Message ID | 4-v1-34e141ddf17e+89-query_device_ex_jgg@nvidia.com (mailing list archive) |
---|---|
State | Not Applicable |
Headers | show |
Series | Simplify query_device() in libibverbs | expand |
On 16/11/2020 22:23, Jason Gunthorpe wrote: > When the user calls efa_query_device_ex() it should not cause the context > values to be mutated, only the attribute shuld be returned. > > Move this code to a dedicated function that is only called during context > setup. > > Cc: Gal Pressman <galpress@amazon.com> > Signed-off-by: Jason Gunthorpe <jgg@nvidia.com> Didn't get a chance to review yet, but this one breaks EFA. I'll try to provide more info today/tomorrow. (BTW: typo in the subject line "intialization", and "shuld" in the commit message)
On 16/11/2020 22:23, Jason Gunthorpe wrote: > When the user calls efa_query_device_ex() it should not cause the context > values to be mutated, only the attribute shuld be returned. > > Move this code to a dedicated function that is only called during context > setup. > > Cc: Gal Pressman <galpress@amazon.com> > Signed-off-by: Jason Gunthorpe <jgg@nvidia.com> > --- > providers/efa/efa.c | 14 +------------ > providers/efa/verbs.c | 46 +++++++++++++++++++++++++++++++++++-------- > providers/efa/verbs.h | 1 + > 3 files changed, 40 insertions(+), 21 deletions(-) > > diff --git a/providers/efa/efa.c b/providers/efa/efa.c > index 35f9b246a711ec..b24c14f7fa1fe1 100644 > --- a/providers/efa/efa.c > +++ b/providers/efa/efa.c > @@ -54,10 +54,7 @@ static struct verbs_context *efa_alloc_context(struct ibv_device *vdev, > { > struct efa_alloc_ucontext_resp resp = {}; > struct efa_alloc_ucontext cmd = {}; > - struct ibv_device_attr_ex attr; > - unsigned int qp_table_sz; > struct efa_context *ctx; > - int err; > > cmd.comp_mask |= EFA_ALLOC_UCONTEXT_CMD_COMP_TX_BATCH; > cmd.comp_mask |= EFA_ALLOC_UCONTEXT_CMD_COMP_MIN_SQ_WR; > @@ -86,17 +83,8 @@ static struct verbs_context *efa_alloc_context(struct ibv_device *vdev, > > verbs_set_ops(&ctx->ibvctx, &efa_ctx_ops); > > - err = efa_query_device_ex(&ctx->ibvctx.context, NULL, &attr, > - sizeof(attr)); > - if (err) > + if (!efa_query_device_ctx(ctx)) > goto err_free_spinlock; This return error on success seems to be the issue, will verify.
On 16/11/2020 22:23, Jason Gunthorpe wrote: > When the user calls efa_query_device_ex() it should not cause the context > values to be mutated, only the attribute shuld be returned. > > Move this code to a dedicated function that is only called during context > setup. > > Cc: Gal Pressman <galpress@amazon.com> > Signed-off-by: Jason Gunthorpe <jgg@nvidia.com> > --- > providers/efa/efa.c | 14 +------------ > providers/efa/verbs.c | 46 +++++++++++++++++++++++++++++++++++-------- > providers/efa/verbs.h | 1 + > 3 files changed, 40 insertions(+), 21 deletions(-) > > diff --git a/providers/efa/efa.c b/providers/efa/efa.c > index 35f9b246a711ec..b24c14f7fa1fe1 100644 > --- a/providers/efa/efa.c > +++ b/providers/efa/efa.c > @@ -54,10 +54,7 @@ static struct verbs_context *efa_alloc_context(struct ibv_device *vdev, > { > struct efa_alloc_ucontext_resp resp = {}; > struct efa_alloc_ucontext cmd = {}; > - struct ibv_device_attr_ex attr; > - unsigned int qp_table_sz; > struct efa_context *ctx; > - int err; > > cmd.comp_mask |= EFA_ALLOC_UCONTEXT_CMD_COMP_TX_BATCH; > cmd.comp_mask |= EFA_ALLOC_UCONTEXT_CMD_COMP_MIN_SQ_WR; > @@ -86,17 +83,8 @@ static struct verbs_context *efa_alloc_context(struct ibv_device *vdev, > > verbs_set_ops(&ctx->ibvctx, &efa_ctx_ops); > > - err = efa_query_device_ex(&ctx->ibvctx.context, NULL, &attr, > - sizeof(attr)); > - if (err) > + if (!efa_query_device_ctx(ctx)) Remove the not. > goto err_free_spinlock; > - > - qp_table_sz = roundup_pow_of_two(attr.orig_attr.max_qp); > - ctx->qp_table_sz_m1 = qp_table_sz - 1; > - ctx->qp_table = calloc(qp_table_sz, sizeof(*ctx->qp_table)); > - if (!ctx->qp_table) > - goto err_free_spinlock; > - > return &ctx->ibvctx; > > err_free_spinlock: > diff --git a/providers/efa/verbs.c b/providers/efa/verbs.c > index 1a9633155c62f8..52d6285f1f409c 100644 > --- a/providers/efa/verbs.c > +++ b/providers/efa/verbs.c > @@ -106,14 +106,6 @@ int efa_query_device_ex(struct ibv_context *context, > if (err) > return err; > > - ctx->device_caps = resp.device_caps; > - ctx->max_sq_wr = resp.max_sq_wr; > - ctx->max_rq_wr = resp.max_rq_wr; > - ctx->max_sq_sge = resp.max_sq_sge; > - ctx->max_rq_sge = resp.max_rq_sge; > - ctx->max_rdma_size = resp.max_rdma_size; > - ctx->max_wr_rdma_sge = a->max_sge_rd; > - > a->max_qp_wr = min_t(int, a->max_qp_wr, > ctx->max_llq_size / sizeof(struct efa_io_tx_wqe)); > snprintf(a->fw_ver, sizeof(a->fw_ver), "%u.%u.%u.%u", > @@ -122,6 +114,44 @@ int efa_query_device_ex(struct ibv_context *context, > return 0; > } > > +int efa_query_device_ctx(struct efa_context *ctx) > +{ > + struct ibv_device_attr_ex attr; > + struct efa_query_device_ex_resp resp; Preferably I would put this first. > + size_t resp_size = sizeof(resp); > + unsigned int qp_table_sz; > + int err; > + > + if (ctx->cmds_supp_udata_mask & EFA_USER_CMDS_SUPP_UDATA_QUERY_DEVICE) { > + err = ibv_cmd_query_device_any(&ctx->ibvctx.context, NULL, > + &attr, sizeof(attr), > + &resp.ibv_resp, &resp_size); > + if (err) > + return err; > + > + ctx->device_caps = resp.device_caps; > + ctx->max_sq_wr = resp.max_sq_wr; > + ctx->max_rq_wr = resp.max_rq_wr; > + ctx->max_sq_sge = resp.max_sq_sge; > + ctx->max_rq_sge = resp.max_rq_sge; > + ctx->max_rdma_size = resp.max_rdma_size; > + ctx->max_wr_rdma_sge = attr.orig_attr.max_sge_rd; max_wr_rdma_sge assignment can be done in the else clause as well. > + } else { > + err = ibv_cmd_query_device_any(&ctx->ibvctx.context, NULL, > + &attr, sizeof(attr.orig_attr), > + NULL, NULL); > + if (err) > + return err; > + } > + > + qp_table_sz = roundup_pow_of_two(attr.orig_attr.max_qp); > + ctx->qp_table_sz_m1 = qp_table_sz - 1; > + ctx->qp_table = calloc(qp_table_sz, sizeof(*ctx->qp_table)); > + if (!ctx->qp_table) > + return ENOMEM; > + return 0; > +}
On Wed, Nov 18, 2020 at 02:45:42PM +0200, Gal Pressman wrote: > > + size_t resp_size = sizeof(resp); > > + unsigned int qp_table_sz; > > + int err; > > + > > + if (ctx->cmds_supp_udata_mask & EFA_USER_CMDS_SUPP_UDATA_QUERY_DEVICE) { > > + err = ibv_cmd_query_device_any(&ctx->ibvctx.context, NULL, > > + &attr, sizeof(attr), > > + &resp.ibv_resp, &resp_size); > > + if (err) > > + return err; > > + > > + ctx->device_caps = resp.device_caps; > > + ctx->max_sq_wr = resp.max_sq_wr; > > + ctx->max_rq_wr = resp.max_rq_wr; > > + ctx->max_sq_sge = resp.max_sq_sge; > > + ctx->max_rq_sge = resp.max_rq_sge; > > + ctx->max_rdma_size = resp.max_rdma_size; > > + ctx->max_wr_rdma_sge = attr.orig_attr.max_sge_rd; > > max_wr_rdma_sge assignment can be done in the else clause as well. Yes, it is the same mistake as I did in mlx5 I updated everything, thanks Jason
diff --git a/providers/efa/efa.c b/providers/efa/efa.c index 35f9b246a711ec..b24c14f7fa1fe1 100644 --- a/providers/efa/efa.c +++ b/providers/efa/efa.c @@ -54,10 +54,7 @@ static struct verbs_context *efa_alloc_context(struct ibv_device *vdev, { struct efa_alloc_ucontext_resp resp = {}; struct efa_alloc_ucontext cmd = {}; - struct ibv_device_attr_ex attr; - unsigned int qp_table_sz; struct efa_context *ctx; - int err; cmd.comp_mask |= EFA_ALLOC_UCONTEXT_CMD_COMP_TX_BATCH; cmd.comp_mask |= EFA_ALLOC_UCONTEXT_CMD_COMP_MIN_SQ_WR; @@ -86,17 +83,8 @@ static struct verbs_context *efa_alloc_context(struct ibv_device *vdev, verbs_set_ops(&ctx->ibvctx, &efa_ctx_ops); - err = efa_query_device_ex(&ctx->ibvctx.context, NULL, &attr, - sizeof(attr)); - if (err) + if (!efa_query_device_ctx(ctx)) goto err_free_spinlock; - - qp_table_sz = roundup_pow_of_two(attr.orig_attr.max_qp); - ctx->qp_table_sz_m1 = qp_table_sz - 1; - ctx->qp_table = calloc(qp_table_sz, sizeof(*ctx->qp_table)); - if (!ctx->qp_table) - goto err_free_spinlock; - return &ctx->ibvctx; err_free_spinlock: diff --git a/providers/efa/verbs.c b/providers/efa/verbs.c index 1a9633155c62f8..52d6285f1f409c 100644 --- a/providers/efa/verbs.c +++ b/providers/efa/verbs.c @@ -106,14 +106,6 @@ int efa_query_device_ex(struct ibv_context *context, if (err) return err; - ctx->device_caps = resp.device_caps; - ctx->max_sq_wr = resp.max_sq_wr; - ctx->max_rq_wr = resp.max_rq_wr; - ctx->max_sq_sge = resp.max_sq_sge; - ctx->max_rq_sge = resp.max_rq_sge; - ctx->max_rdma_size = resp.max_rdma_size; - ctx->max_wr_rdma_sge = a->max_sge_rd; - a->max_qp_wr = min_t(int, a->max_qp_wr, ctx->max_llq_size / sizeof(struct efa_io_tx_wqe)); snprintf(a->fw_ver, sizeof(a->fw_ver), "%u.%u.%u.%u", @@ -122,6 +114,44 @@ int efa_query_device_ex(struct ibv_context *context, return 0; } +int efa_query_device_ctx(struct efa_context *ctx) +{ + struct ibv_device_attr_ex attr; + struct efa_query_device_ex_resp resp; + size_t resp_size = sizeof(resp); + unsigned int qp_table_sz; + int err; + + if (ctx->cmds_supp_udata_mask & EFA_USER_CMDS_SUPP_UDATA_QUERY_DEVICE) { + err = ibv_cmd_query_device_any(&ctx->ibvctx.context, NULL, + &attr, sizeof(attr), + &resp.ibv_resp, &resp_size); + if (err) + return err; + + ctx->device_caps = resp.device_caps; + ctx->max_sq_wr = resp.max_sq_wr; + ctx->max_rq_wr = resp.max_rq_wr; + ctx->max_sq_sge = resp.max_sq_sge; + ctx->max_rq_sge = resp.max_rq_sge; + ctx->max_rdma_size = resp.max_rdma_size; + ctx->max_wr_rdma_sge = attr.orig_attr.max_sge_rd; + } else { + err = ibv_cmd_query_device_any(&ctx->ibvctx.context, NULL, + &attr, sizeof(attr.orig_attr), + NULL, NULL); + if (err) + return err; + } + + qp_table_sz = roundup_pow_of_two(attr.orig_attr.max_qp); + ctx->qp_table_sz_m1 = qp_table_sz - 1; + ctx->qp_table = calloc(qp_table_sz, sizeof(*ctx->qp_table)); + if (!ctx->qp_table) + return ENOMEM; + return 0; +} + int efadv_query_device(struct ibv_context *ibvctx, struct efadv_device_attr *attr, uint32_t inlen) diff --git a/providers/efa/verbs.h b/providers/efa/verbs.h index da022e615af064..3b0e4e0d498761 100644 --- a/providers/efa/verbs.h +++ b/providers/efa/verbs.h @@ -9,6 +9,7 @@ #include <infiniband/driver.h> #include <infiniband/verbs.h> +int efa_query_device_ctx(struct efa_context *ctx); int efa_query_device(struct ibv_context *uctx, struct ibv_device_attr *attr); int efa_query_port(struct ibv_context *uctx, uint8_t port, struct ibv_port_attr *attr);
When the user calls efa_query_device_ex() it should not cause the context values to be mutated, only the attribute shuld be returned. Move this code to a dedicated function that is only called during context setup. Cc: Gal Pressman <galpress@amazon.com> Signed-off-by: Jason Gunthorpe <jgg@nvidia.com> --- providers/efa/efa.c | 14 +------------ providers/efa/verbs.c | 46 +++++++++++++++++++++++++++++++++++-------- providers/efa/verbs.h | 1 + 3 files changed, 40 insertions(+), 21 deletions(-)