Re: [PATCH RESEND rdma-rc] RDMA/mlx5: Add dummy umem to IB_MR_TYPE_DM

[Jason Gunthorpe]

On Tue, Oct 19, 2021 at 01:23:13PM +0300, Leon Romanovsky wrote:
> From: Alaa Hleihel <alaa@xxxxxxxxxx>
> 
> After the cited patch, and for the case of IB_MR_TYPE_DM that doesn't
> have a umem (even though it is a user MR), function mlx5_free_priv_descs()
> will think that it's a kernel MR, leading to wrongly accessing mr->descs
> that will get wrong values in the union which leads to attempting to
> release resources that were not allocated in the first place.
> 
> For example:
>  DMA-API: mlx5_core 0000:08:00.1: device driver tries to free DMA memory it has not allocated [device address=0x0000000000000000] [size=0 bytes]
>  WARNING: CPU: 8 PID: 1021 at kernel/dma/debug.c:961 check_unmap+0x54f/0x8b0
>  RIP: 0010:check_unmap+0x54f/0x8b0
>  Call Trace:
>   debug_dma_unmap_page+0x57/0x60
>   mlx5_free_priv_descs+0x57/0x70 [mlx5_ib]
>   mlx5_ib_dereg_mr+0x1fb/0x3d0 [mlx5_ib]
>   ib_dereg_mr_user+0x60/0x140 [ib_core]
>   uverbs_destroy_uobject+0x59/0x210 [ib_uverbs]
>   uobj_destroy+0x3f/0x80 [ib_uverbs]
>   ib_uverbs_cmd_verbs+0x435/0xd10 [ib_uverbs]
>   ? uverbs_finalize_object+0x50/0x50 [ib_uverbs]
>   ? lock_acquire+0xc4/0x2e0
>   ? lock_acquired+0x12/0x380
>   ? lock_acquire+0xc4/0x2e0
>   ? lock_acquire+0xc4/0x2e0
>   ? ib_uverbs_ioctl+0x7c/0x140 [ib_uverbs]
>   ? lock_release+0x28a/0x400
>   ib_uverbs_ioctl+0xc0/0x140 [ib_uverbs]
>   ? ib_uverbs_ioctl+0x7c/0x140 [ib_uverbs]
>   __x64_sys_ioctl+0x7f/0xb0
>   do_syscall_64+0x38/0x90
> 
> Fix it by adding a dummy umem to IB_MR_TYPE_DM MRs.
> 
> Fixes: f18ec4223117 ("RDMA/mlx5: Use a union inside mlx5_ib_mr")
> Signed-off-by: Alaa Hleihel <alaa@xxxxxxxxxx>
> Signed-off-by: Leon Romanovsky <leonro@xxxxxxxxxx>
> RESEND: https://lore.kernel.org/all/9c6478b70dc23cfec3a7bfc345c30ff817e7e799.1631660866.git.leonro@xxxxxxxxxx
> 
> Our request to drop that original patch was because mr-->umem pointer is checked
> in rereg flow for the DM MRs with expectation to have NULL there. However DM is
> blocked for the rereg path in the commit 5ccbf63f87a3 ("IB/uverbs: Prevent
> reregistration of DM_MR to regular MR"), and the checks in mlx5_ib are redundant.

That logic in the core code is bogus and should be deleted.

It is perfeclty fine to use rereg to change the access flags on a DM
MR and mlx5 now implements that.

So let's not go down a path that blocks it.

Like this instead:

diff --git a/drivers/infiniband/hw/mlx5/mlx5_ib.h b/drivers/infiniband/hw/mlx5/mlx5_ib.h
index e636e954f6bf2a..4a7a56ed740b9b 100644
--- a/drivers/infiniband/hw/mlx5/mlx5_ib.h
+++ b/drivers/infiniband/hw/mlx5/mlx5_ib.h
@@ -664,7 +664,6 @@ struct mlx5_ib_mr {
 
 	/* User MR data */
 	struct mlx5_cache_ent *cache_ent;
-	struct ib_umem *umem;
 
 	/* This is zero'd when the MR is allocated */
 	union {
@@ -676,7 +675,7 @@ struct mlx5_ib_mr {
 			struct list_head list;
 		};
 
-		/* Used only by kernel MRs (umem == NULL) */
+		/* Used only by kernel MRs */
 		struct {
 			void *descs;
 			void *descs_alloc;
@@ -697,8 +696,9 @@ struct mlx5_ib_mr {
 			int data_length;
 		};
 
-		/* Used only by User MRs (umem != NULL) */
+		/* Used only by User MRs */
 		struct {
+			struct ib_umem *umem;
 			unsigned int page_shift;
 			/* Current access_flags */
 			int access_flags;
diff --git a/drivers/infiniband/hw/mlx5/mr.c b/drivers/infiniband/hw/mlx5/mr.c
index 221f0949794e35..997d133d00369d 100644
--- a/drivers/infiniband/hw/mlx5/mr.c
+++ b/drivers/infiniband/hw/mlx5/mr.c
@@ -1904,19 +1904,19 @@ mlx5_alloc_priv_descs(struct ib_device *device,
 	return ret;
 }
 
-static void
-mlx5_free_priv_descs(struct mlx5_ib_mr *mr)
+static void mlx5_free_priv_descs(struct mlx5_ib_mr *mr)
 {
-	if (!mr->umem && mr->descs) {
-		struct ib_device *device = mr->ibmr.device;
-		int size = mr->max_descs * mr->desc_size;
-		struct mlx5_ib_dev *dev = to_mdev(device);
+	struct ib_device *device = mr->ibmr.device;
+	int size = mr->max_descs * mr->desc_size;
+	struct mlx5_ib_dev *dev = to_mdev(device);
 
-		dma_unmap_single(&dev->mdev->pdev->dev, mr->desc_map, size,
-				 DMA_TO_DEVICE);
-		kfree(mr->descs_alloc);
-		mr->descs = NULL;
-	}
+	if (!mr->descs)
+		return;
+
+	dma_unmap_single(&dev->mdev->pdev->dev, mr->desc_map, size,
+			 DMA_TO_DEVICE);
+	kfree(mr->descs_alloc);
+	mr->descs = NULL;
 }
 
 int mlx5_ib_dereg_mr(struct ib_mr *ibmr, struct ib_udata *udata)
@@ -1978,7 +1978,7 @@ int mlx5_ib_dereg_mr(struct ib_mr *ibmr, struct ib_udata *udata)
 			return rc;
 	}
 
-	if (mr->umem) {
+	if (udata && mr->umem) {
 		bool is_odp = is_odp_mr(mr);
 
 		if (!is_odp)
@@ -1992,7 +1992,8 @@ int mlx5_ib_dereg_mr(struct ib_mr *ibmr, struct ib_udata *udata)
 	if (mr->cache_ent) {
 		mlx5_mr_cache_free(dev, mr);
 	} else {
-		mlx5_free_priv_descs(mr);
+		if (!udata)
+			mlx5_free_priv_descs(mr);
 		kfree(mr);
 	}
 	return 0;
@@ -2079,7 +2080,6 @@ static struct mlx5_ib_mr *mlx5_ib_alloc_pi_mr(struct ib_pd *pd,
 	if (err)
 		goto err_free_in;
 
-	mr->umem = NULL;
 	kfree(in);
 
 	return mr;
@@ -2206,7 +2206,6 @@ static struct ib_mr *__mlx5_ib_alloc_mr(struct ib_pd *pd,
 	}
 
 	mr->ibmr.device = pd->device;
-	mr->umem = NULL;
 
 	switch (mr_type) {
 	case IB_MR_TYPE_MEM_REG: