efa: Add support to allocate a parent domain

Parent domains are an extension for PDs and are interchangeable with
them. Add support for parent domain allocations and deallocations.
Since alloc parent domain verb is strictly a userspace verb and it
binds an existing PD to it we can't dealloc the PD if it has parent
domains referencing it so to handle it add PD refcount for parent
domains.
Also parent domains can't be deallocated while there are active objects
(CQs/QPs etc.) using it, to prevent its deallocation add a refcount for
it.

Signed-off-by: Yonatan Nachum <[email protected]>

Author: YonatanNachum

providers/efa/efa.c

@@ -28,6 +28,7 @@ static const struct verbs_match_ent efa_table[] = {
 
 static const struct verbs_context_ops efa_ctx_ops = {
 	.alloc_pd = efa_alloc_pd,
+	.alloc_parent_domain = efa_alloc_parent_domain,
 	.alloc_td = efa_alloc_td,
 	.create_ah = efa_create_ah,
 	.create_cq = efa_create_cq,

providers/efa/efa.h

@@ -51,6 +51,10 @@ struct efa_context {
 struct efa_pd {
 	struct ibv_pd ibvpd;
 	uint16_t pdn;
+	/* Pointer to original PD used to create parent domain */
+	struct efa_pd *orig_pd;
+	/* Number of parent domains referencing this PD */
+	atomic_int refcount;
 };
 
 struct efa_td {
@@ -59,6 +63,14 @@ struct efa_td {
 	atomic_int refcount;
 };
 
+struct efa_parent_domain {
+	struct efa_pd pd;
+	struct efa_td *td;
+	/* Number of objects referencing this parent domain */
+	atomic_int refcount;
+	void *pd_context;
+};
+
 struct efa_sub_cq {
 	uint16_t consumed_cnt;
 	int phase;
@@ -212,6 +224,11 @@ static inline struct efa_td *to_efa_td(struct ibv_td *ibvtd)
 	return container_of(ibvtd, struct efa_td, ibvtd);
 }
 
+static inline struct efa_parent_domain *to_efa_parent_domain(struct ibv_pd *ibvpd)
+{
+	return container_of(ibvpd, struct efa_parent_domain, pd.ibvpd);
+}
+
 bool is_efa_dev(struct ibv_device *device);
 
 #endif /* __EFA_H__ */

providers/efa/verbs.c

@@ -213,6 +213,7 @@ struct ibv_pd *efa_alloc_pd(struct ibv_context *ibvctx)
 		goto out;
 	}
 
+	atomic_init(&pd->refcount, 0);
 	pd->pdn = resp.pdn;
 
 	return &pd->ibvpd;
@@ -223,11 +224,84 @@ struct ibv_pd *efa_alloc_pd(struct ibv_context *ibvctx)
 	return NULL;
 }
 
+struct ibv_pd *efa_alloc_parent_domain(struct ibv_context *ibvctx,
+				       struct ibv_parent_domain_init_attr *attr)
+{
+	struct efa_parent_domain *parent_domain;
+	struct efa_pd *pd;
+
+	if (ibv_check_alloc_parent_domain(attr)) {
+		errno = EINVAL;
+		return NULL;
+	}
+
+	if (!check_comp_mask(attr->comp_mask,
+			     IBV_PARENT_DOMAIN_INIT_ATTR_PD_CONTEXT)) {
+		verbs_err(verbs_get_ctx(ibvctx), "Invalid comp_mask\n");
+		errno = EOPNOTSUPP;
+		return NULL;
+	}
+
+	pd = to_efa_pd(attr->pd);
+
+	/* We don't allow nested parent domains */
+	if (pd->orig_pd) {
+		errno = EINVAL;
+		return NULL;
+	}
+
+	parent_domain = calloc(1, sizeof(*parent_domain));
+	if (!parent_domain) {
+		errno = ENOMEM;
+		return NULL;
+	}
+
+	atomic_init(&parent_domain->refcount, 0);
+
+	if (attr->td) {
+		parent_domain->td = to_efa_td(attr->td);
+		atomic_fetch_add(&parent_domain->td->refcount, 1);
+	}
+
+	parent_domain->pd.orig_pd = pd;
+	atomic_fetch_add(&pd->refcount, 1);
+
+	ibv_initialize_parent_domain(&parent_domain->pd.ibvpd, attr->pd);
+
+	if (attr->comp_mask & IBV_PARENT_DOMAIN_INIT_ATTR_PD_CONTEXT)
+		parent_domain->pd_context = attr->pd_context;
+
+	return &parent_domain->pd.ibvpd;
+}
+
+static int efa_dealloc_parent_domain(struct efa_parent_domain *parent_domain)
+{
+	if (atomic_load(&parent_domain->refcount) > 0)
+		return EBUSY;
+
+	atomic_fetch_sub(&parent_domain->pd.orig_pd->refcount, 1);
+
+	if (parent_domain->td)
+		atomic_fetch_sub(&parent_domain->td->refcount, 1);
+
+	free(parent_domain);
+	return 0;
+}
+
 int efa_dealloc_pd(struct ibv_pd *ibvpd)
 {
+	struct efa_parent_domain *parent_domain;
 	struct efa_pd *pd = to_efa_pd(ibvpd);
 	int err;
 
+	if (pd->orig_pd) {
+		parent_domain = to_efa_parent_domain(ibvpd);
+		return efa_dealloc_parent_domain(parent_domain);
+	}
+
+	if (atomic_load(&pd->refcount) > 0)
+		return EBUSY;
+
 	err = ibv_cmd_dealloc_pd(ibvpd);
 	if (err) {
 		verbs_err(verbs_get_ctx(ibvpd->context),

providers/efa/verbs.h

@@ -16,6 +16,8 @@ int efa_query_device_ex(struct ibv_context *context,
 			const struct ibv_query_device_ex_input *input,
 			struct ibv_device_attr_ex *attr, size_t attr_size);
 struct ibv_pd *efa_alloc_pd(struct ibv_context *uctx);
+struct ibv_pd *efa_alloc_parent_domain(struct ibv_context *ibvctx,
+				       struct ibv_parent_domain_init_attr *attr);
 int efa_dealloc_pd(struct ibv_pd *ibvpd);
 struct ibv_mr *efa_reg_dmabuf_mr(struct ibv_pd *pd, uint64_t offset,
 				 size_t length, uint64_t iova, int fd, int acc);