hw/rdma: Free all receive buffers when QP is destroyed

When QP is destroyed the backend QP is destroyed as well. This ensures
we clean all received buffer we posted to it.
However, a contexts of these buffers are still remain in the device.
Fix it by maintaining a list of buffer's context and free them when QP
is destroyed.

Signed-off-by: Yuval Shaia <yuval.shaia@oracle.com>
Reviewed-by: Marcel Apfelbaum <marcel.apfelbaum@gmail.com>
Message-Id: <1552300155-25216-8-git-send-email-yuval.shaia@oracle.com>
Reviewed-by: Kamal Heib <kamalheib1@gmail.com>
Signed-off-by: Marcel Apfelbaum <marcel.apfelbaum@gmail.com>

diff --git a/hw/rdma/rdma_backend.c b/hw/rdma/rdma_backend.c
index d0bbe57bd2adb6837f377182bd7e7d9d31c968da..e124d8d16bb1a74aa727708b68e77dce68e1763e 100644 (file)
--- a/hw/rdma/rdma_backend.c
+++ b/hw/rdma/rdma_backend.c
@@ -39,6 +39,7 @@
 typedef struct BackendCtx {
     void *up_ctx;
     struct ibv_sge sge; /* Used to save MAD recv buffer */
+    RdmaBackendQP *backend_qp; /* To maintain recv buffers */
 } BackendCtx;
 
 struct backend_umad {
@@ -73,6 +74,7 @@ static void free_cqe_ctx(gpointer data, gpointer user_data)
     bctx = rdma_rm_get_cqe_ctx(rdma_dev_res, cqe_ctx_id);
     if (bctx) {
         rdma_rm_dealloc_cqe_ctx(rdma_dev_res, cqe_ctx_id);
+        atomic_dec(&rdma_dev_res->stats.missing_cqe);
     }
     g_free(bctx);
 }
@@ -85,13 +87,15 @@ static void clean_recv_mads(RdmaBackendDev *backend_dev)
         cqe_ctx_id = rdma_protected_qlist_pop_int64(&backend_dev->
                                                     recv_mads_list);
         if (cqe_ctx_id != -ENOENT) {
+            atomic_inc(&backend_dev->rdma_dev_res->stats.missing_cqe);
             free_cqe_ctx(GINT_TO_POINTER(cqe_ctx_id),
                          backend_dev->rdma_dev_res);
         }
     } while (cqe_ctx_id != -ENOENT);
 }
 
-static int rdma_poll_cq(RdmaDeviceResources *rdma_dev_res, struct ibv_cq *ibcq)
+static int rdma_poll_cq(RdmaBackendDev *backend_dev,
+                        RdmaDeviceResources *rdma_dev_res, struct ibv_cq *ibcq)
 {
     int i, ne, total_ne = 0;
     BackendCtx *bctx;
@@ -113,6 +117,8 @@ static int rdma_poll_cq(RdmaDeviceResources *rdma_dev_res, struct ibv_cq *ibcq)
 
             comp_handler(bctx->up_ctx, &wc[i]);
 
+            rdma_protected_gslist_remove_int32(&bctx->backend_qp->cqe_ctx_list,
+                                               wc[i].wr_id);
             rdma_rm_dealloc_cqe_ctx(rdma_dev_res, wc[i].wr_id);
             g_free(bctx);
         }
@@ -175,14 +181,12 @@ static void *comp_handler_thread(void *arg)
             }
 
             backend_dev->rdma_dev_res->stats.poll_cq_from_bk++;
-            rdma_poll_cq(backend_dev->rdma_dev_res, ev_cq);
+            rdma_poll_cq(backend_dev, backend_dev->rdma_dev_res, ev_cq);
 
             ibv_ack_cq_events(ev_cq, 1);
         }
     }
 
-    /* TODO: Post cqe for all remaining buffs that were posted */
-
     backend_dev->comp_thread.is_running = false;
 
     qemu_thread_exit(0);
@@ -311,7 +315,7 @@ void rdma_backend_poll_cq(RdmaDeviceResources *rdma_dev_res, RdmaBackendCQ *cq)
     int polled;
 
     rdma_dev_res->stats.poll_cq_from_guest++;
-    polled = rdma_poll_cq(rdma_dev_res, cq->ibcq);
+    polled = rdma_poll_cq(cq->backend_dev, rdma_dev_res, cq->ibcq);
     if (!polled) {
         rdma_dev_res->stats.poll_cq_from_guest_empty++;
     }
@@ -501,6 +505,7 @@ void rdma_backend_post_send(RdmaBackendDev *backend_dev,
 
     bctx = g_malloc0(sizeof(*bctx));
     bctx->up_ctx = ctx;
+    bctx->backend_qp = qp;
 
     rc = rdma_rm_alloc_cqe_ctx(backend_dev->rdma_dev_res, &bctx_id, bctx);
     if (unlikely(rc)) {
@@ -508,6 +513,8 @@ void rdma_backend_post_send(RdmaBackendDev *backend_dev,
         goto err_free_bctx;
     }
 
+    rdma_protected_gslist_append_int32(&qp->cqe_ctx_list, bctx_id);
+
     rc = build_host_sge_array(backend_dev->rdma_dev_res, new_sge, sge, num_sge,
                               &backend_dev->rdma_dev_res->stats.tx_len);
     if (rc) {
@@ -616,6 +623,7 @@ void rdma_backend_post_recv(RdmaBackendDev *backend_dev,
 
     bctx = g_malloc0(sizeof(*bctx));
     bctx->up_ctx = ctx;
+    bctx->backend_qp = qp;
 
     rc = rdma_rm_alloc_cqe_ctx(rdma_dev_res, &bctx_id, bctx);
     if (unlikely(rc)) {
@@ -623,6 +631,8 @@ void rdma_backend_post_recv(RdmaBackendDev *backend_dev,
         goto err_free_bctx;
     }
 
+    rdma_protected_gslist_append_int32(&qp->cqe_ctx_list, bctx_id);
+
     rc = build_host_sge_array(rdma_dev_res, new_sge, sge, num_sge,
                               &backend_dev->rdma_dev_res->stats.rx_bufs_len);
     if (rc) {
@@ -762,6 +772,8 @@ int rdma_backend_create_qp(RdmaBackendQP *qp, uint8_t qp_type,
         return -EIO;
     }
 
+    rdma_protected_gslist_init(&qp->cqe_ctx_list);
+
     qp->ibpd = pd->ibpd;
 
     /* TODO: Query QP to get max_inline_data and save it to be used in send */
@@ -919,11 +931,13 @@ int rdma_backend_query_qp(RdmaBackendQP *qp, struct ibv_qp_attr *attr,
     return ibv_query_qp(qp->ibqp, attr, attr_mask, init_attr);
 }
 
-void rdma_backend_destroy_qp(RdmaBackendQP *qp)
+void rdma_backend_destroy_qp(RdmaBackendQP *qp, RdmaDeviceResources *dev_res)
 {
     if (qp->ibqp) {
         ibv_destroy_qp(qp->ibqp);
     }
+    g_slist_foreach(qp->cqe_ctx_list.list, free_cqe_ctx, dev_res);
+    rdma_protected_gslist_destroy(&qp->cqe_ctx_list);
 }
 
 #define CHK_ATTR(req, dev, member, fmt) ({ \

diff --git a/hw/rdma/rdma_backend.h b/hw/rdma/rdma_backend.h
index 8e53a72bf26fbafb224e206d4f0b9332b0558a4d..c54eaf2e4abac2fc1b13cb4f31e8e38126869a6d 100644 (file)
--- a/hw/rdma/rdma_backend.h
+++ b/hw/rdma/rdma_backend.h
@@ -102,7 +102,7 @@ int rdma_backend_qp_state_rts(RdmaBackendQP *qp, uint8_t qp_type,
                               uint32_t sq_psn, uint32_t qkey, bool use_qkey);
 int rdma_backend_query_qp(RdmaBackendQP *qp, struct ibv_qp_attr *attr,
                           int attr_mask, struct ibv_qp_init_attr *init_attr);
-void rdma_backend_destroy_qp(RdmaBackendQP *qp);
+void rdma_backend_destroy_qp(RdmaBackendQP *qp, RdmaDeviceResources *dev_res);
 
 void rdma_backend_post_send(RdmaBackendDev *backend_dev,
                             RdmaBackendQP *qp, uint8_t qp_type,

diff --git a/hw/rdma/rdma_backend_defs.h b/hw/rdma/rdma_backend_defs.h
index a8c15b09abd5b672964396d9b8f3cede91c915b2..817153dc8cf412ca4aa9feeb5c87f363a2c1e3b9 100644 (file)
--- a/hw/rdma/rdma_backend_defs.h
+++ b/hw/rdma/rdma_backend_defs.h
@@ -26,7 +26,6 @@ typedef struct RdmaDeviceResources RdmaDeviceResources;
 
 typedef struct RdmaBackendThread {
     QemuThread thread;
-    QemuMutex mutex;
     bool run; /* Set by thread manager to let thread know it should exit */
     bool is_running; /* Set by the thread to report its status */
 } RdmaBackendThread;
@@ -66,6 +65,7 @@ typedef struct RdmaBackendQP {
     struct ibv_pd *ibpd;
     struct ibv_qp *ibqp;
     uint8_t sgid_idx;
+    RdmaProtectedGSList cqe_ctx_list;
 } RdmaBackendQP;
 
 #endif

diff --git a/hw/rdma/rdma_rm.c b/hw/rdma/rdma_rm.c
index b50e192b4989f96886aa60334076e46f61a7fb38..bac3b2f4a6c3254fa4a6f33945b61e7209c08eda 100644 (file)
--- a/hw/rdma/rdma_rm.c
+++ b/hw/rdma/rdma_rm.c
@@ -537,7 +537,7 @@ void rdma_rm_dealloc_qp(RdmaDeviceResources *dev_res, uint32_t qp_handle)
         return;
     }
 
-    rdma_backend_destroy_qp(&qp->backend_qp);
+    rdma_backend_destroy_qp(&qp->backend_qp, dev_res);
 
     rdma_res_tbl_dealloc(&dev_res->qp_tbl, qp->qpn);
 }

diff --git a/hw/rdma/rdma_utils.c b/hw/rdma/rdma_utils.c
index 0a8abe572d1e0125b92d295dbc93a80b716b50f3..73f279104c37d10630b360ec7405faf691deb438 100644 (file)
--- a/hw/rdma/rdma_utils.c
+++ b/hw/rdma/rdma_utils.c
@@ -90,3 +90,32 @@ int64_t rdma_protected_qlist_pop_int64(RdmaProtectedQList *list)
 
     return qnum_get_uint(qobject_to(QNum, obj));
 }
+
+void rdma_protected_gslist_init(RdmaProtectedGSList *list)
+{
+    qemu_mutex_init(&list->lock);
+}
+
+void rdma_protected_gslist_destroy(RdmaProtectedGSList *list)
+{
+    if (list->list) {
+        g_slist_free(list->list);
+        list->list = NULL;
+    }
+}
+
+void rdma_protected_gslist_append_int32(RdmaProtectedGSList *list,
+                                        int32_t value)
+{
+    qemu_mutex_lock(&list->lock);
+    list->list = g_slist_prepend(list->list, GINT_TO_POINTER(value));
+    qemu_mutex_unlock(&list->lock);
+}
+
+void rdma_protected_gslist_remove_int32(RdmaProtectedGSList *list,
+                                        int32_t value)
+{
+    qemu_mutex_lock(&list->lock);
+    list->list = g_slist_remove(list->list, GINT_TO_POINTER(value));
+    qemu_mutex_unlock(&list->lock);
+}

diff --git a/hw/rdma/rdma_utils.h b/hw/rdma/rdma_utils.h
index a8bf1d4fec0079351b2871c2cd937a451e8d0ee5..2d4224969158906594b8e8b1da465e811529ab13 100644 (file)
--- a/hw/rdma/rdma_utils.h
+++ b/hw/rdma/rdma_utils.h
@@ -34,12 +34,23 @@ typedef struct RdmaProtectedQList {
     QList *list;
 } RdmaProtectedQList;
 
+typedef struct RdmaProtectedGSList {
+    QemuMutex lock;
+    GSList *list;
+} RdmaProtectedGSList;
+
 void *rdma_pci_dma_map(PCIDevice *dev, dma_addr_t addr, dma_addr_t plen);
 void rdma_pci_dma_unmap(PCIDevice *dev, void *buffer, dma_addr_t len);
 void rdma_protected_qlist_init(RdmaProtectedQList *list);
 void rdma_protected_qlist_destroy(RdmaProtectedQList *list);
 void rdma_protected_qlist_append_int64(RdmaProtectedQList *list, int64_t value);
 int64_t rdma_protected_qlist_pop_int64(RdmaProtectedQList *list);
+void rdma_protected_gslist_init(RdmaProtectedGSList *list);
+void rdma_protected_gslist_destroy(RdmaProtectedGSList *list);
+void rdma_protected_gslist_append_int32(RdmaProtectedGSList *list,
+                                        int32_t value);
+void rdma_protected_gslist_remove_int32(RdmaProtectedGSList *list,
+                                        int32_t value);
 
 static inline void addrconf_addr_eui48(uint8_t *eui, const char *addr)
 {