-#include "rds.h"
-#include "ib.h"
+#include "ib_mr.h"
+
+struct workqueue_struct *rds_ib_mr_wq;
static DEFINE_PER_CPU(unsigned long, clean_list_grace);
#define CLEAN_LIST_BUSY_BIT 0
-/*
- * This is stored as mr->r_trans_private.
- */
-struct rds_ib_mr {
- struct rds_ib_device *device;
- struct rds_ib_mr_pool *pool;
- struct ib_fmr *fmr;
-
- struct llist_node llnode;
-
- /* unmap_list is for freeing */
- struct list_head unmap_list;
- unsigned int remap_count;
-
- struct scatterlist *sg;
- unsigned int sg_len;
- u64 *dma;
- int sg_dma_len;
-};
-
-/*
- * Our own little FMR pool
- */
-struct rds_ib_mr_pool {
- unsigned int pool_type;
- struct mutex flush_lock; /* serialize fmr invalidate */
- struct delayed_work flush_worker; /* flush worker */
-
- atomic_t item_count; /* total # of MRs */
- atomic_t dirty_count; /* # dirty of MRs */
-
- struct llist_head drop_list; /* MRs that have reached their max_maps limit */
- struct llist_head free_list; /* unused MRs */
- struct llist_head clean_list; /* global unused & unamapped MRs */
- wait_queue_head_t flush_wait;
-
- atomic_t free_pinned; /* memory pinned by free MRs */
- unsigned long max_items;
- unsigned long max_items_soft;
- unsigned long max_free_pinned;
- struct ib_fmr_attr fmr_attr;
-};
-
-static struct workqueue_struct *rds_ib_fmr_wq;
-
-int rds_ib_fmr_init(void)
-{
- rds_ib_fmr_wq = create_workqueue("rds_fmr_flushd");
- if (!rds_ib_fmr_wq)
- return -ENOMEM;
- return 0;
-}
-
-/* By the time this is called all the IB devices should have been torn down and
- * had their pools freed. As each pool is freed its work struct is waited on,
- * so the pool flushing work queue should be idle by the time we get here.
- */
-void rds_ib_fmr_exit(void)
-{
- destroy_workqueue(rds_ib_fmr_wq);
-}
-
-static int rds_ib_flush_mr_pool(struct rds_ib_mr_pool *pool, int free_all, struct rds_ib_mr **);
-static void rds_ib_teardown_mr(struct rds_ib_mr *ibmr);
-static void rds_ib_mr_pool_flush_worker(struct work_struct *work);
-
static struct rds_ib_device *rds_ib_get_device(__be32 ipaddr)
{
struct rds_ib_device *rds_ibdev;
@@ -235,41 +170,6 @@ void rds_ib_destroy_nodev_conns(void)
rds_conn_destroy(ic->conn);
}
-struct rds_ib_mr_pool *rds_ib_create_mr_pool(struct rds_ib_device *rds_ibdev,
- int pool_type)
-{
- struct rds_ib_mr_pool *pool;
-
- pool = kzalloc(sizeof(*pool), GFP_KERNEL);
- if (!pool)
- return ERR_PTR(-ENOMEM);
-
- pool->pool_type = pool_type;
- init_llist_head(&pool->free_list);
- init_llist_head(&pool->drop_list);
- init_llist_head(&pool->clean_list);
- mutex_init(&pool->flush_lock);
- init_waitqueue_head(&pool->flush_wait);
- INIT_DELAYED_WORK(&pool->flush_worker, rds_ib_mr_pool_flush_worker);
-
- if (pool_type == RDS_IB_MR_1M_POOL) {
- /* +1 allows for unaligned MRs */
- pool->fmr_attr.max_pages = RDS_FMR_1M_MSG_SIZE + 1;
- pool->max_items = RDS_FMR_1M_POOL_SIZE;
- } else {
- /* pool_type == RDS_IB_MR_8K_POOL */
- pool->fmr_attr.max_pages = RDS_FMR_8K_MSG_SIZE + 1;
- pool->max_items = RDS_FMR_8K_POOL_SIZE;
- }
-
- pool->max_free_pinned = pool->max_items * pool->fmr_attr.max_pages / 4;
- pool->fmr_attr.max_maps = rds_ibdev->fmr_max_remaps;
- pool->fmr_attr.page_shift = PAGE_SHIFT;
- pool->max_items_soft = rds_ibdev->max_fmrs * 3 / 4;
-
- return pool;
-}
-
void rds_ib_get_mr_info(struct rds_ib_device *rds_ibdev, struct rds_info_rdma_connection *iinfo)
{
struct rds_ib_mr_pool *pool_1m = rds_ibdev->mr_1m_pool;
@@ -278,16 +178,7 @@ void rds_ib_get_mr_info(struct rds_ib_device *rds_ibdev, struct rds_info_rdma_co
iinfo->rdma_mr_size = pool_1m->fmr_attr.max_pages;
}
-void rds_ib_destroy_mr_pool(struct rds_ib_mr_pool *pool)
-{
- cancel_delayed_work_sync(&pool->flush_worker);
- rds_ib_flush_mr_pool(pool, 1, NULL);
- WARN_ON(atomic_read(&pool->item_count));
- WARN_ON(atomic_read(&pool->free_pinned));
- kfree(pool);
-}
-
-static inline struct rds_ib_mr *rds_ib_reuse_fmr(struct rds_ib_mr_pool *pool)
+struct rds_ib_mr *rds_ib_reuse_mr(struct rds_ib_mr_pool *pool)
{
struct rds_ib_mr *ibmr = NULL;
struct llist_node *ret;
@@ -317,190 +208,6 @@ static inline void wait_clean_list_grace(void)
}
}
-static struct rds_ib_mr *rds_ib_alloc_fmr(struct rds_ib_device *rds_ibdev,
- int npages)
-{
- struct rds_ib_mr_pool *pool;
- struct rds_ib_mr *ibmr = NULL;
- int err = 0, iter = 0;
-
- if (npages <= RDS_FMR_8K_MSG_SIZE)
- pool = rds_ibdev->mr_8k_pool;
- else
- pool = rds_ibdev->mr_1m_pool;
-
- if (atomic_read(&pool->dirty_count) >= pool->max_items / 10)
- queue_delayed_work(rds_ib_fmr_wq, &pool->flush_worker, 10);
-
- /* Switch pools if one of the pool is reaching upper limit */
- if (atomic_read(&pool->dirty_count) >= pool->max_items * 9 / 10) {
- if (pool->pool_type == RDS_IB_MR_8K_POOL)
- pool = rds_ibdev->mr_1m_pool;
- else
- pool = rds_ibdev->mr_8k_pool;
- }
-
- while (1) {
- ibmr = rds_ib_reuse_fmr(pool);
- if (ibmr)
- return ibmr;
-
- /* No clean MRs - now we have the choice of either
- * allocating a fresh MR up to the limit imposed by the
- * driver, or flush any dirty unused MRs.
- * We try to avoid stalling in the send path if possible,
- * so we allocate as long as we're allowed to.
- *
- * We're fussy with enforcing the FMR limit, though. If the driver
- * tells us we can't use more than N fmrs, we shouldn't start
- * arguing with it */
- if (atomic_inc_return(&pool->item_count) <= pool->max_items)
- break;
-
- atomic_dec(&pool->item_count);
-
- if (++iter > 2) {
- if (pool->pool_type == RDS_IB_MR_8K_POOL)
- rds_ib_stats_inc(s_ib_rdma_mr_8k_pool_depleted);
- else
- rds_ib_stats_inc(s_ib_rdma_mr_1m_pool_depleted);
- return ERR_PTR(-EAGAIN);
- }
-
- /* We do have some empty MRs. Flush them out. */
- if (pool->pool_type == RDS_IB_MR_8K_POOL)
- rds_ib_stats_inc(s_ib_rdma_mr_8k_pool_wait);
- else
- rds_ib_stats_inc(s_ib_rdma_mr_1m_pool_wait);
- rds_ib_flush_mr_pool(pool, 0, &ibmr);
- if (ibmr)
- return ibmr;
- }
-
- ibmr = kzalloc_node(sizeof(*ibmr), GFP_KERNEL, rdsibdev_to_node(rds_ibdev));
- if (!ibmr) {
- err = -ENOMEM;
- goto out_no_cigar;
- }
-
- ibmr->fmr = ib_alloc_fmr(rds_ibdev->pd,
- (IB_ACCESS_LOCAL_WRITE |
- IB_ACCESS_REMOTE_READ |
- IB_ACCESS_REMOTE_WRITE|
- IB_ACCESS_REMOTE_ATOMIC),
- &pool->fmr_attr);
- if (IS_ERR(ibmr->fmr)) {
- err = PTR_ERR(ibmr->fmr);
- ibmr->fmr = NULL;
- printk(KERN_WARNING "RDS/IB: ib_alloc_fmr failed (err=%d)\n", err);
- goto out_no_cigar;
- }
-
- ibmr->pool = pool;
- if (pool->pool_type == RDS_IB_MR_8K_POOL)
- rds_ib_stats_inc(s_ib_rdma_mr_8k_alloc);
- else
- rds_ib_stats_inc(s_ib_rdma_mr_1m_alloc);
-
- return ibmr;
-
-out_no_cigar:
- if (ibmr) {
- if (ibmr->fmr)
- ib_dealloc_fmr(ibmr->fmr);
- kfree(ibmr);
- }
- atomic_dec(&pool->item_count);
- return ERR_PTR(err);
-}
-
-static int rds_ib_map_fmr(struct rds_ib_device *rds_ibdev, struct rds_ib_mr *ibmr,
- struct scatterlist *sg, unsigned int nents)
-{
- struct ib_device *dev = rds_ibdev->dev;
- struct scatterlist *scat = sg;
- u64 io_addr = 0;
- u64 *dma_pages;
- u32 len;
- int page_cnt, sg_dma_len;
- int i, j;
- int ret;
-
- sg_dma_len = ib_dma_map_sg(dev, sg, nents,
- DMA_BIDIRECTIONAL);
- if (unlikely(!sg_dma_len)) {
- printk(KERN_WARNING "RDS/IB: dma_map_sg failed!\n");
- return -EBUSY;
- }
-
- len = 0;
- page_cnt = 0;
-
- for (i = 0; i < sg_dma_len; ++i) {
- unsigned int dma_len = ib_sg_dma_len(dev, &scat[i]);
- u64 dma_addr = ib_sg_dma_address(dev, &scat[i]);
-
- if (dma_addr & ~PAGE_MASK) {
- if (i > 0)
- return -EINVAL;
- else
- ++page_cnt;
- }
- if ((dma_addr + dma_len) & ~PAGE_MASK) {
- if (i < sg_dma_len - 1)
- return -EINVAL;
- else
- ++page_cnt;
- }
-
- len += dma_len;
- }
-
- page_cnt += len >> PAGE_SHIFT;
- if (page_cnt > ibmr->pool->fmr_attr.max_pages)
- return -EINVAL;
-
- dma_pages = kmalloc_node(sizeof(u64) * page_cnt, GFP_ATOMIC,
- rdsibdev_to_node(rds_ibdev));
- if (!dma_pages)
- return -ENOMEM;
-
- page_cnt = 0;
- for (i = 0; i < sg_dma_len; ++i) {
- unsigned int dma_len = ib_sg_dma_len(dev, &scat[i]);
- u64 dma_addr = ib_sg_dma_address(dev, &scat[i]);
-
- for (j = 0; j < dma_len; j += PAGE_SIZE)
- dma_pages[page_cnt++] =
- (dma_addr & PAGE_MASK) + j;
- }
-
- ret = ib_map_phys_fmr(ibmr->fmr,
- dma_pages, page_cnt, io_addr);
- if (ret)
- goto out;
-
- /* Success - we successfully remapped the MR, so we can
- * safely tear down the old mapping. */
- rds_ib_teardown_mr(ibmr);
-
- ibmr->sg = scat;
- ibmr->sg_len = nents;
- ibmr->sg_dma_len = sg_dma_len;
- ibmr->remap_count++;
-
- if (ibmr->pool->pool_type == RDS_IB_MR_8K_POOL)
- rds_ib_stats_inc(s_ib_rdma_mr_8k_used);
- else
- rds_ib_stats_inc(s_ib_rdma_mr_1m_used);
- ret = 0;
-
-out:
- kfree(dma_pages);
-
- return ret;
-}
-
void rds_ib_sync_mr(void *trans_private, int direction)
{
struct rds_ib_mr *ibmr = trans_private;
@@ -518,7 +225,7 @@ void rds_ib_sync_mr(void *trans_private, int direction)
}
}
-static void __rds_ib_teardown_mr(struct rds_ib_mr *ibmr)
+void __rds_ib_teardown_mr(struct rds_ib_mr *ibmr)
{
struct rds_ib_device *rds_ibdev = ibmr->device;
@@ -549,7 +256,7 @@ static void __rds_ib_teardown_mr(struct rds_ib_mr *ibmr)
}
}
-static void rds_ib_teardown_mr(struct rds_ib_mr *ibmr)
+void rds_ib_teardown_mr(struct rds_ib_mr *ibmr)
{
unsigned int pinned = ibmr->sg_len;
@@ -623,8 +330,8 @@ static void list_to_llist_nodes(struct rds_ib_mr_pool *pool,
* If the number of MRs allocated exceeds the limit, we also try
* to free as many MRs as needed to get back to this limit.
*/
-static int rds_ib_flush_mr_pool(struct rds_ib_mr_pool *pool,
- int free_all, struct rds_ib_mr **ibmr_ret)
+int rds_ib_flush_mr_pool(struct rds_ib_mr_pool *pool,
+ int free_all, struct rds_ib_mr **ibmr_ret)
{
struct rds_ib_mr *ibmr, *next;
struct llist_node *clean_nodes;
@@ -643,7 +350,7 @@ static int rds_ib_flush_mr_pool(struct rds_ib_mr_pool *pool,
if (ibmr_ret) {
DEFINE_WAIT(wait);
while (!mutex_trylock(&pool->flush_lock)) {
- ibmr = rds_ib_reuse_fmr(pool);
+ ibmr = rds_ib_reuse_mr(pool);
if (ibmr) {
*ibmr_ret = ibmr;
finish_wait(&pool->flush_wait, &wait);
@@ -655,7 +362,7 @@ static int rds_ib_flush_mr_pool(struct rds_ib_mr_pool *pool,
if (llist_empty(&pool->clean_list))
schedule();
- ibmr = rds_ib_reuse_fmr(pool);
+ ibmr = rds_ib_reuse_mr(pool);
if (ibmr) {
*ibmr_ret = ibmr;
finish_wait(&pool->flush_wait, &wait);
@@ -667,7 +374,7 @@ static int rds_ib_flush_mr_pool(struct rds_ib_mr_pool *pool,
mutex_lock(&pool->flush_lock);
if (ibmr_ret) {
- ibmr = rds_ib_reuse_fmr(pool);
+ ibmr = rds_ib_reuse_mr(pool);
if (ibmr) {
*ibmr_ret = ibmr;
goto out;
@@ -773,7 +480,7 @@ void rds_ib_free_mr(void *trans_private, int invalidate)
/* If we've pinned too many pages, request a flush */
if (atomic_read(&pool->free_pinned) >= pool->max_free_pinned ||
atomic_read(&pool->dirty_count) >= pool->max_items / 5)
- queue_delayed_work(rds_ib_fmr_wq, &pool->flush_worker, 10);
+ queue_delayed_work(rds_ib_mr_wq, &pool->flush_worker, 10);
if (invalidate) {
if (likely(!in_interrupt())) {
@@ -782,7 +489,7 @@ void rds_ib_free_mr(void *trans_private, int invalidate)
/* We get here if the user created a MR marked
* as use_once and invalidate at the same time.
*/
- queue_delayed_work(rds_ib_fmr_wq,
+ queue_delayed_work(rds_ib_mr_wq,
&pool->flush_worker, 10);
}
}
@@ -849,3 +556,63 @@ void *rds_ib_get_mr(struct scatterlist *sg, unsigned long nents,
return ibmr;
}
+void rds_ib_destroy_mr_pool(struct rds_ib_mr_pool *pool)
+{
+ cancel_delayed_work_sync(&pool->flush_worker);
+ rds_ib_flush_mr_pool(pool, 1, NULL);
+ WARN_ON(atomic_read(&pool->item_count));
+ WARN_ON(atomic_read(&pool->free_pinned));
+ kfree(pool);
+}
+
+struct rds_ib_mr_pool *rds_ib_create_mr_pool(struct rds_ib_device *rds_ibdev,
+ int pool_type)
+{
+ struct rds_ib_mr_pool *pool;
+
+ pool = kzalloc(sizeof(*pool), GFP_KERNEL);
+ if (!pool)
+ return ERR_PTR(-ENOMEM);
+
+ pool->pool_type = pool_type;
+ init_llist_head(&pool->free_list);
+ init_llist_head(&pool->drop_list);
+ init_llist_head(&pool->clean_list);
+ mutex_init(&pool->flush_lock);
+ init_waitqueue_head(&pool->flush_wait);
+ INIT_DELAYED_WORK(&pool->flush_worker, rds_ib_mr_pool_flush_worker);
+
+ if (pool_type == RDS_IB_MR_1M_POOL) {
+ /* +1 allows for unaligned MRs */
+ pool->fmr_attr.max_pages = RDS_MR_1M_MSG_SIZE + 1;
+ pool->max_items = RDS_MR_1M_POOL_SIZE;
+ } else {
+ /* pool_type == RDS_IB_MR_8K_POOL */
+ pool->fmr_attr.max_pages = RDS_MR_8K_MSG_SIZE + 1;
+ pool->max_items = RDS_MR_8K_POOL_SIZE;
+ }
+
+ pool->max_free_pinned = pool->max_items * pool->fmr_attr.max_pages / 4;
+ pool->fmr_attr.max_maps = rds_ibdev->fmr_max_remaps;
+ pool->fmr_attr.page_shift = PAGE_SHIFT;
+ pool->max_items_soft = rds_ibdev->max_mrs * 3 / 4;
+
+ return pool;
+}
+
+int rds_ib_mr_init(void)
+{
+ rds_ib_mr_wq = create_workqueue("rds_mr_flushd");
+ if (!rds_ib_mr_wq)
+ return -ENOMEM;
+ return 0;
+}
+
+/* By the time this is called all the IB devices should have been torn down and
+ * had their pools freed. As each pool is freed its work struct is waited on,
+ * so the pool flushing work queue should be idle by the time we get here.
+ */
+void rds_ib_mr_exit(void)
+{
+ destroy_workqueue(rds_ib_mr_wq);
+}
--
1.9.1
--
To unsubscribe from this list: send the line "unsubscribe linux-rdma" in
the body of a message to majordomo-u79uwXL29TY76Z2rM5mHXA@public.gmane.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
^ permalink raw reply related [flat|nested] 18+ messages in thread* [net-next][PATCH 06/13] RDS: IB: create struct rds_ib_fmr
[not found] ` <1455967804-27644-1-git-send-email-santosh.shilimkar-QHcLZuEGTsvQT0dZR+AlfA@public.gmane.org>
` (2 preceding siblings ...)
2016-02-20 11:29 ` [net-next][PATCH 05/13] RDS: IB: Re-organise ibmr code Santosh Shilimkar
@ 2016-02-20 11:29 ` Santosh Shilimkar
2016-02-20 11:29 ` [net-next][PATCH 08/13] RDS: IB: add connection info to ibmr Santosh Shilimkar
` (4 subsequent siblings)
8 siblings, 0 replies; 18+ messages in thread
From: Santosh Shilimkar @ 2016-02-20 11:29 UTC (permalink / raw)
To: netdev-u79uwXL29TY76Z2rM5mHXA, davem-fT/PcQaiUtIeIZ0/mPfg9Q
Cc: linux-rdma-u79uwXL29TY76Z2rM5mHXA,
linux-kernel-u79uwXL29TY76Z2rM5mHXA, Santosh Shilimkar
Keep fmr related filed in its own struct. Fastreg MR structure
will be added to the union.
Signed-off-by: Santosh Shilimkar <ssantosh-DgEjT+Ai2ygdnm+yROfE0A@public.gmane.org>
Signed-off-by: Santosh Shilimkar <santosh.shilimkar-QHcLZuEGTsvQT0dZR+AlfA@public.gmane.org>
---
net/rds/ib_fmr.c | 17 ++++++++++-------
net/rds/ib_mr.h | 11 +++++++++--
net/rds/ib_rdma.c | 14 ++++++++++----
3 files changed, 29 insertions(+), 13 deletions(-)
diff --git a/net/rds/ib_fmr.c b/net/rds/ib_fmr.c
index d4f200d..74f2c21 100644
--- a/net/rds/ib_fmr.c
+++ b/net/rds/ib_fmr.c
@@ -36,6 +36,7 @@ struct rds_ib_mr *rds_ib_alloc_fmr(struct rds_ib_device *rds_ibdev, int npages)
{
struct rds_ib_mr_pool *pool;
struct rds_ib_mr *ibmr = NULL;
+ struct rds_ib_fmr *fmr;
int err = 0, iter = 0;
if (npages <= RDS_MR_8K_MSG_SIZE)
@@ -99,15 +100,16 @@ struct rds_ib_mr *rds_ib_alloc_fmr(struct rds_ib_device *rds_ibdev, int npages)
goto out_no_cigar;
}
- ibmr->fmr = ib_alloc_fmr(rds_ibdev->pd,
+ fmr = &ibmr->u.fmr;
+ fmr->fmr = ib_alloc_fmr(rds_ibdev->pd,
(IB_ACCESS_LOCAL_WRITE |
IB_ACCESS_REMOTE_READ |
IB_ACCESS_REMOTE_WRITE |
IB_ACCESS_REMOTE_ATOMIC),
&pool->fmr_attr);
- if (IS_ERR(ibmr->fmr)) {
- err = PTR_ERR(ibmr->fmr);
- ibmr->fmr = NULL;
+ if (IS_ERR(fmr->fmr)) {
+ err = PTR_ERR(fmr->fmr);
+ fmr->fmr = NULL;
pr_warn("RDS/IB: %s failed (err=%d)\n", __func__, err);
goto out_no_cigar;
}
@@ -122,8 +124,8 @@ struct rds_ib_mr *rds_ib_alloc_fmr(struct rds_ib_device *rds_ibdev, int npages)
out_no_cigar:
if (ibmr) {
- if (ibmr->fmr)
- ib_dealloc_fmr(ibmr->fmr);
+ if (fmr->fmr)
+ ib_dealloc_fmr(fmr->fmr);
kfree(ibmr);
}
atomic_dec(&pool->item_count);
@@ -134,6 +136,7 @@ int rds_ib_map_fmr(struct rds_ib_device *rds_ibdev, struct rds_ib_mr *ibmr,
struct scatterlist *sg, unsigned int nents)
{
struct ib_device *dev = rds_ibdev->dev;
+ struct rds_ib_fmr *fmr = &ibmr->u.fmr;
struct scatterlist *scat = sg;
u64 io_addr = 0;
u64 *dma_pages;
@@ -190,7 +193,7 @@ int rds_ib_map_fmr(struct rds_ib_device *rds_ibdev, struct rds_ib_mr *ibmr,
(dma_addr & PAGE_MASK) + j;
}
- ret = ib_map_phys_fmr(ibmr->fmr, dma_pages, page_cnt, io_addr);
+ ret = ib_map_phys_fmr(fmr->fmr, dma_pages, page_cnt, io_addr);
if (ret)
goto out;
diff --git a/net/rds/ib_mr.h b/net/rds/ib_mr.h
index d88724f..309ad59 100644
--- a/net/rds/ib_mr.h
+++ b/net/rds/ib_mr.h
@@ -43,11 +43,15 @@
#define RDS_MR_8K_SCALE (256 / (RDS_MR_8K_MSG_SIZE + 1))
#define RDS_MR_8K_POOL_SIZE (RDS_MR_8K_SCALE * (8192 / 2))
+struct rds_ib_fmr {
+ struct ib_fmr *fmr;
+ u64 *dma;
+};
+
/* This is stored as mr->r_trans_private. */
struct rds_ib_mr {
struct rds_ib_device *device;
struct rds_ib_mr_pool *pool;
- struct ib_fmr *fmr;
struct llist_node llnode;
@@ -57,8 +61,11 @@ struct rds_ib_mr {
struct scatterlist *sg;
unsigned int sg_len;
- u64 *dma;
int sg_dma_len;
+
+ union {
+ struct rds_ib_fmr fmr;
+ } u;
};
/* Our own little MR pool */
diff --git a/net/rds/ib_rdma.c b/net/rds/ib_rdma.c
index c594519..9e608d9 100644
--- a/net/rds/ib_rdma.c
+++ b/net/rds/ib_rdma.c
@@ -334,6 +334,7 @@ int rds_ib_flush_mr_pool(struct rds_ib_mr_pool *pool,
int free_all, struct rds_ib_mr **ibmr_ret)
{
struct rds_ib_mr *ibmr, *next;
+ struct rds_ib_fmr *fmr;
struct llist_node *clean_nodes;
struct llist_node *clean_tail;
LIST_HEAD(unmap_list);
@@ -395,8 +396,10 @@ int rds_ib_flush_mr_pool(struct rds_ib_mr_pool *pool,
goto out;
/* String all ib_mr's onto one list and hand them to ib_unmap_fmr */
- list_for_each_entry(ibmr, &unmap_list, unmap_list)
- list_add(&ibmr->fmr->list, &fmr_list);
+ list_for_each_entry(ibmr, &unmap_list, unmap_list) {
+ fmr = &ibmr->u.fmr;
+ list_add(&fmr->fmr->list, &fmr_list);
+ }
ret = ib_unmap_fmr(&fmr_list);
if (ret)
@@ -405,6 +408,7 @@ int rds_ib_flush_mr_pool(struct rds_ib_mr_pool *pool,
/* Now we can destroy the DMA mapping and unpin any pages */
list_for_each_entry_safe(ibmr, next, &unmap_list, unmap_list) {
unpinned += ibmr->sg_len;
+ fmr = &ibmr->u.fmr;
__rds_ib_teardown_mr(ibmr);
if (nfreed < free_goal ||
ibmr->remap_count >= pool->fmr_attr.max_maps) {
@@ -413,7 +417,7 @@ int rds_ib_flush_mr_pool(struct rds_ib_mr_pool *pool,
else
rds_ib_stats_inc(s_ib_rdma_mr_1m_free);
list_del(&ibmr->unmap_list);
- ib_dealloc_fmr(ibmr->fmr);
+ ib_dealloc_fmr(fmr->fmr);
kfree(ibmr);
nfreed++;
}
@@ -517,6 +521,7 @@ void *rds_ib_get_mr(struct scatterlist *sg, unsigned long nents,
{
struct rds_ib_device *rds_ibdev;
struct rds_ib_mr *ibmr = NULL;
+ struct rds_ib_fmr *fmr;
int ret;
rds_ibdev = rds_ib_get_device(rs->rs_bound_addr);
@@ -536,9 +541,10 @@ void *rds_ib_get_mr(struct scatterlist *sg, unsigned long nents,
return ibmr;
}
+ fmr = &ibmr->u.fmr;
ret = rds_ib_map_fmr(rds_ibdev, ibmr, sg, nents);
if (ret == 0)
- *key_ret = ibmr->fmr->rkey;
+ *key_ret = fmr->fmr->rkey;
else
printk(KERN_WARNING "RDS/IB: map_fmr failed (errno=%d)\n", ret);
--
1.9.1
--
To unsubscribe from this list: send the line "unsubscribe linux-rdma" in
the body of a message to majordomo-u79uwXL29TY76Z2rM5mHXA@public.gmane.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
^ permalink raw reply related [flat|nested] 18+ messages in thread* [net-next][PATCH 08/13] RDS: IB: add connection info to ibmr
[not found] ` <1455967804-27644-1-git-send-email-santosh.shilimkar-QHcLZuEGTsvQT0dZR+AlfA@public.gmane.org>
` (3 preceding siblings ...)
2016-02-20 11:29 ` [net-next][PATCH 06/13] RDS: IB: create struct rds_ib_fmr Santosh Shilimkar
@ 2016-02-20 11:29 ` Santosh Shilimkar
2016-02-20 11:30 ` [net-next][PATCH 10/13] RDS: IB: add mr reused stats Santosh Shilimkar
` (3 subsequent siblings)
8 siblings, 0 replies; 18+ messages in thread
From: Santosh Shilimkar @ 2016-02-20 11:29 UTC (permalink / raw)
To: netdev-u79uwXL29TY76Z2rM5mHXA, davem-fT/PcQaiUtIeIZ0/mPfg9Q
Cc: linux-rdma-u79uwXL29TY76Z2rM5mHXA,
linux-kernel-u79uwXL29TY76Z2rM5mHXA, Santosh Shilimkar
Preperatory patch for FRMR support. From connection info,
we can retrieve cm_id which contains qp handled needed for
work request posting.
We also need to drop the RDS connection on QP error states
where connection handle becomes useful.
Signed-off-by: Santosh Shilimkar <ssantosh-DgEjT+Ai2ygdnm+yROfE0A@public.gmane.org>
Signed-off-by: Santosh Shilimkar <santosh.shilimkar-QHcLZuEGTsvQT0dZR+AlfA@public.gmane.org>
---
net/rds/ib_mr.h | 17 +++++++++--------
1 file changed, 9 insertions(+), 8 deletions(-)
diff --git a/net/rds/ib_mr.h b/net/rds/ib_mr.h
index f5c1fcb..add7725 100644
--- a/net/rds/ib_mr.h
+++ b/net/rds/ib_mr.h
@@ -50,18 +50,19 @@ struct rds_ib_fmr {
/* This is stored as mr->r_trans_private. */
struct rds_ib_mr {
- struct rds_ib_device *device;
- struct rds_ib_mr_pool *pool;
+ struct rds_ib_device *device;
+ struct rds_ib_mr_pool *pool;
+ struct rds_ib_connection *ic;
- struct llist_node llnode;
+ struct llist_node llnode;
/* unmap_list is for freeing */
- struct list_head unmap_list;
- unsigned int remap_count;
+ struct list_head unmap_list;
+ unsigned int remap_count;
- struct scatterlist *sg;
- unsigned int sg_len;
- int sg_dma_len;
+ struct scatterlist *sg;
+ unsigned int sg_len;
+ int sg_dma_len;
union {
struct rds_ib_fmr fmr;
--
1.9.1
--
To unsubscribe from this list: send the line "unsubscribe linux-rdma" in
the body of a message to majordomo-u79uwXL29TY76Z2rM5mHXA@public.gmane.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
^ permalink raw reply related [flat|nested] 18+ messages in thread* [net-next][PATCH 10/13] RDS: IB: add mr reused stats
[not found] ` <1455967804-27644-1-git-send-email-santosh.shilimkar-QHcLZuEGTsvQT0dZR+AlfA@public.gmane.org>
` (4 preceding siblings ...)
2016-02-20 11:29 ` [net-next][PATCH 08/13] RDS: IB: add connection info to ibmr Santosh Shilimkar
@ 2016-02-20 11:30 ` Santosh Shilimkar
2016-02-20 11:30 ` [net-next][PATCH 11/13] RDS: IB: add Fastreg MR (FRMR) detection support Santosh Shilimkar
` (2 subsequent siblings)
8 siblings, 0 replies; 18+ messages in thread
From: Santosh Shilimkar @ 2016-02-20 11:30 UTC (permalink / raw)
To: netdev-u79uwXL29TY76Z2rM5mHXA, davem-fT/PcQaiUtIeIZ0/mPfg9Q
Cc: linux-rdma-u79uwXL29TY76Z2rM5mHXA,
linux-kernel-u79uwXL29TY76Z2rM5mHXA, Santosh Shilimkar
Add MR reuse statistics to RDS IB transport.
Signed-off-by: Santosh Shilimkar <ssantosh-DgEjT+Ai2ygdnm+yROfE0A@public.gmane.org>
Signed-off-by: Santosh Shilimkar <santosh.shilimkar-QHcLZuEGTsvQT0dZR+AlfA@public.gmane.org>
---
net/rds/ib.h | 2 ++
net/rds/ib_rdma.c | 7 ++++++-
net/rds/ib_stats.c | 2 ++
3 files changed, 10 insertions(+), 1 deletion(-)
diff --git a/net/rds/ib.h b/net/rds/ib.h
index c88cb22..62fe7d5 100644
--- a/net/rds/ib.h
+++ b/net/rds/ib.h
@@ -259,6 +259,8 @@ struct rds_ib_statistics {
uint64_t s_ib_rdma_mr_1m_pool_flush;
uint64_t s_ib_rdma_mr_1m_pool_wait;
uint64_t s_ib_rdma_mr_1m_pool_depleted;
+ uint64_t s_ib_rdma_mr_8k_reused;
+ uint64_t s_ib_rdma_mr_1m_reused;
uint64_t s_ib_atomic_cswp;
uint64_t s_ib_atomic_fadd;
};
diff --git a/net/rds/ib_rdma.c b/net/rds/ib_rdma.c
index 20ff191..00e9064 100644
--- a/net/rds/ib_rdma.c
+++ b/net/rds/ib_rdma.c
@@ -188,8 +188,13 @@ struct rds_ib_mr *rds_ib_reuse_mr(struct rds_ib_mr_pool *pool)
flag = this_cpu_ptr(&clean_list_grace);
set_bit(CLEAN_LIST_BUSY_BIT, flag);
ret = llist_del_first(&pool->clean_list);
- if (ret)
+ if (ret) {
ibmr = llist_entry(ret, struct rds_ib_mr, llnode);
+ if (pool->pool_type == RDS_IB_MR_8K_POOL)
+ rds_ib_stats_inc(s_ib_rdma_mr_8k_reused);
+ else
+ rds_ib_stats_inc(s_ib_rdma_mr_1m_reused);
+ }
clear_bit(CLEAN_LIST_BUSY_BIT, flag);
preempt_enable();
diff --git a/net/rds/ib_stats.c b/net/rds/ib_stats.c
index d77e044..7e78dca 100644
--- a/net/rds/ib_stats.c
+++ b/net/rds/ib_stats.c
@@ -73,6 +73,8 @@ static const char *const rds_ib_stat_names[] = {
"ib_rdma_mr_1m_pool_flush",
"ib_rdma_mr_1m_pool_wait",
"ib_rdma_mr_1m_pool_depleted",
+ "ib_rdma_mr_8k_reused",
+ "ib_rdma_mr_1m_reused",
"ib_atomic_cswp",
"ib_atomic_fadd",
};
--
1.9.1
--
To unsubscribe from this list: send the line "unsubscribe linux-rdma" in
the body of a message to majordomo-u79uwXL29TY76Z2rM5mHXA@public.gmane.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
^ permalink raw reply related [flat|nested] 18+ messages in thread* [net-next][PATCH 11/13] RDS: IB: add Fastreg MR (FRMR) detection support
[not found] ` <1455967804-27644-1-git-send-email-santosh.shilimkar-QHcLZuEGTsvQT0dZR+AlfA@public.gmane.org>
` (5 preceding siblings ...)
2016-02-20 11:30 ` [net-next][PATCH 10/13] RDS: IB: add mr reused stats Santosh Shilimkar
@ 2016-02-20 11:30 ` Santosh Shilimkar
[not found] ` <1455967804-27644-12-git-send-email-santosh.shilimkar-QHcLZuEGTsvQT0dZR+AlfA@public.gmane.org>
2016-02-20 11:30 ` [net-next][PATCH 12/13] RDS: IB: allocate extra space on queues for FRMR support Santosh Shilimkar
2016-02-20 11:30 ` [net-next][PATCH 13/13] RDS: IB: Support Fastreg MR (FRMR) memory registration mode Santosh Shilimkar
8 siblings, 1 reply; 18+ messages in thread
From: Santosh Shilimkar @ 2016-02-20 11:30 UTC (permalink / raw)
To: netdev-u79uwXL29TY76Z2rM5mHXA, davem-fT/PcQaiUtIeIZ0/mPfg9Q
Cc: linux-rdma-u79uwXL29TY76Z2rM5mHXA,
linux-kernel-u79uwXL29TY76Z2rM5mHXA, Santosh Shilimkar
Discovere Fast Memmory Registration support using IB device
IB_DEVICE_MEM_MGT_EXTENSIONS. Certain HCA might support just FRMR
or FMR or both FMR and FRWR. In case both mr type are supported,
default FMR is used. Using module parameter 'prefer_frmr',
user can choose its preferred MR method for RDS. Ofcourse the
module parameter has no effect if the HCA support only FRMR
or only FRMR.
Default MR is still kept as FMR against what everyone else
is following. Default will be changed to FRMR once the
RDS performance with FRMR is comparable with FMR. The
work is in progress for the same.
Signed-off-by: Santosh Shilimkar <ssantosh-DgEjT+Ai2ygdnm+yROfE0A@public.gmane.org>
Signed-off-by: Santosh Shilimkar <santosh.shilimkar-QHcLZuEGTsvQT0dZR+AlfA@public.gmane.org>
---
net/rds/ib.c | 14 ++++++++++++++
net/rds/ib.h | 4 ++++
net/rds/ib_mr.h | 1 +
3 files changed, 19 insertions(+)
diff --git a/net/rds/ib.c b/net/rds/ib.c
index bb32cb9..68c94b0 100644
--- a/net/rds/ib.c
+++ b/net/rds/ib.c
@@ -47,6 +47,7 @@
unsigned int rds_ib_mr_1m_pool_size = RDS_MR_1M_POOL_SIZE;
unsigned int rds_ib_mr_8k_pool_size = RDS_MR_8K_POOL_SIZE;
unsigned int rds_ib_retry_count = RDS_IB_DEFAULT_RETRY_COUNT;
+bool prefer_frmr;
module_param(rds_ib_mr_1m_pool_size, int, 0444);
MODULE_PARM_DESC(rds_ib_mr_1m_pool_size, " Max number of 1M mr per HCA");
@@ -54,6 +55,8 @@ module_param(rds_ib_mr_8k_pool_size, int, 0444);
MODULE_PARM_DESC(rds_ib_mr_8k_pool_size, " Max number of 8K mr per HCA");
module_param(rds_ib_retry_count, int, 0444);
MODULE_PARM_DESC(rds_ib_retry_count, " Number of hw retries before reporting an error");
+module_param(prefer_frmr, bool, 0444);
+MODULE_PARM_DESC(prefer_frmr, "Preferred MR method if both FMR and FRMR supported");
/*
* we have a clumsy combination of RCU and a rwsem protecting this list
@@ -140,6 +143,13 @@ static void rds_ib_add_one(struct ib_device *device)
rds_ibdev->max_wrs = device->attrs.max_qp_wr;
rds_ibdev->max_sge = min(device->attrs.max_sge, RDS_IB_MAX_SGE);
+ rds_ibdev->has_fr = (device->attrs.device_cap_flags &
+ IB_DEVICE_MEM_MGT_EXTENSIONS);
+ rds_ibdev->has_fmr = (device->alloc_fmr && device->dealloc_fmr &&
+ device->map_phys_fmr && device->unmap_fmr);
+ rds_ibdev->use_fastreg = (rds_ibdev->has_fr &&
+ (!rds_ibdev->has_fmr || prefer_frmr));
+
rds_ibdev->fmr_max_remaps = device->attrs.max_map_per_fmr?: 32;
rds_ibdev->max_1m_mrs = device->attrs.max_mr ?
min_t(unsigned int, (device->attrs.max_mr / 2),
@@ -178,6 +188,10 @@ static void rds_ib_add_one(struct ib_device *device)
rds_ibdev->fmr_max_remaps, rds_ibdev->max_1m_mrs,
rds_ibdev->max_8k_mrs);
+ pr_info("RDS/IB: %s: %s supported and preferred\n",
+ device->name,
+ rds_ibdev->use_fastreg ? "FRMR" : "FMR");
+
INIT_LIST_HEAD(&rds_ibdev->ipaddr_list);
INIT_LIST_HEAD(&rds_ibdev->conn_list);
diff --git a/net/rds/ib.h b/net/rds/ib.h
index 62fe7d5..c5eddc2 100644
--- a/net/rds/ib.h
+++ b/net/rds/ib.h
@@ -200,6 +200,10 @@ struct rds_ib_device {
struct list_head conn_list;
struct ib_device *dev;
struct ib_pd *pd;
+ bool has_fmr;
+ bool has_fr;
+ bool use_fastreg;
+
unsigned int max_mrs;
struct rds_ib_mr_pool *mr_1m_pool;
struct rds_ib_mr_pool *mr_8k_pool;
diff --git a/net/rds/ib_mr.h b/net/rds/ib_mr.h
index add7725..2f9b9c3 100644
--- a/net/rds/ib_mr.h
+++ b/net/rds/ib_mr.h
@@ -93,6 +93,7 @@ struct rds_ib_mr_pool {
extern struct workqueue_struct *rds_ib_mr_wq;
extern unsigned int rds_ib_mr_1m_pool_size;
extern unsigned int rds_ib_mr_8k_pool_size;
+extern bool prefer_frmr;
struct rds_ib_mr_pool *rds_ib_create_mr_pool(struct rds_ib_device *rds_dev,
int npages);
--
1.9.1
--
To unsubscribe from this list: send the line "unsubscribe linux-rdma" in
the body of a message to majordomo-u79uwXL29TY76Z2rM5mHXA@public.gmane.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
^ permalink raw reply related [flat|nested] 18+ messages in thread* [net-next][PATCH 12/13] RDS: IB: allocate extra space on queues for FRMR support
[not found] ` <1455967804-27644-1-git-send-email-santosh.shilimkar-QHcLZuEGTsvQT0dZR+AlfA@public.gmane.org>
` (6 preceding siblings ...)
2016-02-20 11:30 ` [net-next][PATCH 11/13] RDS: IB: add Fastreg MR (FRMR) detection support Santosh Shilimkar
@ 2016-02-20 11:30 ` Santosh Shilimkar
2016-02-20 11:30 ` [net-next][PATCH 13/13] RDS: IB: Support Fastreg MR (FRMR) memory registration mode Santosh Shilimkar
8 siblings, 0 replies; 18+ messages in thread
From: Santosh Shilimkar @ 2016-02-20 11:30 UTC (permalink / raw)
To: netdev-u79uwXL29TY76Z2rM5mHXA, davem-fT/PcQaiUtIeIZ0/mPfg9Q
Cc: linux-rdma-u79uwXL29TY76Z2rM5mHXA,
linux-kernel-u79uwXL29TY76Z2rM5mHXA, Santosh Shilimkar
Fastreg MR(FRMR) memory registration and invalidation makes use
of work request and completion queues for its operation. Patch
allocates extra queue space towards these operation(s).
Signed-off-by: Santosh Shilimkar <ssantosh-DgEjT+Ai2ygdnm+yROfE0A@public.gmane.org>
Signed-off-by: Santosh Shilimkar <santosh.shilimkar-QHcLZuEGTsvQT0dZR+AlfA@public.gmane.org>
---
net/rds/ib.h | 4 ++++
net/rds/ib_cm.c | 16 ++++++++++++----
2 files changed, 16 insertions(+), 4 deletions(-)
diff --git a/net/rds/ib.h b/net/rds/ib.h
index c5eddc2..eeb0d6c 100644
--- a/net/rds/ib.h
+++ b/net/rds/ib.h
@@ -14,6 +14,7 @@
#define RDS_IB_DEFAULT_RECV_WR 1024
#define RDS_IB_DEFAULT_SEND_WR 256
+#define RDS_IB_DEFAULT_FR_WR 512
#define RDS_IB_DEFAULT_RETRY_COUNT 2
@@ -122,6 +123,9 @@ struct rds_ib_connection {
struct ib_wc i_send_wc[RDS_IB_WC_MAX];
struct ib_wc i_recv_wc[RDS_IB_WC_MAX];
+ /* To control the number of wrs from fastreg */
+ atomic_t i_fastreg_wrs;
+
/* interrupt handling */
struct tasklet_struct i_send_tasklet;
struct tasklet_struct i_recv_tasklet;
diff --git a/net/rds/ib_cm.c b/net/rds/ib_cm.c
index 7f68abc..83f4673 100644
--- a/net/rds/ib_cm.c
+++ b/net/rds/ib_cm.c
@@ -363,7 +363,7 @@ static int rds_ib_setup_qp(struct rds_connection *conn)
struct ib_qp_init_attr attr;
struct ib_cq_init_attr cq_attr = {};
struct rds_ib_device *rds_ibdev;
- int ret;
+ int ret, fr_queue_space;
/*
* It's normal to see a null device if an incoming connection races
@@ -373,6 +373,12 @@ static int rds_ib_setup_qp(struct rds_connection *conn)
if (!rds_ibdev)
return -EOPNOTSUPP;
+ /* The fr_queue_space is currently set to 512, to add extra space on
+ * 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 : 0);
+
/* add the conn now so that connection establishment has the dev */
rds_ib_add_conn(rds_ibdev, conn);
@@ -384,7 +390,7 @@ static int rds_ib_setup_qp(struct rds_connection *conn)
/* Protection domain and memory range */
ic->i_pd = rds_ibdev->pd;
- cq_attr.cqe = ic->i_send_ring.w_nr + 1;
+ cq_attr.cqe = ic->i_send_ring.w_nr + fr_queue_space + 1;
ic->i_send_cq = ib_create_cq(dev, rds_ib_cq_comp_handler_send,
rds_ib_cq_event_handler, conn,
@@ -424,7 +430,7 @@ static int rds_ib_setup_qp(struct rds_connection *conn)
attr.event_handler = rds_ib_qp_event_handler;
attr.qp_context = conn;
/* + 1 to allow for the single ack message */
- attr.cap.max_send_wr = ic->i_send_ring.w_nr + 1;
+ attr.cap.max_send_wr = ic->i_send_ring.w_nr + fr_queue_space + 1;
attr.cap.max_recv_wr = ic->i_recv_ring.w_nr + 1;
attr.cap.max_send_sge = rds_ibdev->max_sge;
attr.cap.max_recv_sge = RDS_IB_RECV_SGE;
@@ -432,6 +438,7 @@ static int rds_ib_setup_qp(struct rds_connection *conn)
attr.qp_type = IB_QPT_RC;
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);
/*
* XXX this can fail if max_*_wr is too large? Are we supposed
@@ -751,7 +758,8 @@ void rds_ib_conn_shutdown(struct rds_connection *conn)
*/
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_signaled_sends) == 0) &&
+ (atomic_read(&ic->i_fastreg_wrs) == RDS_IB_DEFAULT_FR_WR));
tasklet_kill(&ic->i_send_tasklet);
tasklet_kill(&ic->i_recv_tasklet);
--
1.9.1
--
To unsubscribe from this list: send the line "unsubscribe linux-rdma" in
the body of a message to majordomo-u79uwXL29TY76Z2rM5mHXA@public.gmane.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
^ permalink raw reply related [flat|nested] 18+ messages in thread* [net-next][PATCH 13/13] RDS: IB: Support Fastreg MR (FRMR) memory registration mode
[not found] ` <1455967804-27644-1-git-send-email-santosh.shilimkar-QHcLZuEGTsvQT0dZR+AlfA@public.gmane.org>
` (7 preceding siblings ...)
2016-02-20 11:30 ` [net-next][PATCH 12/13] RDS: IB: allocate extra space on queues for FRMR support Santosh Shilimkar
@ 2016-02-20 11:30 ` Santosh Shilimkar
8 siblings, 0 replies; 18+ messages in thread
From: Santosh Shilimkar @ 2016-02-20 11:30 UTC (permalink / raw)
To: netdev-u79uwXL29TY76Z2rM5mHXA, davem-fT/PcQaiUtIeIZ0/mPfg9Q
Cc: linux-rdma-u79uwXL29TY76Z2rM5mHXA,
linux-kernel-u79uwXL29TY76Z2rM5mHXA, Santosh Shilimkar
From: Avinash Repaka <avinash.repaka-QHcLZuEGTsvQT0dZR+AlfA@public.gmane.org>
Fastreg MR(FRMR) is another method with which one can
register memory to HCA. Some of the newer HCAs supports only fastreg
mr mode, so we need to add support for it to RDS to have RDS functional
on them.
Some of the older HCAs support both FMR and FRMR modes. So to try out
FRMR on older HCAs, one can use module parameter 'prefer_frmr'
Signed-off-by: Santosh Shilimkar <ssantosh-DgEjT+Ai2ygdnm+yROfE0A@public.gmane.org>
Signed-off-by: Avinash Repaka <avinash.repaka-QHcLZuEGTsvQT0dZR+AlfA@public.gmane.org>
Signed-off-by: Santosh Shilimkar <santosh.shilimkar-QHcLZuEGTsvQT0dZR+AlfA@public.gmane.org>
---
RDS IB RDMA performance with FRMR is not yet as good as FMR and I do have
some patches in progress to address that. But they are not ready for 4.6
so I left them out of this series.
net/rds/Makefile | 2 +-
net/rds/ib.h | 1 +
net/rds/ib_cm.c | 7 +-
net/rds/ib_frmr.c | 376 ++++++++++++++++++++++++++++++++++++++++++++++++++++++
net/rds/ib_mr.h | 24 ++++
net/rds/ib_rdma.c | 17 ++-
6 files changed, 422 insertions(+), 5 deletions(-)
create mode 100644 net/rds/ib_frmr.c
diff --git a/net/rds/Makefile b/net/rds/Makefile
index bcf5591..0e72bec 100644
--- a/net/rds/Makefile
+++ b/net/rds/Makefile
@@ -6,7 +6,7 @@ rds-y := af_rds.o bind.o cong.o connection.o info.o message.o \
obj-$(CONFIG_RDS_RDMA) += rds_rdma.o
rds_rdma-y := rdma_transport.o \
ib.o ib_cm.o ib_recv.o ib_ring.o ib_send.o ib_stats.o \
- ib_sysctl.o ib_rdma.o ib_fmr.o
+ ib_sysctl.o ib_rdma.o ib_fmr.o ib_frmr.o
obj-$(CONFIG_RDS_TCP) += rds_tcp.o
diff --git a/net/rds/ib.h b/net/rds/ib.h
index eeb0d6c..627fb79 100644
--- a/net/rds/ib.h
+++ b/net/rds/ib.h
@@ -349,6 +349,7 @@ int rds_ib_update_ipaddr(struct rds_ib_device *rds_ibdev, __be32 ipaddr);
void rds_ib_add_conn(struct rds_ib_device *rds_ibdev, struct rds_connection *conn);
void rds_ib_remove_conn(struct rds_ib_device *rds_ibdev, struct rds_connection *conn);
void rds_ib_destroy_nodev_conns(void);
+void rds_ib_mr_cqe_handler(struct rds_ib_connection *ic, struct ib_wc *wc);
/* ib_recv.c */
int rds_ib_recv_init(void);
diff --git a/net/rds/ib_cm.c b/net/rds/ib_cm.c
index 83f4673..8764970 100644
--- a/net/rds/ib_cm.c
+++ b/net/rds/ib_cm.c
@@ -249,7 +249,12 @@ static void poll_scq(struct rds_ib_connection *ic, struct ib_cq *cq,
(unsigned long long)wc->wr_id, wc->status,
wc->byte_len, be32_to_cpu(wc->ex.imm_data));
- rds_ib_send_cqe_handler(ic, wc);
+ if (wc->wr_id <= ic->i_send_ring.w_nr ||
+ wc->wr_id == RDS_IB_ACK_WR_ID)
+ rds_ib_send_cqe_handler(ic, wc);
+ else
+ rds_ib_mr_cqe_handler(ic, wc);
+
}
}
}
diff --git a/net/rds/ib_frmr.c b/net/rds/ib_frmr.c
new file mode 100644
index 0000000..a86de13
--- /dev/null
+++ b/net/rds/ib_frmr.c
@@ -0,0 +1,376 @@
+/*
+ * Copyright (c) 2016 Oracle. All rights reserved.
+ *
+ * This software is available to you under a choice of one of two
+ * licenses. You may choose to be licensed under the terms of the GNU
+ * General Public License (GPL) Version 2, available from the file
+ * COPYING in the main directory of this source tree, or the
+ * OpenIB.org BSD license below:
+ *
+ * Redistribution and use in source and binary forms, with or
+ * without modification, are permitted provided that the following
+ * conditions are met:
+ *
+ * - Redistributions of source code must retain the above
+ * copyright notice, this list of conditions and the following
+ * disclaimer.
+ *
+ * - Redistributions in binary form must reproduce the above
+ * copyright notice, this list of conditions and the following
+ * disclaimer in the documentation and/or other materials
+ * provided with the distribution.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
+ * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
+ * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
+ * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
+ * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
+ * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
+ * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+
+#include "ib_mr.h"
+
+static struct rds_ib_mr *rds_ib_alloc_frmr(struct rds_ib_device *rds_ibdev,
+ int npages)
+{
+ struct rds_ib_mr_pool *pool;
+ struct rds_ib_mr *ibmr = NULL;
+ struct rds_ib_frmr *frmr;
+ int err = 0;
+
+ if (npages <= RDS_MR_8K_MSG_SIZE)
+ pool = rds_ibdev->mr_8k_pool;
+ else
+ pool = rds_ibdev->mr_1m_pool;
+
+ ibmr = rds_ib_try_reuse_ibmr(pool);
+ if (ibmr)
+ return ibmr;
+
+ ibmr = kzalloc_node(sizeof(*ibmr), GFP_KERNEL,
+ rdsibdev_to_node(rds_ibdev));
+ if (!ibmr) {
+ err = -ENOMEM;
+ goto out_no_cigar;
+ }
+
+ frmr = &ibmr->u.frmr;
+ frmr->mr = ib_alloc_mr(rds_ibdev->pd, IB_MR_TYPE_MEM_REG,
+ pool->fmr_attr.max_pages);
+ if (IS_ERR(frmr->mr)) {
+ pr_warn("RDS/IB: %s failed to allocate MR", __func__);
+ goto out_no_cigar;
+ }
+
+ ibmr->pool = pool;
+ if (pool->pool_type == RDS_IB_MR_8K_POOL)
+ rds_ib_stats_inc(s_ib_rdma_mr_8k_alloc);
+ else
+ rds_ib_stats_inc(s_ib_rdma_mr_1m_alloc);
+
+ if (atomic_read(&pool->item_count) > pool->max_items_soft)
+ pool->max_items_soft = pool->max_items;
+
+ frmr->fr_state = FRMR_IS_FREE;
+ return ibmr;
+
+out_no_cigar:
+ kfree(ibmr);
+ atomic_dec(&pool->item_count);
+ return ERR_PTR(err);
+}
+
+static void rds_ib_free_frmr(struct rds_ib_mr *ibmr, bool drop)
+{
+ struct rds_ib_mr_pool *pool = ibmr->pool;
+
+ if (drop)
+ llist_add(&ibmr->llnode, &pool->drop_list);
+ else
+ llist_add(&ibmr->llnode, &pool->free_list);
+ atomic_add(ibmr->sg_len, &pool->free_pinned);
+ atomic_inc(&pool->dirty_count);
+
+ /* If we've pinned too many pages, request a flush */
+ if (atomic_read(&pool->free_pinned) >= pool->max_free_pinned ||
+ atomic_read(&pool->dirty_count) >= pool->max_items / 5)
+ queue_delayed_work(rds_ib_mr_wq, &pool->flush_worker, 10);
+}
+
+static int rds_ib_post_reg_frmr(struct rds_ib_mr *ibmr)
+{
+ struct rds_ib_frmr *frmr = &ibmr->u.frmr;
+ struct ib_send_wr *failed_wr;
+ struct ib_reg_wr reg_wr;
+ int ret;
+
+ while (atomic_dec_return(&ibmr->ic->i_fastreg_wrs) <= 0) {
+ atomic_inc(&ibmr->ic->i_fastreg_wrs);
+ cpu_relax();
+ }
+
+ ret = ib_map_mr_sg_zbva(frmr->mr, ibmr->sg, ibmr->sg_len, PAGE_SIZE);
+ if (unlikely(ret != ibmr->sg_len))
+ return ret < 0 ? ret : -EINVAL;
+
+ /* Perform a WR for the fast_reg_mr. Each individual page
+ * in the sg list is added to the fast reg page list and placed
+ * inside the fast_reg_mr WR. The key used is a rolling 8bit
+ * counter, which should guarantee uniqueness.
+ */
+ ib_update_fast_reg_key(frmr->mr, ibmr->remap_count++);
+ frmr->fr_state = FRMR_IS_INUSE;
+
+ memset(®_wr, 0, sizeof(reg_wr));
+ reg_wr.wr.wr_id = (unsigned long)(void *)ibmr;
+ reg_wr.wr.opcode = IB_WR_REG_MR;
+ reg_wr.wr.num_sge = 0;
+ reg_wr.mr = frmr->mr;
+ reg_wr.key = frmr->mr->rkey;
+ reg_wr.access = IB_ACCESS_LOCAL_WRITE |
+ IB_ACCESS_REMOTE_READ |
+ IB_ACCESS_REMOTE_WRITE;
+ reg_wr.wr.send_flags = IB_SEND_SIGNALED;
+
+ failed_wr = ®_wr.wr;
+ ret = ib_post_send(ibmr->ic->i_cm_id->qp, ®_wr.wr, &failed_wr);
+ WARN_ON(failed_wr != ®_wr.wr);
+ if (unlikely(ret)) {
+ /* Failure here can be because of -ENOMEM as well */
+ frmr->fr_state = FRMR_IS_STALE;
+ atomic_inc(&ibmr->ic->i_fastreg_wrs);
+ if (printk_ratelimit())
+ pr_warn("RDS/IB: %s returned error(%d)\n",
+ __func__, ret);
+ }
+ return ret;
+}
+
+static int rds_ib_map_frmr(struct rds_ib_device *rds_ibdev,
+ struct rds_ib_mr_pool *pool,
+ struct rds_ib_mr *ibmr,
+ struct scatterlist *sg, unsigned int sg_len)
+{
+ struct ib_device *dev = rds_ibdev->dev;
+ struct rds_ib_frmr *frmr = &ibmr->u.frmr;
+ int i;
+ u32 len;
+ int ret = 0;
+
+ /* We want to teardown old ibmr values here and fill it up with
+ * new sg values
+ */
+ rds_ib_teardown_mr(ibmr);
+
+ ibmr->sg = sg;
+ ibmr->sg_len = sg_len;
+ ibmr->sg_dma_len = 0;
+ frmr->sg_byte_len = 0;
+ WARN_ON(ibmr->sg_dma_len);
+ ibmr->sg_dma_len = ib_dma_map_sg(dev, ibmr->sg, ibmr->sg_len,
+ DMA_BIDIRECTIONAL);
+ if (unlikely(!ibmr->sg_dma_len)) {
+ pr_warn("RDS/IB: %s failed!\n", __func__);
+ return -EBUSY;
+ }
+
+ frmr->sg_byte_len = 0;
+ frmr->dma_npages = 0;
+ len = 0;
+
+ ret = -EINVAL;
+ for (i = 0; i < ibmr->sg_dma_len; ++i) {
+ unsigned int dma_len = ib_sg_dma_len(dev, &ibmr->sg[i]);
+ u64 dma_addr = ib_sg_dma_address(dev, &ibmr->sg[i]);
+
+ frmr->sg_byte_len += dma_len;
+ if (dma_addr & ~PAGE_MASK) {
+ if (i > 0)
+ goto out_unmap;
+ else
+ ++frmr->dma_npages;
+ }
+
+ if ((dma_addr + dma_len) & ~PAGE_MASK) {
+ if (i < ibmr->sg_dma_len - 1)
+ goto out_unmap;
+ else
+ ++frmr->dma_npages;
+ }
+
+ len += dma_len;
+ }
+ frmr->dma_npages += len >> PAGE_SHIFT;
+
+ if (frmr->dma_npages > ibmr->pool->fmr_attr.max_pages) {
+ ret = -EMSGSIZE;
+ goto out_unmap;
+ }
+
+ ret = rds_ib_post_reg_frmr(ibmr);
+ if (ret)
+ goto out_unmap;
+
+ if (ibmr->pool->pool_type == RDS_IB_MR_8K_POOL)
+ rds_ib_stats_inc(s_ib_rdma_mr_8k_used);
+ else
+ rds_ib_stats_inc(s_ib_rdma_mr_1m_used);
+
+ return ret;
+
+out_unmap:
+ ib_dma_unmap_sg(rds_ibdev->dev, ibmr->sg, ibmr->sg_len,
+ DMA_BIDIRECTIONAL);
+ ibmr->sg_dma_len = 0;
+ return ret;
+}
+
+static int rds_ib_post_inv(struct rds_ib_mr *ibmr)
+{
+ struct ib_send_wr *s_wr, *failed_wr;
+ struct rds_ib_frmr *frmr = &ibmr->u.frmr;
+ struct rdma_cm_id *i_cm_id = ibmr->ic->i_cm_id;
+ int ret = -EINVAL;
+
+ if (!i_cm_id || !i_cm_id->qp || !frmr->mr)
+ goto out;
+
+ if (frmr->fr_state != FRMR_IS_INUSE)
+ goto out;
+
+ while (atomic_dec_return(&ibmr->ic->i_fastreg_wrs) <= 0) {
+ atomic_inc(&ibmr->ic->i_fastreg_wrs);
+ cpu_relax();
+ }
+
+ frmr->fr_inv = true;
+ s_wr = &frmr->fr_wr;
+
+ memset(s_wr, 0, sizeof(*s_wr));
+ s_wr->wr_id = (unsigned long)(void *)ibmr;
+ s_wr->opcode = IB_WR_LOCAL_INV;
+ s_wr->ex.invalidate_rkey = frmr->mr->rkey;
+ s_wr->send_flags = IB_SEND_SIGNALED;
+
+ failed_wr = s_wr;
+ ret = ib_post_send(i_cm_id->qp, s_wr, &failed_wr);
+ WARN_ON(failed_wr != s_wr);
+ if (unlikely(ret)) {
+ frmr->fr_state = FRMR_IS_STALE;
+ frmr->fr_inv = false;
+ atomic_inc(&ibmr->ic->i_fastreg_wrs);
+ pr_err("RDS/IB: %s returned error(%d)\n", __func__, ret);
+ goto out;
+ }
+out:
+ return ret;
+}
+
+void rds_ib_mr_cqe_handler(struct rds_ib_connection *ic, struct ib_wc *wc)
+{
+ struct rds_ib_mr *ibmr = (void *)wc->wr_id;
+ struct rds_ib_frmr *frmr = &ibmr->u.frmr;
+
+ if (wc->status != IB_WC_SUCCESS) {
+ frmr->fr_state = FRMR_IS_STALE;
+ if (rds_conn_up(ic->conn))
+ rds_ib_conn_error(ic->conn,
+ "frmr completion <%pI4,%pI4> status %u(%s), vendor_err 0x%x, disconnecting and reconnecting\n",
+ &ic->conn->c_laddr,
+ &ic->conn->c_faddr,
+ wc->status,
+ ib_wc_status_msg(wc->status),
+ wc->vendor_err);
+ }
+
+ if (frmr->fr_inv) {
+ frmr->fr_state = FRMR_IS_FREE;
+ frmr->fr_inv = false;
+ }
+
+ atomic_inc(&ic->i_fastreg_wrs);
+}
+
+void rds_ib_unreg_frmr(struct list_head *list, unsigned int *nfreed,
+ unsigned long *unpinned, unsigned int goal)
+{
+ struct rds_ib_mr *ibmr, *next;
+ struct rds_ib_frmr *frmr;
+ int ret = 0;
+ unsigned int freed = *nfreed;
+
+ /* String all ib_mr's onto one list and hand them to ib_unmap_fmr */
+ list_for_each_entry(ibmr, list, unmap_list) {
+ if (ibmr->sg_dma_len)
+ ret |= rds_ib_post_inv(ibmr);
+ }
+ if (ret)
+ pr_warn("RDS/IB: %s failed (err=%d)\n", __func__, ret);
+
+ /* Now we can destroy the DMA mapping and unpin any pages */
+ list_for_each_entry_safe(ibmr, next, list, unmap_list) {
+ *unpinned += ibmr->sg_len;
+ frmr = &ibmr->u.frmr;
+ __rds_ib_teardown_mr(ibmr);
+ if (freed < goal || frmr->fr_state == FRMR_IS_STALE) {
+ /* Don't de-allocate if the MR is not free yet */
+ if (frmr->fr_state == FRMR_IS_INUSE)
+ continue;
+
+ if (ibmr->pool->pool_type == RDS_IB_MR_8K_POOL)
+ rds_ib_stats_inc(s_ib_rdma_mr_8k_free);
+ else
+ rds_ib_stats_inc(s_ib_rdma_mr_1m_free);
+ list_del(&ibmr->unmap_list);
+ if (frmr->mr)
+ ib_dereg_mr(frmr->mr);
+ kfree(ibmr);
+ freed++;
+ }
+ }
+ *nfreed = freed;
+}
+
+struct rds_ib_mr *rds_ib_reg_frmr(struct rds_ib_device *rds_ibdev,
+ struct rds_ib_connection *ic,
+ struct scatterlist *sg,
+ unsigned long nents, u32 *key)
+{
+ struct rds_ib_mr *ibmr = NULL;
+ struct rds_ib_frmr *frmr;
+ int ret;
+
+ do {
+ if (ibmr)
+ rds_ib_free_frmr(ibmr, true);
+ ibmr = rds_ib_alloc_frmr(rds_ibdev, nents);
+ if (IS_ERR(ibmr))
+ return ibmr;
+ frmr = &ibmr->u.frmr;
+ } while (frmr->fr_state != FRMR_IS_FREE);
+
+ ibmr->ic = ic;
+ ibmr->device = rds_ibdev;
+ ret = rds_ib_map_frmr(rds_ibdev, ibmr->pool, ibmr, sg, nents);
+ if (ret == 0) {
+ *key = frmr->mr->rkey;
+ } else {
+ rds_ib_free_frmr(ibmr, false);
+ ibmr = ERR_PTR(ret);
+ }
+
+ return ibmr;
+}
+
+void rds_ib_free_frmr_list(struct rds_ib_mr *ibmr)
+{
+ struct rds_ib_mr_pool *pool = ibmr->pool;
+ struct rds_ib_frmr *frmr = &ibmr->u.frmr;
+
+ if (frmr->fr_state == FRMR_IS_STALE)
+ llist_add(&ibmr->llnode, &pool->drop_list);
+ else
+ llist_add(&ibmr->llnode, &pool->free_list);
+}
diff --git a/net/rds/ib_mr.h b/net/rds/ib_mr.h
index 2f9b9c3..1c754f4 100644
--- a/net/rds/ib_mr.h
+++ b/net/rds/ib_mr.h
@@ -48,6 +48,21 @@ struct rds_ib_fmr {
u64 *dma;
};
+enum rds_ib_fr_state {
+ FRMR_IS_FREE, /* mr invalidated & ready for use */
+ FRMR_IS_INUSE, /* mr is in use or used & can be invalidated */
+ FRMR_IS_STALE, /* Stale MR and needs to be dropped */
+};
+
+struct rds_ib_frmr {
+ struct ib_mr *mr;
+ enum rds_ib_fr_state fr_state;
+ bool fr_inv;
+ struct ib_send_wr fr_wr;
+ unsigned int dma_npages;
+ unsigned int sg_byte_len;
+};
+
/* This is stored as mr->r_trans_private. */
struct rds_ib_mr {
struct rds_ib_device *device;
@@ -66,6 +81,7 @@ struct rds_ib_mr {
union {
struct rds_ib_fmr fmr;
+ struct rds_ib_frmr frmr;
} u;
};
@@ -88,6 +104,7 @@ struct rds_ib_mr_pool {
unsigned long max_items_soft;
unsigned long max_free_pinned;
struct ib_fmr_attr fmr_attr;
+ bool use_fastreg;
};
extern struct workqueue_struct *rds_ib_mr_wq;
@@ -121,4 +138,11 @@ struct rds_ib_mr *rds_ib_try_reuse_ibmr(struct rds_ib_mr_pool *);
void rds_ib_unreg_fmr(struct list_head *, unsigned int *,
unsigned long *, unsigned int);
void rds_ib_free_fmr_list(struct rds_ib_mr *);
+struct rds_ib_mr *rds_ib_reg_frmr(struct rds_ib_device *rds_ibdev,
+ struct rds_ib_connection *ic,
+ struct scatterlist *sg,
+ unsigned long nents, u32 *key);
+void rds_ib_unreg_frmr(struct list_head *list, unsigned int *nfreed,
+ unsigned long *unpinned, unsigned int goal);
+void rds_ib_free_frmr_list(struct rds_ib_mr *);
#endif
diff --git a/net/rds/ib_rdma.c b/net/rds/ib_rdma.c
index 00e9064..6e9ccbd 100644
--- a/net/rds/ib_rdma.c
+++ b/net/rds/ib_rdma.c
@@ -398,7 +398,10 @@ int rds_ib_flush_mr_pool(struct rds_ib_mr_pool *pool,
if (list_empty(&unmap_list))
goto out;
- rds_ib_unreg_fmr(&unmap_list, &nfreed, &unpinned, free_goal);
+ if (pool->use_fastreg)
+ rds_ib_unreg_frmr(&unmap_list, &nfreed, &unpinned, free_goal);
+ else
+ rds_ib_unreg_fmr(&unmap_list, &nfreed, &unpinned, free_goal);
if (!list_empty(&unmap_list)) {
/* we have to make sure that none of the things we're about
@@ -490,7 +493,10 @@ void rds_ib_free_mr(void *trans_private, int invalidate)
rdsdebug("RDS/IB: free_mr nents %u\n", ibmr->sg_len);
/* Return it to the pool's free list */
- rds_ib_free_fmr_list(ibmr);
+ if (rds_ibdev->use_fastreg)
+ rds_ib_free_frmr_list(ibmr);
+ else
+ rds_ib_free_fmr_list(ibmr);
atomic_add(ibmr->sg_len, &pool->free_pinned);
atomic_inc(&pool->dirty_count);
@@ -535,6 +541,7 @@ void *rds_ib_get_mr(struct scatterlist *sg, unsigned long nents,
{
struct rds_ib_device *rds_ibdev;
struct rds_ib_mr *ibmr = NULL;
+ struct rds_ib_connection *ic = rs->rs_conn->c_transport_data;
int ret;
rds_ibdev = rds_ib_get_device(rs->rs_bound_addr);
@@ -548,7 +555,10 @@ void *rds_ib_get_mr(struct scatterlist *sg, unsigned long nents,
goto out;
}
- ibmr = rds_ib_reg_fmr(rds_ibdev, sg, nents, key_ret);
+ if (rds_ibdev->use_fastreg)
+ ibmr = rds_ib_reg_frmr(rds_ibdev, ic, sg, nents, key_ret);
+ else
+ ibmr = rds_ib_reg_fmr(rds_ibdev, sg, nents, key_ret);
if (ibmr)
rds_ibdev = NULL;
@@ -602,6 +612,7 @@ struct rds_ib_mr_pool *rds_ib_create_mr_pool(struct rds_ib_device *rds_ibdev,
pool->fmr_attr.max_maps = rds_ibdev->fmr_max_remaps;
pool->fmr_attr.page_shift = PAGE_SHIFT;
pool->max_items_soft = rds_ibdev->max_mrs * 3 / 4;
+ pool->use_fastreg = rds_ibdev->use_fastreg;
return pool;
}
--
1.9.1
--
To unsubscribe from this list: send the line "unsubscribe linux-rdma" in
the body of a message to majordomo-u79uwXL29TY76Z2rM5mHXA@public.gmane.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
^ permalink raw reply related [flat|nested] 18+ messages in thread