From: zzry <1245464216(a)qq.com>
Add support for thread domain (TD) and parent domain (PAD).
Extend the orginal hns_roce_pd struct to hns_roce_pad by
adding the new hns_roce_td struct. When a parent domain
holds a thread domain, the associated data path will be set
to lock-free mode to improve performance.
---
providers/hns/hns_roce_u.c | 5 +-
providers/hns/hns_roce_u.h | 69 +++++++++++++-
providers/hns/hns_roce_u_verbs.c | 156 ++++++++++++++++++++++++++++---
3 files changed, 215 insertions(+), 15 deletions(-)
diff --git a/providers/hns/hns_roce_u.c b/providers/hns/hns_roce_u.c
index 266e73e..e3c72bb 100644
--- a/providers/hns/hns_roce_u.c
+++ b/providers/hns/hns_roce_u.c
@@ -67,7 +67,7 @@ static const struct verbs_context_ops hns_common_ops = {
.create_qp = hns_roce_u_create_qp,
.create_qp_ex = hns_roce_u_create_qp_ex,
.dealloc_mw = hns_roce_u_dealloc_mw,
- .dealloc_pd = hns_roce_u_free_pd,
+ .dealloc_pd = hns_roce_u_dealloc_pd,
.dereg_mr = hns_roce_u_dereg_mr,
.destroy_cq = hns_roce_u_destroy_cq,
.modify_cq = hns_roce_u_modify_cq,
@@ -88,6 +88,9 @@ static const struct verbs_context_ops hns_common_ops = {
.close_xrcd = hns_roce_u_close_xrcd,
.open_qp = hns_roce_u_open_qp,
.get_srq_num = hns_roce_u_get_srq_num,
+ .alloc_td = hns_roce_u_alloc_td,
+ .dealloc_td = hns_roce_u_dealloc_td,
+ .alloc_parent_domain = hns_roce_u_alloc_pad,
};
static uint32_t calc_table_shift(uint32_t entry_count, uint32_t size_shift)
diff --git a/providers/hns/hns_roce_u.h b/providers/hns/hns_roce_u.h
index afb68fe..338f162 100644
--- a/providers/hns/hns_roce_u.h
+++ b/providers/hns/hns_roce_u.h
@@ -190,6 +190,11 @@ struct hns_roce_db_page {
unsigned long *bitmap;
};
+struct hns_roce_spinlock {
+ pthread_spinlock_t lock;
+ int need_lock;
+};
+
struct hns_roce_context {
struct verbs_context ibv_ctx;
void *uar;
@@ -224,9 +229,21 @@ struct hns_roce_context {
unsigned int max_inline_data;
};
+struct hns_roce_td {
+ struct ibv_td ibv_td;
+ atomic_int refcount;
+};
+
struct hns_roce_pd {
struct ibv_pd ibv_pd;
unsigned int pdn;
+ atomic_int refcount;
+ struct hns_roce_pd *protection_domain;
+};
+
+struct hns_roce_pad {
+ struct hns_roce_pd pd;
+ struct hns_roce_td *td;
};
struct hns_roce_cq {
@@ -398,9 +415,35 @@ static inline struct hns_roce_context *to_hr_ctx(struct ibv_context *ibv_ctx)
return container_of(ibv_ctx, struct hns_roce_context, ibv_ctx.context);
}
+static inline struct hns_roce_td *to_hr_td(struct ibv_td *ibv_td)
+{
+ return container_of(ibv_td, struct hns_roce_td, ibv_td);
+}
+
+/* to_hr_pd always returns the real hns_roce_pd obj. */
static inline struct hns_roce_pd *to_hr_pd(struct ibv_pd *ibv_pd)
{
- return container_of(ibv_pd, struct hns_roce_pd, ibv_pd);
+ struct hns_roce_pd *pd =
+ container_of(ibv_pd, struct hns_roce_pd, ibv_pd);
+
+ if (pd->protection_domain)
+ return pd->protection_domain;
+
+ return pd;
+}
+
+static inline struct hns_roce_pad *to_hr_pad(struct ibv_pd *ibv_pd)
+{
+ struct hns_roce_pad *pad =
+ ibv_pd ?
+ container_of(ibv_pd, struct hns_roce_pad, pd.ibv_pd) :
+ NULL;
+
+ if (pad && pad->pd.protection_domain)
+ return pad;
+
+ /* Otherwise ibv_pd isn't a parent_domain */
+ return NULL;
}
static inline struct hns_roce_cq *to_hr_cq(struct ibv_cq *ibv_cq)
@@ -423,14 +466,35 @@ static inline struct hns_roce_ah *to_hr_ah(struct ibv_ah *ibv_ah)
return container_of(ibv_ah, struct hns_roce_ah, ibv_ah);
}
+static inline int hns_roce_spin_lock(struct hns_roce_spinlock *hr_lock)
+{
+ if (hr_lock->need_lock)
+ return pthread_spin_lock(&hr_lock->lock);
+
+ return 0;
+}
+
+static inline int hns_roce_spin_unlock(struct hns_roce_spinlock *hr_lock)
+{
+ if (hr_lock->need_lock)
+ return pthread_spin_unlock(&hr_lock->lock);
+
+ return 0;
+}
+
int hns_roce_u_query_device(struct ibv_context *context,
const struct ibv_query_device_ex_input *input,
struct ibv_device_attr_ex *attr, size_t attr_size);
int hns_roce_u_query_port(struct ibv_context *context, uint8_t port,
struct ibv_port_attr *attr);
+struct ibv_td *hns_roce_u_alloc_td(struct ibv_context *context,
+ struct ibv_td_init_attr *attr);
+int hns_roce_u_dealloc_td(struct ibv_td *ibv_td);
+struct ibv_pd *hns_roce_u_alloc_pad(struct ibv_context *context,
+ struct ibv_parent_domain_init_attr *attr);
struct ibv_pd *hns_roce_u_alloc_pd(struct ibv_context *context);
-int hns_roce_u_free_pd(struct ibv_pd *pd);
+int hns_roce_u_dealloc_pd(struct ibv_pd *pd);
struct ibv_mr *hns_roce_u_reg_mr(struct ibv_pd *pd, void *addr, size_t length,
uint64_t hca_va, int access);
@@ -489,6 +553,7 @@ int hns_roce_u_close_xrcd(struct ibv_xrcd *ibv_xrcd);
int hns_roce_alloc_buf(struct hns_roce_buf *buf, unsigned int size,
int page_size);
void hns_roce_free_buf(struct hns_roce_buf *buf);
+void hns_roce_qp_spinlock_destroy(struct hns_roce_qp *qp);
void hns_roce_free_qp_buf(struct hns_roce_qp *qp, struct hns_roce_context *ctx);
diff --git a/providers/hns/hns_roce_u_verbs.c b/providers/hns/hns_roce_u_verbs.c
index 34f7ee4..89114f3 100644
--- a/providers/hns/hns_roce_u_verbs.c
+++ b/providers/hns/hns_roce_u_verbs.c
@@ -33,6 +33,7 @@
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
+#include <math.h>
#include <errno.h>
#include <pthread.h>
#include <sys/mman.h>
@@ -42,6 +43,37 @@
#include "hns_roce_u_db.h"
#include "hns_roce_u_hw_v2.h"
+static bool hns_roce_whether_need_lock(struct ibv_pd *pd)
+{
+ struct hns_roce_pad *pad;
+
+ pad = to_hr_pad(pd);
+ if (pad && pad->td)
+ return false;
+
+ return true;
+}
+
+static int hns_roce_spinlock_init(struct hns_roce_spinlock *hr_lock,
+ bool need_lock)
+{
+ hr_lock->need_lock = need_lock;
+
+ if (need_lock)
+ return pthread_spin_init(&hr_lock->lock,
+ PTHREAD_PROCESS_PRIVATE);
+
+ return 0;
+}
+
+static int hns_roce_spinlock_destroy(struct hns_roce_spinlock *hr_lock)
+{
+ if (hr_lock->need_lock)
+ return pthread_spin_destroy(&hr_lock->lock);
+
+ return 0;
+}
+
void hns_roce_init_qp_indices(struct hns_roce_qp *qp)
{
qp->sq.head = 0;
@@ -85,38 +117,138 @@ int hns_roce_u_query_port(struct ibv_context *context, uint8_t port,
return ibv_cmd_query_port(context, port, attr, &cmd, sizeof(cmd));
}
+struct ibv_td *hns_roce_u_alloc_td(struct ibv_context *context,
+ struct ibv_td_init_attr *attr)
+{
+ struct hns_roce_td *td;
+
+ if (attr->comp_mask) {
+ errno = EOPNOTSUPP;
+ return NULL;
+ }
+
+ td = calloc(1, sizeof(*td));
+ if (!td) {
+ errno = ENOMEM;
+ return NULL;
+ }
+
+ td->ibv_td.context = context;
+ atomic_init(&td->refcount, 1);
+
+ return &td->ibv_td;
+}
+
+int hns_roce_u_dealloc_td(struct ibv_td *ibv_td)
+{
+ struct hns_roce_td *td;
+
+ td = to_hr_td(ibv_td);
+ if (atomic_load(&td->refcount) > 1)
+ return EBUSY;
+
+ free(td);
+
+ return 0;
+}
+
struct ibv_pd *hns_roce_u_alloc_pd(struct ibv_context *context)
{
+ struct hns_roce_alloc_pd_resp resp = {};
struct ibv_alloc_pd cmd;
struct hns_roce_pd *pd;
- struct hns_roce_alloc_pd_resp resp = {};
-
- pd = malloc(sizeof(*pd));
- if (!pd)
- return NULL;
- if (ibv_cmd_alloc_pd(context, &pd->ibv_pd, &cmd, sizeof(cmd),
- &resp.ibv_resp, sizeof(resp))) {
- free(pd);
+ pd = calloc(1, sizeof(*pd));
+ if (!pd) {
+ errno = ENOMEM;
return NULL;
}
+ errno = ibv_cmd_alloc_pd(context, &pd->ibv_pd, &cmd, sizeof(cmd),
+ &resp.ibv_resp, sizeof(resp));
+ if (errno)
+ goto err;
+
+ atomic_init(&pd->refcount, 1);
pd->pdn = resp.pdn;
return &pd->ibv_pd;
+
+err:
+ free(pd);
+ return NULL;
+}
+
+struct ibv_pd *hns_roce_u_alloc_pad(struct ibv_context *context,
+ struct ibv_parent_domain_init_attr *attr)
+{
+ struct hns_roce_pad *pad;
+
+ if (ibv_check_alloc_parent_domain(attr))
+ return NULL;
+
+ if (attr->comp_mask) {
+ errno = EOPNOTSUPP;
+ return NULL;
+ }
+
+ pad = calloc(1, sizeof(*pad));
+ if (!pad) {
+ errno = ENOMEM;
+ return NULL;
+ }
+
+ if (attr->td) {
+ pad->td = to_hr_td(attr->td);
+ atomic_fetch_add(&pad->td->refcount, 1);
+ }
+
+ pad->pd.protection_domain = to_hr_pd(attr->pd);
+ atomic_fetch_add(&pad->pd.protection_domain->refcount, 1);
+
+ atomic_init(&pad->pd.refcount, 1);
+ ibv_initialize_parent_domain(&pad->pd.ibv_pd,
+ &pad->pd.protection_domain->ibv_pd);
+
+ return &pad->pd.ibv_pd;
+}
+
+static void hns_roce_free_pad(struct hns_roce_pad *pad)
+{
+ atomic_fetch_sub(&pad->pd.protection_domain->refcount, 1);
+
+ if (pad->td)
+ atomic_fetch_sub(&pad->td->refcount, 1);
+
+ free(pad);
}
-int hns_roce_u_free_pd(struct ibv_pd *pd)
+static int hns_roce_free_pd(struct hns_roce_pd *pd)
{
int ret;
- ret = ibv_cmd_dealloc_pd(pd);
+ if (atomic_load(&pd->refcount) > 1)
+ return EBUSY;
+
+ ret = ibv_cmd_dealloc_pd(&pd->ibv_pd);
if (ret)
return ret;
- free(to_hr_pd(pd));
+ free(pd);
+ return 0;
+}
- return ret;
+int hns_roce_u_dealloc_pd(struct ibv_pd *ibv_pd)
+{
+ struct hns_roce_pad *pad = to_hr_pad(ibv_pd);
+ struct hns_roce_pd *pd = to_hr_pd(ibv_pd);
+
+ if (pad) {
+ hns_roce_free_pad(pad);
+ return 0;
+ }
+
+ return hns_roce_free_pd(pd);
}
struct ibv_xrcd *hns_roce_u_open_xrcd(struct ibv_context *context,
--
2.30.0