[PATCH rdma-next 0/2] RDMA/efa: Synchronize IRQ before freeing admin queue and EQ

[PATCH rdma-next 0/2] RDMA/efa: Synchronize IRQ before freeing admin queue and EQ

[Leon Romanovsky]

Two similar patches for pre-existing issues.

Signed-off-by: Leon Romanovsky <leonro@nvidia.com>
---
Leon Romanovsky (2):
      RDMA/efa: Keep admin queues alive while IRQ is registered
      RDMA/efa: Keep EQ resources alive while IRQ is registered

 drivers/infiniband/hw/efa/efa_com.c  |  7 ++-----
 drivers/infiniband/hw/efa/efa_com.h  |  1 +
 drivers/infiniband/hw/efa/efa_main.c | 33 +++++++++++++++++++--------------
 3 files changed, 22 insertions(+), 19 deletions(-)
---
base-commit: 63367df6e7255067ad6a83abe0d2799dfa491876
change-id: 20260907-use-after-free-of-admin-queue-struct-5c10e817ce3a

Best regards,
--  
Leon Romanovsky <leonro@nvidia.com>

[PATCH rdma-next 1/2] RDMA/efa: Keep admin queues alive while IRQ is registered

[Leon Romanovsky]

From: Leon Romanovsky <leonro@nvidia.com>

The management IRQ handler accesses both the admin completion queue and the
async event queue. The driver registered the IRQ before constructing these
queues and destroyed them before freeing the IRQ, so the handler's lifetime
was not contained by the resources it accesses.

Initialize the queues with interrupts masked, request the IRQ, and then
switch to interrupt mode. On removal, reset the device and free the IRQ
before destroying the queues. Also reset the device before destroying the
queues if IRQ registration fails, because the device already has their DMA
addresses.

Fixes: b7f5e880f377 ("RDMA/efa: Add the efa module")
Signed-off-by: Leon Romanovsky <leonro@nvidia.com>
---
 drivers/infiniband/hw/efa/efa_com.c  |  4 +---
 drivers/infiniband/hw/efa/efa_main.c | 15 +++++++++------
 2 files changed, 10 insertions(+), 9 deletions(-)

diff --git a/drivers/infiniband/hw/efa/efa_com.c b/drivers/infiniband/hw/efa/efa_com.c
index 583b1cf0d721..cb2d448db2d7 100644
--- a/drivers/infiniband/hw/efa/efa_com.c
+++ b/drivers/infiniband/hw/efa/efa_com.c
@@ -850,7 +850,7 @@ int efa_com_admin_init(struct efa_com_dev *edev,
 
 	aq->dmadev = edev->dmadev;
 	aq->efa_dev = edev->efa_dev;
-	set_bit(EFA_AQ_STATE_POLLING_BIT, &aq->state);
+	efa_com_set_admin_polling_mode(edev, true);
 
 	sema_init(&aq->avail_cmds, aq->depth);
 
@@ -868,8 +868,6 @@ int efa_com_admin_init(struct efa_com_dev *edev,
 	if (err)
 		goto err_destroy_sq;
 
-	efa_com_set_admin_polling_mode(edev, false);
-
 	err = efa_com_admin_init_aenq(edev, aenq_handlers);
 	if (err)
 		goto err_destroy_cq;
diff --git a/drivers/infiniband/hw/efa/efa_main.c b/drivers/infiniband/hw/efa/efa_main.c
index 4cd80727dcd2..b1df9bb8fac2 100644
--- a/drivers/infiniband/hw/efa/efa_main.c
+++ b/drivers/infiniband/hw/efa/efa_main.c
@@ -619,18 +619,21 @@ static struct efa_dev *efa_probe_device(struct pci_dev *pdev)
 	edev->aq.msix_vector_idx = dev->admin_msix_vector_idx;
 	edev->aenq.msix_vector_idx = dev->admin_msix_vector_idx;
 
-	err = efa_set_mgmnt_irq(dev);
+	err = efa_com_admin_init(edev, &aenq_handlers);
 	if (err)
 		goto err_disable_msix;
 
-	err = efa_com_admin_init(edev, &aenq_handlers);
+	err = efa_set_mgmnt_irq(dev);
 	if (err)
-		goto err_free_mgmnt_irq;
+		goto err_destroy_admin;
+
+	efa_com_set_admin_polling_mode(edev, false);
 
 	return dev;
 
-err_free_mgmnt_irq:
-	efa_free_irq(dev, &dev->admin_irq);
+err_destroy_admin:
+	efa_com_dev_reset(edev, EFA_REGS_RESET_INIT_ERR);
+	efa_com_admin_destroy(edev);
 err_disable_msix:
 	efa_disable_msix(dev);
 err_reg_read_destroy:
@@ -654,8 +657,8 @@ static void efa_remove_device(struct pci_dev *pdev,
 
 	edev = &dev->edev;
 	efa_com_dev_reset(edev, reset_reason);
-	efa_com_admin_destroy(edev);
 	efa_free_irq(dev, &dev->admin_irq);
+	efa_com_admin_destroy(edev);
 	efa_disable_msix(dev);
 	efa_com_mmio_reg_read_destroy(edev);
 	devm_iounmap(&pdev->dev, edev->reg_bar);

-- 
2.55.0

[PATCH rdma-next 2/2] RDMA/efa: Keep EQ resources alive while IRQ is registered

[Leon Romanovsky]

From: Leon Romanovsky <leonro@nvidia.com>

The completion IRQ handler accesses the EQ state and DMA buffer. Its IRQ was
registered before that state was initialized, while teardown released the
buffer before free_irq() synchronized the handler.

Initialize the EQ without arming it, register the IRQ, and then arm it.
Reverse the resource order during teardown by freeing the IRQ before
destroying the EQ.

Fixes: 2a152512a155 ("RDMA/efa: CQ notifications")
Signed-off-by: Leon Romanovsky <leonro@nvidia.com>
---
 drivers/infiniband/hw/efa/efa_com.c  |  3 +--
 drivers/infiniband/hw/efa/efa_com.h  |  1 +
 drivers/infiniband/hw/efa/efa_main.c | 18 ++++++++++--------
 3 files changed, 12 insertions(+), 10 deletions(-)

diff --git a/drivers/infiniband/hw/efa/efa_com.c b/drivers/infiniband/hw/efa/efa_com.c
index cb2d448db2d7..04c6d63c449e 100644
--- a/drivers/infiniband/hw/efa/efa_com.c
+++ b/drivers/infiniband/hw/efa/efa_com.c
@@ -1252,7 +1252,7 @@ static void efa_com_destroy_eq(struct efa_com_dev *edev,
 				      err);
 }
 
-static void efa_com_arm_eq(struct efa_com_dev *edev, struct efa_com_eq *eeq)
+void efa_com_arm_eq(struct efa_com_dev *edev, struct efa_com_eq *eeq)
 {
 	u32 val = 0;
 
@@ -1341,7 +1341,6 @@ int efa_com_eq_init(struct efa_com_dev *edev, struct efa_com_eq *eeq,
 	eeq->phase = 1;
 	eeq->depth = params.depth;
 	eeq->cb = cb;
-	efa_com_arm_eq(edev, eeq);
 
 	return 0;
 
diff --git a/drivers/infiniband/hw/efa/efa_com.h b/drivers/infiniband/hw/efa/efa_com.h
index 0341704d0921..98fb6a42a6cb 100644
--- a/drivers/infiniband/hw/efa/efa_com.h
+++ b/drivers/infiniband/hw/efa/efa_com.h
@@ -169,6 +169,7 @@ int efa_com_admin_init(struct efa_com_dev *edev,
 void efa_com_admin_destroy(struct efa_com_dev *edev);
 int efa_com_eq_init(struct efa_com_dev *edev, struct efa_com_eq *eeq,
 		    efa_eqe_handler cb, u16 depth, u8 msix_vec);
+void efa_com_arm_eq(struct efa_com_dev *edev, struct efa_com_eq *eeq);
 void efa_com_eq_destroy(struct efa_com_dev *edev, struct efa_com_eq *eeq);
 int efa_com_dev_reset(struct efa_com_dev *edev,
 		      enum efa_regs_reset_reason_types reset_reason);
diff --git a/drivers/infiniband/hw/efa/efa_main.c b/drivers/infiniband/hw/efa/efa_main.c
index b1df9bb8fac2..753ed8430b69 100644
--- a/drivers/infiniband/hw/efa/efa_main.c
+++ b/drivers/infiniband/hw/efa/efa_main.c
@@ -302,28 +302,30 @@ static void efa_set_host_info(struct efa_dev *dev)
 
 static void efa_destroy_eq(struct efa_dev *dev, struct efa_eq *eq)
 {
-	efa_com_eq_destroy(&dev->edev, &eq->eeq);
 	efa_free_irq(dev, &eq->irq);
+	efa_com_eq_destroy(&dev->edev, &eq->eeq);
 }
 
 static int efa_create_eq(struct efa_dev *dev, struct efa_eq *eq, u32 msix_vec)
 {
 	int err;
 
-	efa_setup_comp_irq(dev, eq, msix_vec);
-	err = efa_request_irq(dev, &eq->irq);
+	err = efa_com_eq_init(&dev->edev, &eq->eeq, efa_process_eqe,
+			      dev->dev_attr.max_eq_depth, msix_vec);
 	if (err)
 		return err;
 
-	err = efa_com_eq_init(&dev->edev, &eq->eeq, efa_process_eqe,
-			      dev->dev_attr.max_eq_depth, msix_vec);
+	efa_setup_comp_irq(dev, eq, msix_vec);
+	err = efa_request_irq(dev, &eq->irq);
 	if (err)
-		goto err_free_comp_irq;
+		goto err_destroy_eq;
+
+	efa_com_arm_eq(&dev->edev, &eq->eeq);
 
 	return 0;
 
-err_free_comp_irq:
-	efa_free_irq(dev, &eq->irq);
+err_destroy_eq:
+	efa_com_eq_destroy(&dev->edev, &eq->eeq);
 	return err;
 }
 

-- 
2.55.0

Re: [PATCH rdma-next 1/2] RDMA/efa: Keep admin queues alive while IRQ is registered

[Michael Margolin]

On Mon, Sep 07, 2026 at 09:44:46PM +0300, Leon Romanovsky wrote:
> From: Leon Romanovsky <leonro@nvidia.com>
> 
> The management IRQ handler accesses both the admin completion queue and the
> async event queue. The driver registered the IRQ before constructing these
> queues and destroyed them before freeing the IRQ, so the handler's lifetime
> was not contained by the resources it accesses.
> 
> Initialize the queues with interrupts masked, request the IRQ, and then
> switch to interrupt mode. On removal, reset the device and free the IRQ
> before destroying the queues. Also reset the device before destroying the
> queues if IRQ registration fails, because the device already has their DMA
> addresses.
> 
> Fixes: b7f5e880f377 ("RDMA/efa: Add the efa module")
> Signed-off-by: Leon Romanovsky <leonro@nvidia.com>
> ---
>  drivers/infiniband/hw/efa/efa_com.c  |  4 +---
>  drivers/infiniband/hw/efa/efa_main.c | 15 +++++++++------
>  2 files changed, 10 insertions(+), 9 deletions(-)
> 
> diff --git a/drivers/infiniband/hw/efa/efa_com.c b/drivers/infiniband/hw/efa/efa_com.c
> index 583b1cf0d721..cb2d448db2d7 100644
> --- a/drivers/infiniband/hw/efa/efa_com.c
> +++ b/drivers/infiniband/hw/efa/efa_com.c
> @@ -850,7 +850,7 @@ int efa_com_admin_init(struct efa_com_dev *edev,
>  
>  	aq->dmadev = edev->dmadev;
>  	aq->efa_dev = edev->efa_dev;
> -	set_bit(EFA_AQ_STATE_POLLING_BIT, &aq->state);
> +	efa_com_set_admin_polling_mode(edev, true);
>  
>  	sema_init(&aq->avail_cmds, aq->depth);
>  
> @@ -868,8 +868,6 @@ int efa_com_admin_init(struct efa_com_dev *edev,
>  	if (err)
>  		goto err_destroy_sq;
>  
> -	efa_com_set_admin_polling_mode(edev, false);
> -
>  	err = efa_com_admin_init_aenq(edev, aenq_handlers);
>  	if (err)
>  		goto err_destroy_cq;
> diff --git a/drivers/infiniband/hw/efa/efa_main.c b/drivers/infiniband/hw/efa/efa_main.c
> index 4cd80727dcd2..b1df9bb8fac2 100644
> --- a/drivers/infiniband/hw/efa/efa_main.c
> +++ b/drivers/infiniband/hw/efa/efa_main.c
> @@ -619,18 +619,21 @@ static struct efa_dev *efa_probe_device(struct pci_dev *pdev)
>  	edev->aq.msix_vector_idx = dev->admin_msix_vector_idx;
>  	edev->aenq.msix_vector_idx = dev->admin_msix_vector_idx;
>  
> -	err = efa_set_mgmnt_irq(dev);
> +	err = efa_com_admin_init(edev, &aenq_handlers);
>  	if (err)
>  		goto err_disable_msix;
>  
> -	err = efa_com_admin_init(edev, &aenq_handlers);
> +	err = efa_set_mgmnt_irq(dev);
>  	if (err)
> -		goto err_free_mgmnt_irq;
> +		goto err_destroy_admin;
> +
> +	efa_com_set_admin_polling_mode(edev, false);
>  
>  	return dev;
>  
> -err_free_mgmnt_irq:
> -	efa_free_irq(dev, &dev->admin_irq);
> +err_destroy_admin:
> +	efa_com_dev_reset(edev, EFA_REGS_RESET_INIT_ERR);
> +	efa_com_admin_destroy(edev);
>  err_disable_msix:
>  	efa_disable_msix(dev);
>  err_reg_read_destroy:
> @@ -654,8 +657,8 @@ static void efa_remove_device(struct pci_dev *pdev,
>  
>  	edev = &dev->edev;
>  	efa_com_dev_reset(edev, reset_reason);
> -	efa_com_admin_destroy(edev);
>  	efa_free_irq(dev, &dev->admin_irq);
> +	efa_com_admin_destroy(edev);
>  	efa_disable_msix(dev);
>  	efa_com_mmio_reg_read_destroy(edev);
>  	devm_iounmap(&pdev->dev, edev->reg_bar);
> 
> -- 
> 2.55.0
>

Looks good to me, thanks!

Reviewed-by: Michael Margolin <mrgolin@amazon.com>

Re: [PATCH rdma-next 2/2] RDMA/efa: Keep EQ resources alive while IRQ is registered

[Michael Margolin]

On Mon, Sep 07, 2026 at 09:44:47PM +0300, Leon Romanovsky wrote:
> From: Leon Romanovsky <leonro@nvidia.com>
> 
> The completion IRQ handler accesses the EQ state and DMA buffer. Its IRQ was
> registered before that state was initialized, while teardown released the
> buffer before free_irq() synchronized the handler.
> 
> Initialize the EQ without arming it, register the IRQ, and then arm it.
> Reverse the resource order during teardown by freeing the IRQ before
> destroying the EQ.
> 
> Fixes: 2a152512a155 ("RDMA/efa: CQ notifications")
> Signed-off-by: Leon Romanovsky <leonro@nvidia.com>
> ---
>  drivers/infiniband/hw/efa/efa_com.c  |  3 +--
>  drivers/infiniband/hw/efa/efa_com.h  |  1 +
>  drivers/infiniband/hw/efa/efa_main.c | 18 ++++++++++--------
>  3 files changed, 12 insertions(+), 10 deletions(-)
> 
> diff --git a/drivers/infiniband/hw/efa/efa_com.c b/drivers/infiniband/hw/efa/efa_com.c
> index cb2d448db2d7..04c6d63c449e 100644
> --- a/drivers/infiniband/hw/efa/efa_com.c
> +++ b/drivers/infiniband/hw/efa/efa_com.c
> @@ -1252,7 +1252,7 @@ static void efa_com_destroy_eq(struct efa_com_dev *edev,
>  				      err);
>  }
>  
> -static void efa_com_arm_eq(struct efa_com_dev *edev, struct efa_com_eq *eeq)
> +void efa_com_arm_eq(struct efa_com_dev *edev, struct efa_com_eq *eeq)
>  {
>  	u32 val = 0;
>  
> @@ -1341,7 +1341,6 @@ int efa_com_eq_init(struct efa_com_dev *edev, struct efa_com_eq *eeq,
>  	eeq->phase = 1;
>  	eeq->depth = params.depth;
>  	eeq->cb = cb;
> -	efa_com_arm_eq(edev, eeq);
>  
>  	return 0;
>  
> diff --git a/drivers/infiniband/hw/efa/efa_com.h b/drivers/infiniband/hw/efa/efa_com.h
> index 0341704d0921..98fb6a42a6cb 100644
> --- a/drivers/infiniband/hw/efa/efa_com.h
> +++ b/drivers/infiniband/hw/efa/efa_com.h
> @@ -169,6 +169,7 @@ int efa_com_admin_init(struct efa_com_dev *edev,
>  void efa_com_admin_destroy(struct efa_com_dev *edev);
>  int efa_com_eq_init(struct efa_com_dev *edev, struct efa_com_eq *eeq,
>  		    efa_eqe_handler cb, u16 depth, u8 msix_vec);
> +void efa_com_arm_eq(struct efa_com_dev *edev, struct efa_com_eq *eeq);
>  void efa_com_eq_destroy(struct efa_com_dev *edev, struct efa_com_eq *eeq);
>  int efa_com_dev_reset(struct efa_com_dev *edev,
>  		      enum efa_regs_reset_reason_types reset_reason);
> diff --git a/drivers/infiniband/hw/efa/efa_main.c b/drivers/infiniband/hw/efa/efa_main.c
> index b1df9bb8fac2..753ed8430b69 100644
> --- a/drivers/infiniband/hw/efa/efa_main.c
> +++ b/drivers/infiniband/hw/efa/efa_main.c
> @@ -302,28 +302,30 @@ static void efa_set_host_info(struct efa_dev *dev)
>  
>  static void efa_destroy_eq(struct efa_dev *dev, struct efa_eq *eq)
>  {
> -	efa_com_eq_destroy(&dev->edev, &eq->eeq);
>  	efa_free_irq(dev, &eq->irq);
> +	efa_com_eq_destroy(&dev->edev, &eq->eeq);
>  }
>  
>  static int efa_create_eq(struct efa_dev *dev, struct efa_eq *eq, u32 msix_vec)
>  {
>  	int err;
>  
> -	efa_setup_comp_irq(dev, eq, msix_vec);
> -	err = efa_request_irq(dev, &eq->irq);
> +	err = efa_com_eq_init(&dev->edev, &eq->eeq, efa_process_eqe,
> +			      dev->dev_attr.max_eq_depth, msix_vec);
>  	if (err)
>  		return err;
>  
> -	err = efa_com_eq_init(&dev->edev, &eq->eeq, efa_process_eqe,
> -			      dev->dev_attr.max_eq_depth, msix_vec);
> +	efa_setup_comp_irq(dev, eq, msix_vec);
> +	err = efa_request_irq(dev, &eq->irq);
>  	if (err)
> -		goto err_free_comp_irq;
> +		goto err_destroy_eq;
> +
> +	efa_com_arm_eq(&dev->edev, &eq->eeq);
>  
>  	return 0;
>  
> -err_free_comp_irq:
> -	efa_free_irq(dev, &eq->irq);
> +err_destroy_eq:
> +	efa_com_eq_destroy(&dev->edev, &eq->eeq);
>  	return err;
>  }
>  
> 
> -- 
> 2.55.0
>

Looks good to me, thanks!

Reviewed-by: Michael Margolin <mrgolin@amazon.com>

Re: [PATCH rdma-next 0/2] RDMA/efa: Synchronize IRQ before freeing admin queue and EQ

[Leon Romanovsky]

On Mon, 07 Sep 2026 21:44:45 +0300, Leon Romanovsky wrote:
> Two similar patches for pre-existing issues.

Applied, thanks!

[1/2] RDMA/efa: Keep admin queues alive while IRQ is registered
      https://git.kernel.org/rdma/rdma/c/e08aca85c02ff2
[2/2] RDMA/efa: Keep EQ resources alive while IRQ is registered
      https://git.kernel.org/rdma/rdma/c/e22a3627b71517

Best regards,
-- 
Leon Romanovsky <leonro@nvidia.com>