@@ -667,7 +667,7 @@ static inline struct siw_sqe *orq_get_free(struct siw_qp *qp)
{
struct siw_sqe *orq_e = orq_get_tail(qp);
- if (orq_e && READ_ONCE(orq_e->flags) == 0)
+ if (READ_ONCE(orq_e->flags) == 0)
return orq_e;
return NULL;
@@ -199,26 +199,26 @@ void siw_qp_llp_write_space(struct sock *sk)
static int siw_qp_readq_init(struct siw_qp *qp, int irq_size, int orq_size)
{
- irq_size = roundup_pow_of_two(irq_size);
- orq_size = roundup_pow_of_two(orq_size);
-
- qp->attrs.irq_size = irq_size;
- qp->attrs.orq_size = orq_size;
-
- qp->irq = vzalloc(irq_size * sizeof(struct siw_sqe));
- if (!qp->irq) {
- siw_dbg_qp(qp, "irq malloc for %d failed\n", irq_size);
- qp->attrs.irq_size = 0;
- return -ENOMEM;
+ if (irq_size) {
+ irq_size = roundup_pow_of_two(irq_size);
+ qp->irq = vzalloc(irq_size * sizeof(struct siw_sqe));
+ if (!qp->irq) {
+ qp->attrs.irq_size = 0;
+ return -ENOMEM;
+ }
}
- qp->orq = vzalloc(orq_size * sizeof(struct siw_sqe));
- if (!qp->orq) {
- siw_dbg_qp(qp, "orq malloc for %d failed\n", orq_size);
- qp->attrs.orq_size = 0;
- qp->attrs.irq_size = 0;
- vfree(qp->irq);
- return -ENOMEM;
+ if (orq_size) {
+ orq_size = roundup_pow_of_two(orq_size);
+ qp->orq = vzalloc(orq_size * sizeof(struct siw_sqe));
+ if (!qp->orq) {
+ qp->attrs.orq_size = 0;
+ qp->attrs.irq_size = 0;
+ vfree(qp->irq);
+ return -ENOMEM;
+ }
}
+ qp->attrs.irq_size = irq_size;
+ qp->attrs.orq_size = orq_size;
siw_dbg_qp(qp, "ORD %d, IRD %d\n", orq_size, irq_size);
return 0;
}
@@ -288,13 +288,14 @@ int siw_qp_mpa_rts(struct siw_qp *qp, enum mpa_v2_ctrl ctrl)
if (ctrl & MPA_V2_RDMA_WRITE_RTR)
wqe->sqe.opcode = SIW_OP_WRITE;
else if (ctrl & MPA_V2_RDMA_READ_RTR) {
- struct siw_sqe *rreq;
+ struct siw_sqe *rreq = NULL;
wqe->sqe.opcode = SIW_OP_READ;
spin_lock(&qp->orq_lock);
- rreq = orq_get_free(qp);
+ if (qp->attrs.orq_size)
+ rreq = orq_get_free(qp);
if (rreq) {
siw_read_to_orq(rreq, &wqe->sqe);
qp->orq_put++;
@@ -877,135 +878,88 @@ void siw_read_to_orq(struct siw_sqe *rreq, struct siw_sqe *sqe)
rreq->num_sge = 1;
}
-/*
- * Must be called with SQ locked.
- * To avoid complete SQ starvation by constant inbound READ requests,
- * the active IRQ will not be served after qp->irq_burst, if the
- * SQ has pending work.
- */
-int siw_activate_tx(struct siw_qp *qp)
+static int siw_activate_tx_from_sq(struct siw_qp *qp)
{
- struct siw_sqe *irqe, *sqe;
+ struct siw_sqe *sqe;
struct siw_wqe *wqe = tx_wqe(qp);
int rv = 1;
- irqe = &qp->irq[qp->irq_get % qp->attrs.irq_size];
-
- if (irqe->flags & SIW_WQE_VALID) {
- sqe = sq_get_next(qp);
-
- /*
- * Avoid local WQE processing starvation in case
- * of constant inbound READ request stream
- */
- if (sqe && ++qp->irq_burst >= SIW_IRQ_MAXBURST_SQ_ACTIVE) {
- qp->irq_burst = 0;
- goto skip_irq;
- }
- memset(wqe->mem, 0, sizeof(*wqe->mem) * SIW_MAX_SGE);
- wqe->wr_status = SIW_WR_QUEUED;
-
- /* start READ RESPONSE */
- wqe->sqe.opcode = SIW_OP_READ_RESPONSE;
- wqe->sqe.flags = 0;
- if (irqe->num_sge) {
- wqe->sqe.num_sge = 1;
- wqe->sqe.sge[0].length = irqe->sge[0].length;
- wqe->sqe.sge[0].laddr = irqe->sge[0].laddr;
- wqe->sqe.sge[0].lkey = irqe->sge[0].lkey;
- } else {
- wqe->sqe.num_sge = 0;
- }
-
- /* Retain original RREQ's message sequence number for
- * potential error reporting cases.
- */
- wqe->sqe.sge[1].length = irqe->sge[1].length;
-
- wqe->sqe.rkey = irqe->rkey;
- wqe->sqe.raddr = irqe->raddr;
+ sqe = sq_get_next(qp);
+ if (!sqe)
+ return 0;
- wqe->processed = 0;
- qp->irq_get++;
+ memset(wqe->mem, 0, sizeof(*wqe->mem) * SIW_MAX_SGE);
+ wqe->wr_status = SIW_WR_QUEUED;
- /* mark current IRQ entry free */
- smp_store_mb(irqe->flags, 0);
+ /* First copy SQE to kernel private memory */
+ memcpy(&wqe->sqe, sqe, sizeof(*sqe));
+ if (wqe->sqe.opcode >= SIW_NUM_OPCODES) {
+ rv = -EINVAL;
goto out;
}
- sqe = sq_get_next(qp);
- if (sqe) {
-skip_irq:
- memset(wqe->mem, 0, sizeof(*wqe->mem) * SIW_MAX_SGE);
- wqe->wr_status = SIW_WR_QUEUED;
-
- /* First copy SQE to kernel private memory */
- memcpy(&wqe->sqe, sqe, sizeof(*sqe));
-
- if (wqe->sqe.opcode >= SIW_NUM_OPCODES) {
+ if (wqe->sqe.flags & SIW_WQE_INLINE) {
+ if (wqe->sqe.opcode != SIW_OP_SEND &&
+ wqe->sqe.opcode != SIW_OP_WRITE) {
rv = -EINVAL;
goto out;
}
- if (wqe->sqe.flags & SIW_WQE_INLINE) {
- if (wqe->sqe.opcode != SIW_OP_SEND &&
- wqe->sqe.opcode != SIW_OP_WRITE) {
- rv = -EINVAL;
- goto out;
- }
- if (wqe->sqe.sge[0].length > SIW_MAX_INLINE) {
- rv = -EINVAL;
- goto out;
- }
- wqe->sqe.sge[0].laddr = (uintptr_t)&wqe->sqe.sge[1];
- wqe->sqe.sge[0].lkey = 0;
- wqe->sqe.num_sge = 1;
+ if (wqe->sqe.sge[0].length > SIW_MAX_INLINE) {
+ rv = -EINVAL;
+ goto out;
}
- if (wqe->sqe.flags & SIW_WQE_READ_FENCE) {
- /* A READ cannot be fenced */
- if (unlikely(wqe->sqe.opcode == SIW_OP_READ ||
- wqe->sqe.opcode ==
- SIW_OP_READ_LOCAL_INV)) {
- siw_dbg_qp(qp, "cannot fence read\n");
- rv = -EINVAL;
- goto out;
- }
- spin_lock(&qp->orq_lock);
+ wqe->sqe.sge[0].laddr = (uintptr_t)&wqe->sqe.sge[1];
+ wqe->sqe.sge[0].lkey = 0;
+ wqe->sqe.num_sge = 1;
+ }
+ if (wqe->sqe.flags & SIW_WQE_READ_FENCE) {
+ /* A READ cannot be fenced */
+ if (unlikely(wqe->sqe.opcode == SIW_OP_READ ||
+ wqe->sqe.opcode ==
+ SIW_OP_READ_LOCAL_INV)) {
+ siw_dbg_qp(qp, "cannot fence read\n");
+ rv = -EINVAL;
+ goto out;
+ }
+ spin_lock(&qp->orq_lock);
- if (!siw_orq_empty(qp)) {
- qp->tx_ctx.orq_fence = 1;
- rv = 0;
- }
- spin_unlock(&qp->orq_lock);
+ if (qp->attrs.orq_size && !siw_orq_empty(qp)) {
+ qp->tx_ctx.orq_fence = 1;
+ rv = 0;
+ }
+ spin_unlock(&qp->orq_lock);
- } else if (wqe->sqe.opcode == SIW_OP_READ ||
- wqe->sqe.opcode == SIW_OP_READ_LOCAL_INV) {
- struct siw_sqe *rreq;
+ } else if (wqe->sqe.opcode == SIW_OP_READ ||
+ wqe->sqe.opcode == SIW_OP_READ_LOCAL_INV) {
+ struct siw_sqe *rreq;
- wqe->sqe.num_sge = 1;
+ if (unlikely(!qp->attrs.orq_size)) {
+ /* We negotiated not to send READ req's */
+ rv = -EINVAL;
+ goto out;
+ }
+ wqe->sqe.num_sge = 1;
- spin_lock(&qp->orq_lock);
+ spin_lock(&qp->orq_lock);
- rreq = orq_get_free(qp);
- if (rreq) {
- /*
- * Make an immediate copy in ORQ to be ready
- * to process loopback READ reply
- */
- siw_read_to_orq(rreq, &wqe->sqe);
- qp->orq_put++;
- } else {
- qp->tx_ctx.orq_fence = 1;
- rv = 0;
- }
- spin_unlock(&qp->orq_lock);
+ rreq = orq_get_free(qp);
+ if (rreq) {
+ /*
+ * Make an immediate copy in ORQ to be ready
+ * to process loopback READ reply
+ */
+ siw_read_to_orq(rreq, &wqe->sqe);
+ qp->orq_put++;
+ } else {
+ qp->tx_ctx.orq_fence = 1;
+ rv = 0;
}
-
- /* Clear SQE, can be re-used by application */
- smp_store_mb(sqe->flags, 0);
- qp->sq_get++;
- } else {
- rv = 0;
+ spin_unlock(&qp->orq_lock);
}
+
+ /* Clear SQE, can be re-used by application */
+ smp_store_mb(sqe->flags, 0);
+ qp->sq_get++;
out:
if (unlikely(rv < 0)) {
siw_dbg_qp(qp, "error %d\n", rv);
@@ -1014,6 +968,65 @@ out:
return rv;
}
+/*
+ * Must be called with SQ locked.
+ * To avoid complete SQ starvation by constant inbound READ requests,
+ * the active IRQ will not be served after qp->irq_burst, if the
+ * SQ has pending work.
+ */
+int siw_activate_tx(struct siw_qp *qp)
+{
+ struct siw_sqe *irqe;
+ struct siw_wqe *wqe = tx_wqe(qp);
+
+ if (!qp->attrs.irq_size)
+ return siw_activate_tx_from_sq(qp);
+
+ irqe = &qp->irq[qp->irq_get % qp->attrs.irq_size];
+
+ if (!(irqe->flags & SIW_WQE_VALID))
+ return siw_activate_tx_from_sq(qp);
+
+ /*
+ * Avoid local WQE processing starvation in case
+ * of constant inbound READ request stream
+ */
+ if (sq_get_next(qp) && ++qp->irq_burst >= SIW_IRQ_MAXBURST_SQ_ACTIVE) {
+ qp->irq_burst = 0;
+ return siw_activate_tx_from_sq(qp);
+ }
+ memset(wqe->mem, 0, sizeof(*wqe->mem) * SIW_MAX_SGE);
+ wqe->wr_status = SIW_WR_QUEUED;
+
+ /* start READ RESPONSE */
+ wqe->sqe.opcode = SIW_OP_READ_RESPONSE;
+ wqe->sqe.flags = 0;
+ if (irqe->num_sge) {
+ wqe->sqe.num_sge = 1;
+ wqe->sqe.sge[0].length = irqe->sge[0].length;
+ wqe->sqe.sge[0].laddr = irqe->sge[0].laddr;
+ wqe->sqe.sge[0].lkey = irqe->sge[0].lkey;
+ } else {
+ wqe->sqe.num_sge = 0;
+ }
+
+ /* Retain original RREQ's message sequence number for
+ * potential error reporting cases.
+ */
+ wqe->sqe.sge[1].length = irqe->sge[1].length;
+
+ wqe->sqe.rkey = irqe->rkey;
+ wqe->sqe.raddr = irqe->raddr;
+
+ wqe->processed = 0;
+ qp->irq_get++;
+
+ /* mark current IRQ entry free */
+ smp_store_mb(irqe->flags, 0);
+
+ return 1;
+}
+
/*
* Check if current CQ state qualifies for calling CQ completion
* handler. Must be called with CQ lock held.
@@ -680,6 +680,10 @@ static int siw_init_rresp(struct siw_qp *qp, struct siw_rx_stream *srx)
}
spin_lock_irqsave(&qp->sq_lock, flags);
+ if (unlikely(!qp->attrs.irq_size)) {
+ run_sq = 0;
+ goto error_irq;
+ }
if (tx_work->wr_status == SIW_WR_IDLE) {
/*
* immediately schedule READ response w/o
@@ -712,8 +716,9 @@ static int siw_init_rresp(struct siw_qp *qp, struct siw_rx_stream *srx)
/* RRESP now valid as current TX wqe or placed into IRQ */
smp_store_mb(resp->flags, SIW_WQE_VALID);
} else {
- pr_warn("siw: [QP %u]: irq %d exceeded %d\n", qp_id(qp),
- qp->irq_put % qp->attrs.irq_size, qp->attrs.irq_size);
+error_irq:
+ pr_warn("siw: [QP %u]: IRQ exceeded or null, size %d\n",
+ qp_id(qp), qp->attrs.irq_size);
siw_init_terminate(qp, TERM_ERROR_LAYER_RDMAP,
RDMAP_ETYPE_REMOTE_OPERATION,
@@ -740,6 +745,9 @@ static int siw_orqe_start_rx(struct siw_qp *qp)
struct siw_sqe *orqe;
struct siw_wqe *wqe = NULL;
+ if (unlikely(!qp->attrs.orq_size))
+ return -EPROTO;
+
/* make sure ORQ indices are current */
smp_mb();
@@ -796,8 +804,8 @@ int siw_proc_rresp(struct siw_qp *qp)
*/
rv = siw_orqe_start_rx(qp);
if (rv) {
- pr_warn("siw: [QP %u]: ORQ empty at idx %d\n",
- qp_id(qp), qp->orq_get % qp->attrs.orq_size);
+ pr_warn("siw: [QP %u]: ORQ empty, size %d\n",
+ qp_id(qp), qp->attrs.orq_size);
goto error_term;
}
rv = siw_rresp_check_ntoh(srx, frx);
@@ -1290,11 +1298,13 @@ static int siw_rdmap_complete(struct siw_qp *qp, int error)
wc_status);
siw_wqe_put_mem(wqe, SIW_OP_READ);
- if (!error)
+ if (!error) {
rv = siw_check_tx_fence(qp);
- else
- /* Disable current ORQ eleement */
- WRITE_ONCE(orq_get_current(qp)->flags, 0);
+ } else {
+ /* Disable current ORQ element */
+ if (qp->attrs.orq_size)
+ WRITE_ONCE(orq_get_current(qp)->flags, 0);
+ }
break;
case RDMAP_RDMA_READ_REQ:
@@ -1107,8 +1107,8 @@ next_wqe:
/*
* RREQ may have already been completed by inbound RRESP!
*/
- if (tx_type == SIW_OP_READ ||
- tx_type == SIW_OP_READ_LOCAL_INV) {
+ if ((tx_type == SIW_OP_READ ||
+ tx_type == SIW_OP_READ_LOCAL_INV) && qp->attrs.orq_size) {
/* Cleanup pending entry in ORQ */
qp->orq_put--;
qp->orq[qp->orq_put % qp->attrs.orq_size].flags = 0;
@@ -387,13 +387,23 @@ struct ib_qp *siw_create_qp(struct ib_pd *pd,
if (rv)
goto err_out;
+ num_sqe = attrs->cap.max_send_wr;
+ num_rqe = attrs->cap.max_recv_wr;
+
/* All queue indices are derived from modulo operations
* on a free running 'get' (consumer) and 'put' (producer)
* unsigned counter. Having queue sizes at power of two
* avoids handling counter wrap around.
*/
- num_sqe = roundup_pow_of_two(attrs->cap.max_send_wr);
- num_rqe = roundup_pow_of_two(attrs->cap.max_recv_wr);
+ if (num_sqe)
+ num_sqe = roundup_pow_of_two(num_sqe);
+ else {
+ /* Zero sized SQ is not supported */
+ rv = -EINVAL;
+ goto err_out;
+ }
+ if (num_rqe)
+ num_rqe = roundup_pow_of_two(num_rqe);
if (qp->kernel_verbs)
qp->sendq = vzalloc(num_sqe * sizeof(struct siw_sqe));
@@ -401,7 +411,6 @@ struct ib_qp *siw_create_qp(struct ib_pd *pd,
qp->sendq = vmalloc_user(num_sqe * sizeof(struct siw_sqe));
if (qp->sendq == NULL) {
- siw_dbg(base_dev, "SQ size %d alloc failed\n", num_sqe);
rv = -ENOMEM;
goto err_out_xa;
}
@@ -434,7 +443,6 @@ struct ib_qp *siw_create_qp(struct ib_pd *pd,
vmalloc_user(num_rqe * sizeof(struct siw_rqe));
if (qp->recvq == NULL) {
- siw_dbg(base_dev, "RQ size %d alloc failed\n", num_rqe);
rv = -ENOMEM;
goto err_out_xa;
}
@@ -982,9 +990,9 @@ int siw_post_receive(struct ib_qp *base_qp, const struct ib_recv_wr *wr,
unsigned long flags;
int rv = 0;
- if (qp->srq) {
+ if (qp->srq || qp->attrs.rq_size == 0) {
*bad_wr = wr;
- return -EOPNOTSUPP; /* what else from errno.h? */
+ return -EINVAL;
}
if (!qp->kernel_verbs) {
siw_dbg_qp(qp, "no kernel post_recv for user mapped sq\n");