IB/mlx5: Reset flow support for IB kernel ULPs
The driver exposes interfaces that directly relate to HW state.
Upon fatal error, consumers of these interfaces (ULPs) that rely
on completion of all their posted work-request could hang, thereby
introducing dependencies in shutdown order. To prevent this from
happening, we manage the relevant resources (CQs, QPs) that are used
by the device. Upon a fatal error, we now generate simulated
completions for outstanding WQEs that were not completed at the
time the HW was reset.
It includes invoking the completion event handler for all involved
CQs so that the ULPs will poll those CQs. When polled we return
simulated CQEs with IB_WC_WR_FLUSH_ERR return code enabling ULPs
to clean up their resources and not wait forever for completions
upon receiving remove_one.
The above change requires an extra check in the data path to make
sure that when device is in error state, the simulated CQEs will
be returned and no further WQEs will be posted.
Signed-off-by: Maor Gottlieb <maorg@mellanox.com>
Signed-off-by: Leon Romanovsky <leon@kernel.org>
Signed-off-by: Doug Ledford <dledford@redhat.com>
diff --git a/drivers/infiniband/hw/mlx5/cq.c b/drivers/infiniband/hw/mlx5/cq.c
index 9c0e67b..308a358 100644
--- a/drivers/infiniband/hw/mlx5/cq.c
+++ b/drivers/infiniband/hw/mlx5/cq.c
@@ -424,6 +424,83 @@
item->key = be32_to_cpu(cqe->mkey);
}
+static void sw_send_comp(struct mlx5_ib_qp *qp, int num_entries,
+ struct ib_wc *wc, int *npolled)
+{
+ struct mlx5_ib_wq *wq;
+ unsigned int cur;
+ unsigned int idx;
+ int np;
+ int i;
+
+ wq = &qp->sq;
+ cur = wq->head - wq->tail;
+ np = *npolled;
+
+ if (cur == 0)
+ return;
+
+ for (i = 0; i < cur && np < num_entries; i++) {
+ idx = wq->last_poll & (wq->wqe_cnt - 1);
+ wc->wr_id = wq->wrid[idx];
+ wc->status = IB_WC_WR_FLUSH_ERR;
+ wc->vendor_err = MLX5_CQE_SYNDROME_WR_FLUSH_ERR;
+ wq->tail++;
+ np++;
+ wc->qp = &qp->ibqp;
+ wc++;
+ wq->last_poll = wq->w_list[idx].next;
+ }
+ *npolled = np;
+}
+
+static void sw_recv_comp(struct mlx5_ib_qp *qp, int num_entries,
+ struct ib_wc *wc, int *npolled)
+{
+ struct mlx5_ib_wq *wq;
+ unsigned int cur;
+ int np;
+ int i;
+
+ wq = &qp->rq;
+ cur = wq->head - wq->tail;
+ np = *npolled;
+
+ if (cur == 0)
+ return;
+
+ for (i = 0; i < cur && np < num_entries; i++) {
+ wc->wr_id = wq->wrid[wq->tail & (wq->wqe_cnt - 1)];
+ wc->status = IB_WC_WR_FLUSH_ERR;
+ wc->vendor_err = MLX5_CQE_SYNDROME_WR_FLUSH_ERR;
+ wq->tail++;
+ np++;
+ wc->qp = &qp->ibqp;
+ wc++;
+ }
+ *npolled = np;
+}
+
+static void mlx5_ib_poll_sw_comp(struct mlx5_ib_cq *cq, int num_entries,
+ struct ib_wc *wc, int *npolled)
+{
+ struct mlx5_ib_qp *qp;
+
+ *npolled = 0;
+ /* Find uncompleted WQEs belonging to that cq and retrun mmics ones */
+ list_for_each_entry(qp, &cq->list_send_qp, cq_send_list) {
+ sw_send_comp(qp, num_entries, wc + *npolled, npolled);
+ if (*npolled >= num_entries)
+ return;
+ }
+
+ list_for_each_entry(qp, &cq->list_recv_qp, cq_recv_list) {
+ sw_recv_comp(qp, num_entries, wc + *npolled, npolled);
+ if (*npolled >= num_entries)
+ return;
+ }
+}
+
static int mlx5_poll_one(struct mlx5_ib_cq *cq,
struct mlx5_ib_qp **cur_qp,
struct ib_wc *wc)
@@ -594,12 +671,18 @@
{
struct mlx5_ib_cq *cq = to_mcq(ibcq);
struct mlx5_ib_qp *cur_qp = NULL;
+ struct mlx5_ib_dev *dev = to_mdev(cq->ibcq.device);
+ struct mlx5_core_dev *mdev = dev->mdev;
unsigned long flags;
int soft_polled = 0;
int npolled;
int err = 0;
spin_lock_irqsave(&cq->lock, flags);
+ if (mdev->state == MLX5_DEVICE_STATE_INTERNAL_ERROR) {
+ mlx5_ib_poll_sw_comp(cq, num_entries, wc, &npolled);
+ goto out;
+ }
if (unlikely(!list_empty(&cq->wc_list)))
soft_polled = poll_soft_wc(cq, num_entries, wc);
@@ -612,7 +695,7 @@
if (npolled)
mlx5_cq_set_ci(&cq->mcq);
-
+out:
spin_unlock_irqrestore(&cq->lock, flags);
if (err == 0 || err == -EAGAIN)
@@ -843,6 +926,8 @@
cq->resize_buf = NULL;
cq->resize_umem = NULL;
cq->create_flags = attr->flags;
+ INIT_LIST_HEAD(&cq->list_send_qp);
+ INIT_LIST_HEAD(&cq->list_recv_qp);
if (context) {
err = create_cq_user(dev, udata, context, cq, entries,
diff --git a/drivers/infiniband/hw/mlx5/main.c b/drivers/infiniband/hw/mlx5/main.c
index 21acee4..9b6d283 100644
--- a/drivers/infiniband/hw/mlx5/main.c
+++ b/drivers/infiniband/hw/mlx5/main.c
@@ -1980,6 +1980,65 @@
mutex_unlock(&ports->devr->mutex);
}
+static void mlx5_ib_handle_internal_error(struct mlx5_ib_dev *ibdev)
+{
+ struct mlx5_ib_qp *mqp;
+ struct mlx5_ib_cq *send_mcq, *recv_mcq;
+ struct mlx5_core_cq *mcq;
+ struct list_head cq_armed_list;
+ unsigned long flags_qp;
+ unsigned long flags_cq;
+ unsigned long flags;
+
+ INIT_LIST_HEAD(&cq_armed_list);
+
+ /* Go over qp list reside on that ibdev, sync with create/destroy qp.*/
+ spin_lock_irqsave(&ibdev->reset_flow_resource_lock, flags);
+ list_for_each_entry(mqp, &ibdev->qp_list, qps_list) {
+ spin_lock_irqsave(&mqp->sq.lock, flags_qp);
+ if (mqp->sq.tail != mqp->sq.head) {
+ send_mcq = to_mcq(mqp->ibqp.send_cq);
+ spin_lock_irqsave(&send_mcq->lock, flags_cq);
+ if (send_mcq->mcq.comp &&
+ mqp->ibqp.send_cq->comp_handler) {
+ if (!send_mcq->mcq.reset_notify_added) {
+ send_mcq->mcq.reset_notify_added = 1;
+ list_add_tail(&send_mcq->mcq.reset_notify,
+ &cq_armed_list);
+ }
+ }
+ spin_unlock_irqrestore(&send_mcq->lock, flags_cq);
+ }
+ spin_unlock_irqrestore(&mqp->sq.lock, flags_qp);
+ spin_lock_irqsave(&mqp->rq.lock, flags_qp);
+ /* no handling is needed for SRQ */
+ if (!mqp->ibqp.srq) {
+ if (mqp->rq.tail != mqp->rq.head) {
+ recv_mcq = to_mcq(mqp->ibqp.recv_cq);
+ spin_lock_irqsave(&recv_mcq->lock, flags_cq);
+ if (recv_mcq->mcq.comp &&
+ mqp->ibqp.recv_cq->comp_handler) {
+ if (!recv_mcq->mcq.reset_notify_added) {
+ recv_mcq->mcq.reset_notify_added = 1;
+ list_add_tail(&recv_mcq->mcq.reset_notify,
+ &cq_armed_list);
+ }
+ }
+ spin_unlock_irqrestore(&recv_mcq->lock,
+ flags_cq);
+ }
+ }
+ spin_unlock_irqrestore(&mqp->rq.lock, flags_qp);
+ }
+ /*At that point all inflight post send were put to be executed as of we
+ * lock/unlock above locks Now need to arm all involved CQs.
+ */
+ list_for_each_entry(mcq, &cq_armed_list, reset_notify) {
+ mcq->comp(mcq);
+ }
+ spin_unlock_irqrestore(&ibdev->reset_flow_resource_lock, flags);
+}
+
static void mlx5_ib_event(struct mlx5_core_dev *dev, void *context,
enum mlx5_dev_event event, unsigned long param)
{
@@ -1992,6 +2051,7 @@
case MLX5_DEV_EVENT_SYS_ERROR:
ibdev->ib_active = false;
ibev.event = IB_EVENT_DEVICE_FATAL;
+ mlx5_ib_handle_internal_error(ibdev);
break;
case MLX5_DEV_EVENT_PORT_UP:
@@ -2595,6 +2655,8 @@
mutex_init(&dev->flow_db.lock);
mutex_init(&dev->cap_mask_mutex);
+ INIT_LIST_HEAD(&dev->qp_list);
+ spin_lock_init(&dev->reset_flow_resource_lock);
if (ll == IB_LINK_LAYER_ETHERNET) {
err = mlx5_enable_roce(dev);
diff --git a/drivers/infiniband/hw/mlx5/mlx5_ib.h b/drivers/infiniband/hw/mlx5/mlx5_ib.h
index 391588e..0001ed5 100644
--- a/drivers/infiniband/hw/mlx5/mlx5_ib.h
+++ b/drivers/infiniband/hw/mlx5/mlx5_ib.h
@@ -380,6 +380,9 @@
spinlock_t disable_page_faults_lock;
struct mlx5_ib_pfault pagefaults[MLX5_IB_PAGEFAULT_CONTEXTS];
#endif
+ struct list_head qps_list;
+ struct list_head cq_recv_list;
+ struct list_head cq_send_list;
};
struct mlx5_ib_cq_buf {
@@ -441,6 +444,8 @@
struct mlx5_ib_cq_buf *resize_buf;
struct ib_umem *resize_umem;
int cqe_size;
+ struct list_head list_send_qp;
+ struct list_head list_recv_qp;
u32 create_flags;
struct list_head wc_list;
enum ib_cq_notify_flags notify_flags;
@@ -621,6 +626,9 @@
struct srcu_struct mr_srcu;
#endif
struct mlx5_ib_flow_db flow_db;
+ /* protect resources needed as part of reset flow */
+ spinlock_t reset_flow_resource_lock;
+ struct list_head qp_list;
};
static inline struct mlx5_ib_cq *to_mibcq(struct mlx5_core_cq *mcq)
diff --git a/drivers/infiniband/hw/mlx5/mr.c b/drivers/infiniband/hw/mlx5/mr.c
index 8cf2ce5..4b02130 100644
--- a/drivers/infiniband/hw/mlx5/mr.c
+++ b/drivers/infiniband/hw/mlx5/mr.c
@@ -1193,12 +1193,16 @@
static int unreg_umr(struct mlx5_ib_dev *dev, struct mlx5_ib_mr *mr)
{
+ struct mlx5_core_dev *mdev = dev->mdev;
struct umr_common *umrc = &dev->umrc;
struct mlx5_ib_umr_context umr_context;
struct mlx5_umr_wr umrwr = {};
struct ib_send_wr *bad;
int err;
+ if (mdev->state == MLX5_DEVICE_STATE_INTERNAL_ERROR)
+ return 0;
+
mlx5_ib_init_umr_context(&umr_context);
umrwr.wr.wr_cqe = &umr_context.cqe;
diff --git a/drivers/infiniband/hw/mlx5/qp.c b/drivers/infiniband/hw/mlx5/qp.c
index f9df4b52..9004905 100644
--- a/drivers/infiniband/hw/mlx5/qp.c
+++ b/drivers/infiniband/hw/mlx5/qp.c
@@ -77,6 +77,10 @@
u8 rsvd0[16];
};
+static void get_cqs(enum ib_qp_type qp_type,
+ struct ib_cq *ib_send_cq, struct ib_cq *ib_recv_cq,
+ struct mlx5_ib_cq **send_cq, struct mlx5_ib_cq **recv_cq);
+
static int is_qp0(enum ib_qp_type qp_type)
{
return qp_type == IB_QPT_SMI;
@@ -609,6 +613,11 @@
}
}
+static void mlx5_ib_lock_cqs(struct mlx5_ib_cq *send_cq,
+ struct mlx5_ib_cq *recv_cq);
+static void mlx5_ib_unlock_cqs(struct mlx5_ib_cq *send_cq,
+ struct mlx5_ib_cq *recv_cq);
+
static int uuarn_to_uar_index(struct mlx5_uuar_info *uuari, int uuarn)
{
return uuari->uars[uuarn / MLX5_BF_REGS_PER_PAGE].index;
@@ -1457,6 +1466,9 @@
struct mlx5_ib_create_qp_resp resp;
struct mlx5_create_qp_mbox_in *in;
struct mlx5_ib_create_qp ucmd;
+ struct mlx5_ib_cq *send_cq;
+ struct mlx5_ib_cq *recv_cq;
+ unsigned long flags;
int inlen = sizeof(*in);
int err;
u32 uidx = MLX5_IB_DEFAULT_UIDX;
@@ -1714,6 +1726,23 @@
base->container_mibqp = qp;
base->mqp.event = mlx5_ib_qp_event;
+ get_cqs(init_attr->qp_type, init_attr->send_cq, init_attr->recv_cq,
+ &send_cq, &recv_cq);
+ spin_lock_irqsave(&dev->reset_flow_resource_lock, flags);
+ mlx5_ib_lock_cqs(send_cq, recv_cq);
+ /* Maintain device to QPs access, needed for further handling via reset
+ * flow
+ */
+ list_add_tail(&qp->qps_list, &dev->qp_list);
+ /* Maintain CQ to QPs access, needed for further handling via reset flow
+ */
+ if (send_cq)
+ list_add_tail(&qp->cq_send_list, &send_cq->list_send_qp);
+ if (recv_cq)
+ list_add_tail(&qp->cq_recv_list, &recv_cq->list_recv_qp);
+ mlx5_ib_unlock_cqs(send_cq, recv_cq);
+ spin_unlock_irqrestore(&dev->reset_flow_resource_lock, flags);
+
return 0;
err_create:
@@ -1732,23 +1761,23 @@
if (send_cq) {
if (recv_cq) {
if (send_cq->mcq.cqn < recv_cq->mcq.cqn) {
- spin_lock_irq(&send_cq->lock);
+ spin_lock(&send_cq->lock);
spin_lock_nested(&recv_cq->lock,
SINGLE_DEPTH_NESTING);
} else if (send_cq->mcq.cqn == recv_cq->mcq.cqn) {
- spin_lock_irq(&send_cq->lock);
+ spin_lock(&send_cq->lock);
__acquire(&recv_cq->lock);
} else {
- spin_lock_irq(&recv_cq->lock);
+ spin_lock(&recv_cq->lock);
spin_lock_nested(&send_cq->lock,
SINGLE_DEPTH_NESTING);
}
} else {
- spin_lock_irq(&send_cq->lock);
+ spin_lock(&send_cq->lock);
__acquire(&recv_cq->lock);
}
} else if (recv_cq) {
- spin_lock_irq(&recv_cq->lock);
+ spin_lock(&recv_cq->lock);
__acquire(&send_cq->lock);
} else {
__acquire(&send_cq->lock);
@@ -1763,21 +1792,21 @@
if (recv_cq) {
if (send_cq->mcq.cqn < recv_cq->mcq.cqn) {
spin_unlock(&recv_cq->lock);
- spin_unlock_irq(&send_cq->lock);
+ spin_unlock(&send_cq->lock);
} else if (send_cq->mcq.cqn == recv_cq->mcq.cqn) {
__release(&recv_cq->lock);
- spin_unlock_irq(&send_cq->lock);
+ spin_unlock(&send_cq->lock);
} else {
spin_unlock(&send_cq->lock);
- spin_unlock_irq(&recv_cq->lock);
+ spin_unlock(&recv_cq->lock);
}
} else {
__release(&recv_cq->lock);
- spin_unlock_irq(&send_cq->lock);
+ spin_unlock(&send_cq->lock);
}
} else if (recv_cq) {
__release(&send_cq->lock);
- spin_unlock_irq(&recv_cq->lock);
+ spin_unlock(&recv_cq->lock);
} else {
__release(&recv_cq->lock);
__release(&send_cq->lock);
@@ -1789,17 +1818,18 @@
return to_mpd(qp->ibqp.pd);
}
-static void get_cqs(struct mlx5_ib_qp *qp,
+static void get_cqs(enum ib_qp_type qp_type,
+ struct ib_cq *ib_send_cq, struct ib_cq *ib_recv_cq,
struct mlx5_ib_cq **send_cq, struct mlx5_ib_cq **recv_cq)
{
- switch (qp->ibqp.qp_type) {
+ switch (qp_type) {
case IB_QPT_XRC_TGT:
*send_cq = NULL;
*recv_cq = NULL;
break;
case MLX5_IB_QPT_REG_UMR:
case IB_QPT_XRC_INI:
- *send_cq = to_mcq(qp->ibqp.send_cq);
+ *send_cq = ib_send_cq ? to_mcq(ib_send_cq) : NULL;
*recv_cq = NULL;
break;
@@ -1811,8 +1841,8 @@
case IB_QPT_RAW_IPV6:
case IB_QPT_RAW_ETHERTYPE:
case IB_QPT_RAW_PACKET:
- *send_cq = to_mcq(qp->ibqp.send_cq);
- *recv_cq = to_mcq(qp->ibqp.recv_cq);
+ *send_cq = ib_send_cq ? to_mcq(ib_send_cq) : NULL;
+ *recv_cq = ib_recv_cq ? to_mcq(ib_recv_cq) : NULL;
break;
case IB_QPT_MAX:
@@ -1831,6 +1861,7 @@
struct mlx5_ib_cq *send_cq, *recv_cq;
struct mlx5_ib_qp_base *base = &qp->trans_qp.base;
struct mlx5_modify_qp_mbox_in *in;
+ unsigned long flags;
int err;
if (qp->ibqp.rwq_ind_tbl) {
@@ -1861,17 +1892,28 @@
base->mqp.qpn);
}
- get_cqs(qp, &send_cq, &recv_cq);
+ get_cqs(qp->ibqp.qp_type, qp->ibqp.send_cq, qp->ibqp.recv_cq,
+ &send_cq, &recv_cq);
+
+ spin_lock_irqsave(&dev->reset_flow_resource_lock, flags);
+ mlx5_ib_lock_cqs(send_cq, recv_cq);
+ /* del from lists under both locks above to protect reset flow paths */
+ list_del(&qp->qps_list);
+ if (send_cq)
+ list_del(&qp->cq_send_list);
+
+ if (recv_cq)
+ list_del(&qp->cq_recv_list);
if (qp->create_type == MLX5_QP_KERNEL) {
- mlx5_ib_lock_cqs(send_cq, recv_cq);
__mlx5_ib_cq_clean(recv_cq, base->mqp.qpn,
qp->ibqp.srq ? to_msrq(qp->ibqp.srq) : NULL);
if (send_cq != recv_cq)
__mlx5_ib_cq_clean(send_cq, base->mqp.qpn,
NULL);
- mlx5_ib_unlock_cqs(send_cq, recv_cq);
}
+ mlx5_ib_unlock_cqs(send_cq, recv_cq);
+ spin_unlock_irqrestore(&dev->reset_flow_resource_lock, flags);
if (qp->ibqp.qp_type == IB_QPT_RAW_PACKET) {
destroy_raw_packet_qp(dev, qp);
@@ -2559,7 +2601,8 @@
}
pd = get_pd(qp);
- get_cqs(qp, &send_cq, &recv_cq);
+ get_cqs(qp->ibqp.qp_type, qp->ibqp.send_cq, qp->ibqp.recv_cq,
+ &send_cq, &recv_cq);
context->flags_pd = cpu_to_be32(pd ? pd->pdn : to_mpd(dev->devr.p0)->pdn);
context->cqn_send = send_cq ? cpu_to_be32(send_cq->mcq.cqn) : 0;
@@ -3658,6 +3701,7 @@
{
struct mlx5_wqe_ctrl_seg *ctrl = NULL; /* compiler warning */
struct mlx5_ib_dev *dev = to_mdev(ibqp->device);
+ struct mlx5_core_dev *mdev = dev->mdev;
struct mlx5_ib_qp *qp;
struct mlx5_ib_mr *mr;
struct mlx5_wqe_data_seg *dpseg;
@@ -3685,6 +3729,13 @@
spin_lock_irqsave(&qp->sq.lock, flags);
+ if (mdev->state == MLX5_DEVICE_STATE_INTERNAL_ERROR) {
+ err = -EIO;
+ *bad_wr = wr;
+ nreq = 0;
+ goto out;
+ }
+
for (nreq = 0; wr; nreq++, wr = wr->next) {
if (unlikely(wr->opcode >= ARRAY_SIZE(mlx5_ib_opcode))) {
mlx5_ib_warn(dev, "\n");
@@ -3986,6 +4037,8 @@
struct mlx5_ib_qp *qp = to_mqp(ibqp);
struct mlx5_wqe_data_seg *scat;
struct mlx5_rwqe_sig *sig;
+ struct mlx5_ib_dev *dev = to_mdev(ibqp->device);
+ struct mlx5_core_dev *mdev = dev->mdev;
unsigned long flags;
int err = 0;
int nreq;
@@ -3997,6 +4050,13 @@
spin_lock_irqsave(&qp->rq.lock, flags);
+ if (mdev->state == MLX5_DEVICE_STATE_INTERNAL_ERROR) {
+ err = -EIO;
+ *bad_wr = wr;
+ nreq = 0;
+ goto out;
+ }
+
ind = qp->rq.head & (qp->rq.wqe_cnt - 1);
for (nreq = 0; wr; nreq++, wr = wr->next) {
diff --git a/drivers/infiniband/hw/mlx5/srq.c b/drivers/infiniband/hw/mlx5/srq.c
index 3b2ddd6..55efb34 100644
--- a/drivers/infiniband/hw/mlx5/srq.c
+++ b/drivers/infiniband/hw/mlx5/srq.c
@@ -458,6 +458,8 @@
struct mlx5_ib_srq *srq = to_msrq(ibsrq);
struct mlx5_wqe_srq_next_seg *next;
struct mlx5_wqe_data_seg *scat;
+ struct mlx5_ib_dev *dev = to_mdev(ibsrq->device);
+ struct mlx5_core_dev *mdev = dev->mdev;
unsigned long flags;
int err = 0;
int nreq;
@@ -465,6 +467,12 @@
spin_lock_irqsave(&srq->lock, flags);
+ if (mdev->state == MLX5_DEVICE_STATE_INTERNAL_ERROR) {
+ err = -EIO;
+ *bad_wr = wr;
+ goto out;
+ }
+
for (nreq = 0; wr; nreq++, wr = wr->next) {
if (unlikely(wr->num_sge > srq->msrq.max_gs)) {
err = -EINVAL;
@@ -507,7 +515,7 @@
*srq->db.db = cpu_to_be32(srq->wqe_ctr);
}
-
+out:
spin_unlock_irqrestore(&srq->lock, flags);
return err;
diff --git a/include/linux/mlx5/cq.h b/include/linux/mlx5/cq.h
index 2be976d..2566f6d 100644
--- a/include/linux/mlx5/cq.h
+++ b/include/linux/mlx5/cq.h
@@ -58,6 +58,8 @@
void (*comp)(struct mlx5_core_cq *);
void *priv;
} tasklet_ctx;
+ int reset_notify_added;
+ struct list_head reset_notify;
};