| Message ID | 20190131130850.6850-10-yuval.shaia@oracle.com (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
| Series | Misc fixes to pvrdma device | expand |
On 1/31/19 3:08 PM, Yuval Shaia wrote: > 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. I think without 'a' and without 'are' > 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> > --- > hw/rdma/rdma_backend.c | 25 +++++++++++++++++++------ > hw/rdma/rdma_backend.h | 2 +- > hw/rdma/rdma_backend_defs.h | 1 + > hw/rdma/rdma_rm.c | 2 +- > 4 files changed, 22 insertions(+), 8 deletions(-) > > diff --git a/hw/rdma/rdma_backend.c b/hw/rdma/rdma_backend.c > index 65b28aa5b2..3283461b15 100644 > --- 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 { > @@ -90,7 +91,8 @@ static void clean_recv_mads(RdmaBackendDev *backend_dev) > } 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 +115,9 @@ static int rdma_poll_cq(RdmaDeviceResources *rdma_dev_res, struct ibv_cq *ibcq) > > comp_handler(bctx->up_ctx, &wc[i]); > > + bctx->backend_qp->cqe_ctx_list = > + g_slist_remove(bctx->backend_qp->cqe_ctx_list, > + GINT_TO_POINTER(wc[i].wr_id)); No synchronization needed ? Thanks, Marcel > rdma_rm_dealloc_cqe_ctx(rdma_dev_res, wc[i].wr_id); > g_free(bctx); > } > @@ -173,14 +178,12 @@ 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); > @@ -307,7 +310,7 @@ int rdma_backend_query_port(RdmaBackendDev *backend_dev, > int rdma_backend_poll_cq(RdmaDeviceResources *rdma_dev_res, RdmaBackendCQ *cq) > { > rdma_dev_res->stats.poll_cq_from_guest++; > - return rdma_poll_cq(rdma_dev_res, cq->ibcq); > + return rdma_poll_cq(cq->backend_dev, rdma_dev_res, cq->ibcq); > } > > static GHashTable *ah_hash; > @@ -494,6 +497,7 @@ 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)) { > @@ -501,6 +505,9 @@ void rdma_backend_post_send(RdmaBackendDev *backend_dev, > goto err_free_bctx; > } > > + qp->cqe_ctx_list = g_slist_append(qp->cqe_ctx_list, > + GINT_TO_POINTER(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) { > @@ -608,6 +615,7 @@ 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)) { > @@ -615,6 +623,9 @@ void rdma_backend_post_recv(RdmaBackendDev *backend_dev, > goto err_free_bctx; > } > > + qp->cqe_ctx_list = g_slist_append(qp->cqe_ctx_list, > + GINT_TO_POINTER(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) { > @@ -910,11 +921,13 @@ 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, free_cqe_ctx, dev_res); > + g_slist_free(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 6abc367a52..798a12f28f 100644 > --- a/hw/rdma/rdma_backend.h > +++ b/hw/rdma/rdma_backend.h > @@ -103,7 +103,7 @@ 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 bec0457f25..2306bbd18e 100644 > --- a/hw/rdma/rdma_backend_defs.h > +++ b/hw/rdma/rdma_backend_defs.h > @@ -66,6 +66,7 @@ typedef struct RdmaBackendQP { > struct ibv_pd *ibpd; > struct ibv_qp *ibqp; > uint8_t sgid_idx; > + GSList *cqe_ctx_list; > } RdmaBackendQP; > > #endif > diff --git a/hw/rdma/rdma_rm.c b/hw/rdma/rdma_rm.c > index ff536e356b..c7b9ef17f8 100644 > --- a/hw/rdma/rdma_rm.c > +++ b/hw/rdma/rdma_rm.c > @@ -485,7 +485,7 @@ 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); > }
On Wed, Feb 06, 2019 at 12:23:25PM +0200, Marcel Apfelbaum wrote: > > > On 1/31/19 3:08 PM, Yuval Shaia wrote: > > 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. > > I think without 'a' and without 'are' > > > 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> > > --- > > hw/rdma/rdma_backend.c | 25 +++++++++++++++++++------ > > hw/rdma/rdma_backend.h | 2 +- > > hw/rdma/rdma_backend_defs.h | 1 + > > hw/rdma/rdma_rm.c | 2 +- > > 4 files changed, 22 insertions(+), 8 deletions(-) > > > > diff --git a/hw/rdma/rdma_backend.c b/hw/rdma/rdma_backend.c > > index 65b28aa5b2..3283461b15 100644 > > --- 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 { > > @@ -90,7 +91,8 @@ static void clean_recv_mads(RdmaBackendDev *backend_dev) > > } 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 +115,9 @@ static int rdma_poll_cq(RdmaDeviceResources *rdma_dev_res, struct ibv_cq *ibcq) > > comp_handler(bctx->up_ctx, &wc[i]); > > + bctx->backend_qp->cqe_ctx_list = > > + g_slist_remove(bctx->backend_qp->cqe_ctx_list, > > + GINT_TO_POINTER(wc[i].wr_id)); > > No synchronization needed ? > > Thanks, > Marcel Agree, thanks! > > > rdma_rm_dealloc_cqe_ctx(rdma_dev_res, wc[i].wr_id); > > g_free(bctx); > > } > > @@ -173,14 +178,12 @@ 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); > > @@ -307,7 +310,7 @@ int rdma_backend_query_port(RdmaBackendDev *backend_dev, > > int rdma_backend_poll_cq(RdmaDeviceResources *rdma_dev_res, RdmaBackendCQ *cq) > > { > > rdma_dev_res->stats.poll_cq_from_guest++; > > - return rdma_poll_cq(rdma_dev_res, cq->ibcq); > > + return rdma_poll_cq(cq->backend_dev, rdma_dev_res, cq->ibcq); > > } > > static GHashTable *ah_hash; > > @@ -494,6 +497,7 @@ 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)) { > > @@ -501,6 +505,9 @@ void rdma_backend_post_send(RdmaBackendDev *backend_dev, > > goto err_free_bctx; > > } > > + qp->cqe_ctx_list = g_slist_append(qp->cqe_ctx_list, > > + GINT_TO_POINTER(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) { > > @@ -608,6 +615,7 @@ 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)) { > > @@ -615,6 +623,9 @@ void rdma_backend_post_recv(RdmaBackendDev *backend_dev, > > goto err_free_bctx; > > } > > + qp->cqe_ctx_list = g_slist_append(qp->cqe_ctx_list, > > + GINT_TO_POINTER(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) { > > @@ -910,11 +921,13 @@ 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, free_cqe_ctx, dev_res); > > + g_slist_free(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 6abc367a52..798a12f28f 100644 > > --- a/hw/rdma/rdma_backend.h > > +++ b/hw/rdma/rdma_backend.h > > @@ -103,7 +103,7 @@ 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 bec0457f25..2306bbd18e 100644 > > --- a/hw/rdma/rdma_backend_defs.h > > +++ b/hw/rdma/rdma_backend_defs.h > > @@ -66,6 +66,7 @@ typedef struct RdmaBackendQP { > > struct ibv_pd *ibpd; > > struct ibv_qp *ibqp; > > uint8_t sgid_idx; > > + GSList *cqe_ctx_list; > > } RdmaBackendQP; > > #endif > > diff --git a/hw/rdma/rdma_rm.c b/hw/rdma/rdma_rm.c > > index ff536e356b..c7b9ef17f8 100644 > > --- a/hw/rdma/rdma_rm.c > > +++ b/hw/rdma/rdma_rm.c > > @@ -485,7 +485,7 @@ 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_backend.c b/hw/rdma/rdma_backend.c index 65b28aa5b2..3283461b15 100644 --- 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 { @@ -90,7 +91,8 @@ static void clean_recv_mads(RdmaBackendDev *backend_dev) } 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 +115,9 @@ static int rdma_poll_cq(RdmaDeviceResources *rdma_dev_res, struct ibv_cq *ibcq) comp_handler(bctx->up_ctx, &wc[i]); + bctx->backend_qp->cqe_ctx_list = + g_slist_remove(bctx->backend_qp->cqe_ctx_list, + GINT_TO_POINTER(wc[i].wr_id)); rdma_rm_dealloc_cqe_ctx(rdma_dev_res, wc[i].wr_id); g_free(bctx); } @@ -173,14 +178,12 @@ 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); @@ -307,7 +310,7 @@ int rdma_backend_query_port(RdmaBackendDev *backend_dev, int rdma_backend_poll_cq(RdmaDeviceResources *rdma_dev_res, RdmaBackendCQ *cq) { rdma_dev_res->stats.poll_cq_from_guest++; - return rdma_poll_cq(rdma_dev_res, cq->ibcq); + return rdma_poll_cq(cq->backend_dev, rdma_dev_res, cq->ibcq); } static GHashTable *ah_hash; @@ -494,6 +497,7 @@ 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)) { @@ -501,6 +505,9 @@ void rdma_backend_post_send(RdmaBackendDev *backend_dev, goto err_free_bctx; } + qp->cqe_ctx_list = g_slist_append(qp->cqe_ctx_list, + GINT_TO_POINTER(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) { @@ -608,6 +615,7 @@ 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)) { @@ -615,6 +623,9 @@ void rdma_backend_post_recv(RdmaBackendDev *backend_dev, goto err_free_bctx; } + qp->cqe_ctx_list = g_slist_append(qp->cqe_ctx_list, + GINT_TO_POINTER(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) { @@ -910,11 +921,13 @@ 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, free_cqe_ctx, dev_res); + g_slist_free(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 6abc367a52..798a12f28f 100644 --- a/hw/rdma/rdma_backend.h +++ b/hw/rdma/rdma_backend.h @@ -103,7 +103,7 @@ 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 bec0457f25..2306bbd18e 100644 --- a/hw/rdma/rdma_backend_defs.h +++ b/hw/rdma/rdma_backend_defs.h @@ -66,6 +66,7 @@ typedef struct RdmaBackendQP { struct ibv_pd *ibpd; struct ibv_qp *ibqp; uint8_t sgid_idx; + GSList *cqe_ctx_list; } RdmaBackendQP; #endif diff --git a/hw/rdma/rdma_rm.c b/hw/rdma/rdma_rm.c index ff536e356b..c7b9ef17f8 100644 --- a/hw/rdma/rdma_rm.c +++ b/hw/rdma/rdma_rm.c @@ -485,7 +485,7 @@ 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); }
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> --- hw/rdma/rdma_backend.c | 25 +++++++++++++++++++------ hw/rdma/rdma_backend.h | 2 +- hw/rdma/rdma_backend_defs.h | 1 + hw/rdma/rdma_rm.c | 2 +- 4 files changed, 22 insertions(+), 8 deletions(-)