@@ -521,6 +521,9 @@ static const struct {
[IB_QPT_RC] = (IB_QP_PKEY_INDEX |
IB_QP_PORT |
IB_QP_ACCESS_FLAGS),
+ [IB_QPT_DC_INI] = (IB_QP_PKEY_INDEX |
+ IB_QP_PORT |
+ IB_QP_DC_KEY),
[IB_QPT_XRC_INI] = (IB_QP_PKEY_INDEX |
IB_QP_PORT |
IB_QP_ACCESS_FLAGS),
@@ -549,6 +552,9 @@ static const struct {
[IB_QPT_RC] = (IB_QP_PKEY_INDEX |
IB_QP_PORT |
IB_QP_ACCESS_FLAGS),
+ [IB_QPT_DC_INI] = (IB_QP_PKEY_INDEX |
+ IB_QP_PORT |
+ IB_QP_DC_KEY),
[IB_QPT_XRC_INI] = (IB_QP_PKEY_INDEX |
IB_QP_PORT |
IB_QP_ACCESS_FLAGS),
@@ -574,6 +580,8 @@ static const struct {
IB_QP_RQ_PSN |
IB_QP_MAX_DEST_RD_ATOMIC |
IB_QP_MIN_RNR_TIMER),
+ [IB_QPT_DC_INI] = (IB_QP_AV |
+ IB_QP_PATH_MTU),
[IB_QPT_XRC_INI] = (IB_QP_AV |
IB_QP_PATH_MTU |
IB_QP_DEST_QPN |
@@ -600,6 +608,8 @@ static const struct {
[IB_QPT_RC] = (IB_QP_ALT_PATH |
IB_QP_ACCESS_FLAGS |
IB_QP_PKEY_INDEX),
+ [IB_QPT_DC_INI] = (IB_QP_PKEY_INDEX |
+ IB_QP_DC_KEY),
[IB_QPT_XRC_INI] = (IB_QP_ALT_PATH |
IB_QP_ACCESS_FLAGS |
IB_QP_PKEY_INDEX),
@@ -640,6 +650,10 @@ static const struct {
IB_QP_RNR_RETRY |
IB_QP_SQ_PSN |
IB_QP_MAX_QP_RD_ATOMIC),
+ [IB_QPT_DC_INI] = (IB_QP_TIMEOUT |
+ IB_QP_RETRY_CNT |
+ IB_QP_RNR_RETRY |
+ IB_QP_MAX_QP_RD_ATOMIC),
[IB_QPT_XRC_INI] = (IB_QP_TIMEOUT |
IB_QP_RETRY_CNT |
IB_QP_RNR_RETRY |
@@ -662,6 +676,10 @@ static const struct {
IB_QP_ACCESS_FLAGS |
IB_QP_MIN_RNR_TIMER |
IB_QP_PATH_MIG_STATE),
+ [IB_QPT_DC_INI] = (IB_QP_CUR_STATE |
+ IB_QP_ALT_PATH |
+ IB_QP_MIN_RNR_TIMER |
+ IB_QP_PATH_MIG_STATE),
[IB_QPT_XRC_INI] = (IB_QP_CUR_STATE |
IB_QP_ALT_PATH |
IB_QP_ACCESS_FLAGS |
@@ -695,6 +713,10 @@ static const struct {
IB_QP_ALT_PATH |
IB_QP_PATH_MIG_STATE |
IB_QP_MIN_RNR_TIMER),
+ [IB_QPT_DC_INI] = (IB_QP_CUR_STATE |
+ IB_QP_ALT_PATH |
+ IB_QP_PATH_MIG_STATE |
+ IB_QP_MIN_RNR_TIMER),
[IB_QPT_XRC_INI] = (IB_QP_CUR_STATE |
IB_QP_ACCESS_FLAGS |
IB_QP_ALT_PATH |
@@ -1438,6 +1460,71 @@ int ib_destroy_flow(struct ib_flow *flow_id)
}
EXPORT_SYMBOL(ib_destroy_flow);
+struct ib_dct *ib_create_dct(struct ib_pd *pd, struct ib_dct_init_attr *attr,
+ struct ib_udata *uhw)
+{
+ struct ib_dct *dct;
+
+ if (!pd->device->create_dct)
+ return ERR_PTR(-ENOSYS);
+
+ dct = pd->device->create_dct(pd, attr, uhw);
+ if (!IS_ERR(dct)) {
+ dct->pd = pd;
+ dct->srq = attr->srq;
+ dct->cq = attr->cq;
+ atomic_inc(&dct->srq->usecnt);
+ atomic_inc(&dct->cq->usecnt);
+ atomic_inc(&dct->pd->usecnt);
+ }
+
+ return dct;
+}
+EXPORT_SYMBOL(ib_create_dct);
+
+int ib_destroy_dct(struct ib_dct *dct, struct ib_udata *uhw)
+{
+ struct ib_srq *srq;
+ struct ib_cq *cq;
+ struct ib_pd *pd;
+ int err;
+
+ if (!dct->device->destroy_dct)
+ return -ENOSYS;
+
+ srq = dct->srq;
+ cq = dct->cq;
+ pd = dct->pd;
+ err = dct->device->destroy_dct(dct, uhw);
+ if (!err) {
+ atomic_dec(&srq->usecnt);
+ atomic_dec(&cq->usecnt);
+ atomic_dec(&pd->usecnt);
+ }
+
+ return err;
+}
+EXPORT_SYMBOL(ib_destroy_dct);
+
+int ib_query_dct(struct ib_dct *dct, struct ib_dct_attr *attr,
+ struct ib_udata *uhw)
+{
+ if (!dct->device->query_dct)
+ return -ENOSYS;
+
+ return dct->device->query_dct(dct, attr, uhw);
+}
+EXPORT_SYMBOL(ib_query_dct);
+
+int ib_arm_dct(struct ib_dct *dct, struct ib_udata *uhw)
+{
+ if (!dct->device->arm_dct)
+ return -ENOSYS;
+
+ return dct->device->arm_dct(dct, uhw);
+}
+EXPORT_SYMBOL(ib_arm_dct);
+
int ib_check_mr_status(struct ib_mr *mr, u32 check_mask,
struct ib_mr_status *mr_status)
{
@@ -123,7 +123,14 @@ enum ib_device_cap_flags {
IB_DEVICE_MEM_WINDOW_TYPE_2A = (1<<23),
IB_DEVICE_MEM_WINDOW_TYPE_2B = (1<<24),
IB_DEVICE_MANAGED_FLOW_STEERING = (1<<29),
- IB_DEVICE_SIGNATURE_HANDOVER = (1<<30)
+ IB_DEVICE_SIGNATURE_HANDOVER = (1<<30),
+ IB_DEVICE_DC_TRANSPORT = ((u64)1<<32),
+};
+
+enum {
+ IB_DCT_STATE_ACTIVE,
+ IB_DCT_STATE_DRAINING,
+ IB_DCT_STATE_DRAINED
};
enum ib_signature_prot_cap {
@@ -155,7 +162,7 @@ struct ib_device_attr {
u32 hw_ver;
int max_qp;
int max_qp_wr;
- int device_cap_flags;
+ u64 device_cap_flags;
int max_sge;
int max_sge_rd;
int max_cq;
@@ -191,6 +198,8 @@ struct ib_device_attr {
u32 log_atomic_arg_sizes; /* bit-mask of supported sizes */
u32 max_fa_bit_boundary;
u32 log_max_atomic_inline;
+ uint32_t dc_rd_req;
+ uint32_t dc_rd_res;
};
enum ib_mtu {
@@ -391,6 +400,9 @@ enum ib_event_type {
IB_EVENT_QP_LAST_WQE_REACHED,
IB_EVENT_CLIENT_REREGISTER,
IB_EVENT_GID_CHANGE,
+ IB_EVENT_DCT_KEY_VIOLATION,
+ IB_EVENT_DCT_ACCESS_ERR,
+ IB_EVENT_DCT_REQ_ERR,
};
struct ib_event {
@@ -399,6 +411,7 @@ struct ib_event {
struct ib_cq *cq;
struct ib_qp *qp;
struct ib_srq *srq;
+ struct ib_dct *dct;
u8 port_num;
} element;
enum ib_event_type event;
@@ -762,6 +775,7 @@ enum ib_qp_type {
IB_QPT_RAW_PACKET = 8,
IB_QPT_XRC_INI = 9,
IB_QPT_XRC_TGT,
+ IB_QPT_DC_INI,
IB_QPT_MAX,
/* Reserve a range for qp types internal to the low level driver.
* These qp types will not be visible at the IB core layer, so the
@@ -878,6 +892,7 @@ enum ib_qp_attr_mask {
IB_QP_ALT_SMAC = (1<<22),
IB_QP_VID = (1<<23),
IB_QP_ALT_VID = (1<<24),
+ IB_QP_DC_KEY = (1<<25),
};
enum ib_qp_state {
@@ -931,6 +946,7 @@ struct ib_qp_attr {
u8 alt_smac[ETH_ALEN];
u16 vlan_id;
u16 alt_vlan_id;
+ u64 dct_key;
};
enum ib_wr_opcode {
@@ -1393,6 +1409,58 @@ struct ib_cache {
u8 *lmc_cache;
};
+struct ib_dct {
+ struct ib_device *device;
+ struct ib_uobject *uobject;
+ struct ib_pd *pd;
+ struct ib_cq *cq;
+ struct ib_srq *srq;
+ void *dct_context;
+ u32 dct_num;
+
+ void (*event_handler)(struct ib_event *, void *);
+};
+
+enum {
+ IB_DCT_CREATE_FLAGS_MASK = 0,
+};
+
+struct ib_dct_init_attr {
+ struct ib_pd *pd;
+ struct ib_cq *cq;
+ struct ib_srq *srq;
+ u64 dc_key;
+ u8 port;
+ u32 access_flags;
+ u8 min_rnr_timer;
+ u8 tclass;
+ u32 flow_label;
+ enum ib_mtu mtu;
+ u8 pkey_index;
+ u8 gid_index;
+ u8 hop_limit;
+ u32 create_flags;
+ u32 inline_size;
+ void *dct_context;
+
+ void (*event_handler)(struct ib_event *, void *);
+};
+
+struct ib_dct_attr {
+ u64 dc_key;
+ u8 port;
+ u32 access_flags;
+ u8 min_rnr_timer;
+ u8 tclass;
+ u32 flow_label;
+ enum ib_mtu mtu;
+ u8 pkey_index;
+ u8 gid_index;
+ u8 hop_limit;
+ u32 key_violations;
+ u8 state;
+};
+
struct ib_dma_mapping_ops {
int (*mapping_error)(struct ib_device *dev,
u64 dma_addr);
@@ -1612,6 +1680,15 @@ struct ib_device {
int (*destroy_flow)(struct ib_flow *flow_id);
int (*check_mr_status)(struct ib_mr *mr, u32 check_mask,
struct ib_mr_status *mr_status);
+ struct ib_dct * (*create_dct)(struct ib_pd *pd,
+ struct ib_dct_init_attr *attr,
+ struct ib_udata *uhw);
+ int (*destroy_dct)(struct ib_dct *dct,
+ struct ib_udata *uhw);
+ int (*query_dct)(struct ib_dct *dct,
+ struct ib_dct_attr *attr,
+ struct ib_udata *uhw);
+ int (*arm_dct)(struct ib_dct *dct, struct ib_udata *uhw);
struct ib_dma_mapping_ops *dma_ops;
@@ -2602,6 +2679,12 @@ int ib_dealloc_xrcd(struct ib_xrcd *xrcd);
struct ib_flow *ib_create_flow(struct ib_qp *qp,
struct ib_flow_attr *flow_attr, int domain);
int ib_destroy_flow(struct ib_flow *flow_id);
+struct ib_dct *ib_create_dct(struct ib_pd *pd, struct ib_dct_init_attr *attr,
+ struct ib_udata *uhw);
+int ib_destroy_dct(struct ib_dct *dct, struct ib_udata *uhw);
+int ib_query_dct(struct ib_dct *dct, struct ib_dct_attr *attr,
+ struct ib_udata *uhw);
+int ib_arm_dct(struct ib_dct *dct, struct ib_udata *uhw);
static inline int ib_check_mr_access(int flags)
{
The Dynamically Connected (DC) Transport Service provides a reliable datagram-like model that allows a single sender to target multiple destinations from the same QP, keeping the communication resource footprint essentially independent of system size. DC supports RDMA read and write operations, as well as atomic variable updates. With this transport a DC initiator QP may be used to target multiple remote DC Targets, in one or more remote processes. As far as reachability is concerned, the DC model is somewhat similar to the Unreliable Datagram (UD) model in the sense that each WR submitted to the DC SQ carries the information that identifies the remote destination. DC contexts are then dynamically tied to each other across the network to create a temporary RC-equivalent connection that is used to reliably deliver one or more messages. This dynamic connection is created in-band and pipelined with the subsequent data communication thus eliminating most of the cost associated with the 3-way handshake off the Connection Manager protocol used for connecting RC QPs. When all WRs posted to that remote network address are acknowledged, the initiator sends a disconnect request to the responder, thereby releasing the responder resources. A DC initiator is yet another type of QP identified by a new transport type, IB_QPT_DC_INI. The target is end is presented by a new object of type ib_dct. This patch extend the verbs API with the following new APIs: ib_create_dc - Create a DC target ib_destroy_dct - Destroy a DC target ib_query_dct - query DC target ib_arm_dct - Arm a DC target to generate asynchronous event on DC key violation. Once a event is generated, the DC target moves to a fired state and will not generated further key violation events unless re-armed. ib_modify_qp_ex - is an extension to ib_modify_qp which allows to pass the 64 bit DC key. Signed-off-by: Eli Cohen <eli@mellanox.com> --- drivers/infiniband/core/verbs.c | 87 +++++++++++++++++++++++++++++++++++++++++ include/rdma/ib_verbs.h | 87 ++++++++++++++++++++++++++++++++++++++++- 2 files changed, 172 insertions(+), 2 deletions(-)