RDMA/cxgb3: Fixes for zero STag
Handling the zero STag in receive work request requires some extra
logic in the driver:
- Only set the QP_PRIV bit for kernel mode QPs.
- Add a zero STag build function for recv wrs. The uP needs a PBL
allocated and passed down in the recv WR so it can construct a HW
PBL for the zero STag S/G entries. Note: we need to place a few
restrictions on zero STag usage because of this:
1) all SGEs in a recv WR must either be zero STag or not. No mixing.
2) an individual SGE length cannot exceed 128MB for a zero-stag SGE.
This should be OK since it's not really practical to allocate
such a large chunk of pinned contiguous DMA mapped memory.
- Add an optimized non-zero-STag recv wr format for kernel users.
This is needed to optimize both zero and non-zero STag cracking in
the recv path for kernel users.
- Remove the iwch_ prefix from the static build functions.
- Bump required FW version.
Signed-off-by: Steve Wise <swise@opengridcomputing.com>
diff --git a/drivers/infiniband/hw/cxgb3/iwch_qp.c b/drivers/infiniband/hw/cxgb3/iwch_qp.c
index 3b44300..9a3be3a 100644
--- a/drivers/infiniband/hw/cxgb3/iwch_qp.c
+++ b/drivers/infiniband/hw/cxgb3/iwch_qp.c
@@ -33,10 +33,11 @@
#include "iwch.h"
#include "iwch_cm.h"
#include "cxio_hal.h"
+#include "cxio_resource.h"
#define NO_SUPPORT -1
-static int iwch_build_rdma_send(union t3_wr *wqe, struct ib_send_wr *wr,
+static int build_rdma_send(union t3_wr *wqe, struct ib_send_wr *wr,
u8 * flit_cnt)
{
int i;
@@ -81,7 +82,7 @@
return 0;
}
-static int iwch_build_rdma_write(union t3_wr *wqe, struct ib_send_wr *wr,
+static int build_rdma_write(union t3_wr *wqe, struct ib_send_wr *wr,
u8 *flit_cnt)
{
int i;
@@ -122,7 +123,7 @@
return 0;
}
-static int iwch_build_rdma_read(union t3_wr *wqe, struct ib_send_wr *wr,
+static int build_rdma_read(union t3_wr *wqe, struct ib_send_wr *wr,
u8 *flit_cnt)
{
if (wr->num_sge > 1)
@@ -143,7 +144,7 @@
return 0;
}
-static int iwch_build_fastreg(union t3_wr *wqe, struct ib_send_wr *wr,
+static int build_fastreg(union t3_wr *wqe, struct ib_send_wr *wr,
u8 *flit_cnt, int *wr_cnt, struct t3_wq *wq)
{
int i;
@@ -185,7 +186,7 @@
return 0;
}
-static int iwch_build_inv_stag(union t3_wr *wqe, struct ib_send_wr *wr,
+static int build_inv_stag(union t3_wr *wqe, struct ib_send_wr *wr,
u8 *flit_cnt)
{
wqe->local_inv.stag = cpu_to_be32(wr->ex.invalidate_rkey);
@@ -244,23 +245,106 @@
return 0;
}
-static int iwch_build_rdma_recv(struct iwch_dev *rhp, union t3_wr *wqe,
+static int build_rdma_recv(struct iwch_qp *qhp, union t3_wr *wqe,
struct ib_recv_wr *wr)
{
- int i;
- if (wr->num_sge > T3_MAX_SGE)
- return -EINVAL;
+ int i, err = 0;
+ u32 pbl_addr[T3_MAX_SGE];
+ u8 page_size[T3_MAX_SGE];
+
+ err = iwch_sgl2pbl_map(qhp->rhp, wr->sg_list, wr->num_sge, pbl_addr,
+ page_size);
+ if (err)
+ return err;
+ wqe->recv.pagesz[0] = page_size[0];
+ wqe->recv.pagesz[1] = page_size[1];
+ wqe->recv.pagesz[2] = page_size[2];
+ wqe->recv.pagesz[3] = page_size[3];
wqe->recv.num_sgle = cpu_to_be32(wr->num_sge);
for (i = 0; i < wr->num_sge; i++) {
wqe->recv.sgl[i].stag = cpu_to_be32(wr->sg_list[i].lkey);
wqe->recv.sgl[i].len = cpu_to_be32(wr->sg_list[i].length);
- wqe->recv.sgl[i].to = cpu_to_be64(wr->sg_list[i].addr);
+
+ /* to in the WQE == the offset into the page */
+ wqe->recv.sgl[i].to = cpu_to_be64(((u32) wr->sg_list[i].addr) %
+ (1UL << (12 + page_size[i])));
+
+ /* pbl_addr is the adapters address in the PBL */
+ wqe->recv.pbl_addr[i] = cpu_to_be32(pbl_addr[i]);
}
for (; i < T3_MAX_SGE; i++) {
wqe->recv.sgl[i].stag = 0;
wqe->recv.sgl[i].len = 0;
wqe->recv.sgl[i].to = 0;
+ wqe->recv.pbl_addr[i] = 0;
}
+ qhp->wq.rq[Q_PTR2IDX(qhp->wq.rq_wptr,
+ qhp->wq.rq_size_log2)].wr_id = wr->wr_id;
+ qhp->wq.rq[Q_PTR2IDX(qhp->wq.rq_wptr,
+ qhp->wq.rq_size_log2)].pbl_addr = 0;
+ return 0;
+}
+
+static int build_zero_stag_recv(struct iwch_qp *qhp, union t3_wr *wqe,
+ struct ib_recv_wr *wr)
+{
+ int i;
+ u32 pbl_addr;
+ u32 pbl_offset;
+
+
+ /*
+ * The T3 HW requires the PBL in the HW recv descriptor to reference
+ * a PBL entry. So we allocate the max needed PBL memory here and pass
+ * it to the uP in the recv WR. The uP will build the PBL and setup
+ * the HW recv descriptor.
+ */
+ pbl_addr = cxio_hal_pblpool_alloc(&qhp->rhp->rdev, T3_STAG0_PBL_SIZE);
+ if (!pbl_addr)
+ return -ENOMEM;
+
+ /*
+ * Compute the 8B aligned offset.
+ */
+ pbl_offset = (pbl_addr - qhp->rhp->rdev.rnic_info.pbl_base) >> 3;
+
+ wqe->recv.num_sgle = cpu_to_be32(wr->num_sge);
+
+ for (i = 0; i < wr->num_sge; i++) {
+
+ /*
+ * Use a 128MB page size. This and an imposed 128MB
+ * sge length limit allows us to require only a 2-entry HW
+ * PBL for each SGE. This restriction is acceptable since
+ * since it is not possible to allocate 128MB of contiguous
+ * DMA coherent memory!
+ */
+ if (wr->sg_list[i].length > T3_STAG0_MAX_PBE_LEN)
+ return -EINVAL;
+ wqe->recv.pagesz[i] = T3_STAG0_PAGE_SHIFT;
+
+ /*
+ * T3 restricts a recv to all zero-stag or all non-zero-stag.
+ */
+ if (wr->sg_list[i].lkey != 0)
+ return -EINVAL;
+ wqe->recv.sgl[i].stag = 0;
+ wqe->recv.sgl[i].len = cpu_to_be32(wr->sg_list[i].length);
+ wqe->recv.sgl[i].to = cpu_to_be64(wr->sg_list[i].addr);
+ wqe->recv.pbl_addr[i] = cpu_to_be32(pbl_offset);
+ pbl_offset += 2;
+ }
+ for (; i < T3_MAX_SGE; i++) {
+ wqe->recv.pagesz[i] = 0;
+ wqe->recv.sgl[i].stag = 0;
+ wqe->recv.sgl[i].len = 0;
+ wqe->recv.sgl[i].to = 0;
+ wqe->recv.pbl_addr[i] = 0;
+ }
+ qhp->wq.rq[Q_PTR2IDX(qhp->wq.rq_wptr,
+ qhp->wq.rq_size_log2)].wr_id = wr->wr_id;
+ qhp->wq.rq[Q_PTR2IDX(qhp->wq.rq_wptr,
+ qhp->wq.rq_size_log2)].pbl_addr = pbl_addr;
return 0;
}
@@ -312,18 +396,18 @@
if (wr->send_flags & IB_SEND_FENCE)
t3_wr_flags |= T3_READ_FENCE_FLAG;
t3_wr_opcode = T3_WR_SEND;
- err = iwch_build_rdma_send(wqe, wr, &t3_wr_flit_cnt);
+ err = build_rdma_send(wqe, wr, &t3_wr_flit_cnt);
break;
case IB_WR_RDMA_WRITE:
case IB_WR_RDMA_WRITE_WITH_IMM:
t3_wr_opcode = T3_WR_WRITE;
- err = iwch_build_rdma_write(wqe, wr, &t3_wr_flit_cnt);
+ err = build_rdma_write(wqe, wr, &t3_wr_flit_cnt);
break;
case IB_WR_RDMA_READ:
case IB_WR_RDMA_READ_WITH_INV:
t3_wr_opcode = T3_WR_READ;
t3_wr_flags = 0; /* T3 reads are always signaled */
- err = iwch_build_rdma_read(wqe, wr, &t3_wr_flit_cnt);
+ err = build_rdma_read(wqe, wr, &t3_wr_flit_cnt);
if (err)
break;
sqp->read_len = wqe->read.local_len;
@@ -332,14 +416,14 @@
break;
case IB_WR_FAST_REG_MR:
t3_wr_opcode = T3_WR_FASTREG;
- err = iwch_build_fastreg(wqe, wr, &t3_wr_flit_cnt,
+ err = build_fastreg(wqe, wr, &t3_wr_flit_cnt,
&wr_cnt, &qhp->wq);
break;
case IB_WR_LOCAL_INV:
if (wr->send_flags & IB_SEND_FENCE)
t3_wr_flags |= T3_LOCAL_FENCE_FLAG;
t3_wr_opcode = T3_WR_INV_STAG;
- err = iwch_build_inv_stag(wqe, wr, &t3_wr_flit_cnt);
+ err = build_inv_stag(wqe, wr, &t3_wr_flit_cnt);
break;
default:
PDBG("%s post of type=%d TBD!\n", __func__,
@@ -398,18 +482,24 @@
return -EINVAL;
}
while (wr) {
+ if (wr->num_sge > T3_MAX_SGE) {
+ err = -EINVAL;
+ *bad_wr = wr;
+ break;
+ }
idx = Q_PTR2IDX(qhp->wq.wptr, qhp->wq.size_log2);
wqe = (union t3_wr *) (qhp->wq.queue + idx);
if (num_wrs)
- err = iwch_build_rdma_recv(qhp->rhp, wqe, wr);
+ if (wr->sg_list[0].lkey)
+ err = build_rdma_recv(qhp, wqe, wr);
+ else
+ err = build_zero_stag_recv(qhp, wqe, wr);
else
err = -ENOMEM;
if (err) {
*bad_wr = wr;
break;
}
- qhp->wq.rq[Q_PTR2IDX(qhp->wq.rq_wptr, qhp->wq.rq_size_log2)] =
- wr->wr_id;
build_fw_riwrh((void *) wqe, T3_WR_RCV, T3_COMPLETION_FLAG,
Q_GENBIT(qhp->wq.wptr, qhp->wq.size_log2),
0, sizeof(struct t3_receive_wr) >> 3, T3_SOPEOP);
@@ -810,7 +900,8 @@
init_attr.qp_dma_size = (1UL << qhp->wq.size_log2);
init_attr.rqe_count = iwch_rqes_posted(qhp);
init_attr.flags = qhp->attr.mpa_attr.initiator ? MPA_INITIATOR : 0;
- init_attr.flags |= capable(CAP_NET_BIND_SERVICE) ? PRIV_QP : 0;
+ if (!qhp->ibqp.uobject)
+ init_attr.flags |= PRIV_QP;
if (peer2peer) {
init_attr.rtr_type = RTR_READ;
if (init_attr.ord == 0 && qhp->attr.mpa_attr.initiator)