Merge tag 'for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/rdma/rdma
authorLinus Torvalds <torvalds@linux-foundation.org>
Sat, 9 Apr 2022 04:29:02 +0000 (18:29 -1000)
committerLinus Torvalds <torvalds@linux-foundation.org>
Sat, 9 Apr 2022 04:29:02 +0000 (18:29 -1000)
Pull rdma fixes from Jason Gunthorpe:
 "Several bug fixes for old bugs:

   - Welcome Leon as co-maintainer for RDMA so we are back to having two
     people

   - Some corner cases are fixed in mlx5's MR code

   - Long standing CM bug where a DREQ at the wrong time can result in a
     long timeout

   - Missing locking and refcounting in hf1"

* tag 'for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/rdma/rdma:
  RDMA/hfi1: Fix use-after-free bug for mm struct
  IB/rdmavt: add lock to call to rvt_error_qp to prevent a race condition
  IB/cm: Cancel mad on the DREQ event when the state is MRA_REP_RCVD
  RDMA/mlx5: Add a missing update of cache->last_add
  RDMA/mlx5: Don't remove cache MRs when a delay is needed
  MAINTAINERS: Update qib and hfi1 related drivers
  MAINTAINERS: Add Leon Romanovsky to RDMA maintainers

MAINTAINERS
drivers/infiniband/core/cm.c
drivers/infiniband/hw/hfi1/mmu_rb.c
drivers/infiniband/hw/mlx5/mr.c
drivers/infiniband/sw/rdmavt/qp.c

index 17156b0..61d9f11 100644 (file)
@@ -8676,7 +8676,6 @@ F:        include/linux/cciss*.h
 F:     include/uapi/linux/cciss*.h
 
 HFI1 DRIVER
-M:     Mike Marciniszyn <mike.marciniszyn@cornelisnetworks.com>
 M:     Dennis Dalessandro <dennis.dalessandro@cornelisnetworks.com>
 L:     linux-rdma@vger.kernel.org
 S:     Supported
@@ -9599,6 +9598,7 @@ F:        drivers/iio/pressure/dps310.c
 
 INFINIBAND SUBSYSTEM
 M:     Jason Gunthorpe <jgg@nvidia.com>
+M:     Leon Romanovsky <leonro@nvidia.com>
 L:     linux-rdma@vger.kernel.org
 S:     Supported
 W:     https://github.com/linux-rdma/rdma-core
@@ -14657,7 +14657,6 @@ F:      drivers/rtc/rtc-optee.c
 
 OPA-VNIC DRIVER
 M:     Dennis Dalessandro <dennis.dalessandro@cornelisnetworks.com>
-M:     Mike Marciniszyn <mike.marciniszyn@cornelisnetworks.com>
 L:     linux-rdma@vger.kernel.org
 S:     Supported
 F:     drivers/infiniband/ulp/opa_vnic
@@ -16099,7 +16098,6 @@ F:      include/uapi/linux/qemu_fw_cfg.h
 
 QIB DRIVER
 M:     Dennis Dalessandro <dennis.dalessandro@cornelisnetworks.com>
-M:     Mike Marciniszyn <mike.marciniszyn@cornelisnetworks.com>
 L:     linux-rdma@vger.kernel.org
 S:     Supported
 F:     drivers/infiniband/hw/qib/
@@ -16617,7 +16615,6 @@ F:      drivers/net/ethernet/rdc/r6040.c
 
 RDMAVT - RDMA verbs software
 M:     Dennis Dalessandro <dennis.dalessandro@cornelisnetworks.com>
-M:     Mike Marciniszyn <mike.marciniszyn@cornelisnetworks.com>
 L:     linux-rdma@vger.kernel.org
 S:     Supported
 F:     drivers/infiniband/sw/rdmavt
index 35f0d5e..1c107d6 100644 (file)
@@ -2824,6 +2824,7 @@ static int cm_dreq_handler(struct cm_work *work)
        switch (cm_id_priv->id.state) {
        case IB_CM_REP_SENT:
        case IB_CM_DREQ_SENT:
+       case IB_CM_MRA_REP_RCVD:
                ib_cancel_mad(cm_id_priv->msg);
                break;
        case IB_CM_ESTABLISHED:
@@ -2831,8 +2832,6 @@ static int cm_dreq_handler(struct cm_work *work)
                    cm_id_priv->id.lap_state == IB_CM_MRA_LAP_RCVD)
                        ib_cancel_mad(cm_id_priv->msg);
                break;
-       case IB_CM_MRA_REP_RCVD:
-               break;
        case IB_CM_TIMEWAIT:
                atomic_long_inc(&work->port->counters[CM_RECV_DUPLICATES]
                                                     [CM_DREQ_COUNTER]);
index 876cc78..7333646 100644 (file)
@@ -80,6 +80,9 @@ void hfi1_mmu_rb_unregister(struct mmu_rb_handler *handler)
        unsigned long flags;
        struct list_head del_list;
 
+       /* Prevent freeing of mm until we are completely finished. */
+       mmgrab(handler->mn.mm);
+
        /* Unregister first so we don't get any more notifications. */
        mmu_notifier_unregister(&handler->mn, handler->mn.mm);
 
@@ -102,6 +105,9 @@ void hfi1_mmu_rb_unregister(struct mmu_rb_handler *handler)
 
        do_remove(handler, &del_list);
 
+       /* Now the mm may be freed. */
+       mmdrop(handler->mn.mm);
+
        kfree(handler);
 }
 
index 956f8e8..32ef67e 100644 (file)
@@ -574,8 +574,10 @@ static void __cache_work_func(struct mlx5_cache_ent *ent)
                spin_lock_irq(&ent->lock);
                if (ent->disabled)
                        goto out;
-               if (need_delay)
+               if (need_delay) {
                        queue_delayed_work(cache->wq, &ent->dwork, 300 * HZ);
+                       goto out;
+               }
                remove_cache_mr_locked(ent);
                queue_adjust_cache_locked(ent);
        }
@@ -625,6 +627,7 @@ static void mlx5_mr_cache_free(struct mlx5_ib_dev *dev, struct mlx5_ib_mr *mr)
 {
        struct mlx5_cache_ent *ent = mr->cache_ent;
 
+       WRITE_ONCE(dev->cache.last_add, jiffies);
        spin_lock_irq(&ent->lock);
        list_add_tail(&mr->list, &ent->head);
        ent->available_mrs++;
index ae50b56..8ef112f 100644 (file)
@@ -3190,7 +3190,11 @@ serr_no_r_lock:
        spin_lock_irqsave(&sqp->s_lock, flags);
        rvt_send_complete(sqp, wqe, send_status);
        if (sqp->ibqp.qp_type == IB_QPT_RC) {
-               int lastwqe = rvt_error_qp(sqp, IB_WC_WR_FLUSH_ERR);
+               int lastwqe;
+
+               spin_lock(&sqp->r_lock);
+               lastwqe = rvt_error_qp(sqp, IB_WC_WR_FLUSH_ERR);
+               spin_unlock(&sqp->r_lock);
 
                sqp->s_flags &= ~RVT_S_BUSY;
                spin_unlock_irqrestore(&sqp->s_lock, flags);