Linux RDMA and InfiniBand development
 help / color / mirror / Atom feed
From: sashiko-bot@kernel.org
To: "Abhijit Gangurde" <abhijit.gangurde@amd.com>
Cc: linux-rdma@vger.kernel.org
Subject: Re: [for-next v1 2/2] RDMA/ionic: Accept QP SQ and RQ rings as ioctl umem attributes
Date: Wed, 18 Nov 2026 15:51:34 +0000	[thread overview]
Message-ID: <sashiko-outbox-221381@kernel.org> (raw)
In-Reply-To: <20261118152313.113714-3-abhijit.gangurde@amd.com>

Thank you for your contribution! Sashiko AI review found 2 potential
issues to consider:
- [Medium] RDMA/ionic: double-fetch of buffer descriptor in
  ionic_get_user_q_umem()
- [Medium] RDMA/ionic: missing DMABUF addr==0 handling in CQ and QP
  qdesc validation
It also noticed 1 related pre-existing issue this patch does not
introduce:
- [Critical] RDMA/ionic: validation bypass via undefined behavior in
  ionic_validate_qdesc() [1]
--

commit 2a0ba584011ea872720029ef7dae9322c2160a45
Author: Abhijit Gangurde <abhijit.gangurde@amd.com>

RDMA/ionic: Accept QP SQ and RQ rings as ioctl umem attributes

This commit adds support for pinning QP SQ and RQ rings from the
UVERBS_ATTR_CREATE_QP_SQ_BUF_UMEM and UVERBS_ATTR_CREATE_QP_RQ_BUF_UMEM
ioctl buffer descriptor attributes, splitting queue geometry validation into
ionic_validate_qdesc_geom() for dma-buf descriptors.

> diff --git a/drivers/infiniband/hw/ionic/ionic_controlpath.c b/drivers/infiniband/hw/ionic/ionic_controlpath.c
> index 11458262f280d..02f1972126ed0 100644
> --- a/drivers/infiniband/hw/ionic/ionic_controlpath.c
> +++ b/drivers/infiniband/hw/ionic/ionic_controlpath.c
> @@ -14,13 +14,9 @@
>  #define ionic_set_ecn(tos)   (((tos) | 2u) & ~1u)
>  #define ionic_clear_ecn(tos)  ((tos) & ~3u)
>  
> -static int ionic_validate_qdesc(struct ionic_qdesc *q)
> +static int ionic_validate_qdesc_geom(struct ionic_qdesc *q)
>  {
> -	if (!q->addr || !q->size || !q->mask ||
> -	    !q->depth_log2 || !q->stride_log2)
> -		return -EINVAL;
> -
> -	if (q->addr & (PAGE_SIZE - 1))
> +	if (!q->size || !q->mask || !q->depth_log2 || !q->stride_log2)
>  		return -EINVAL;
>  
>  	if (q->mask != BIT(q->depth_log2) - 1)
> @@ -32,6 +28,14 @@ static int ionic_validate_qdesc(struct ionic_qdesc *q)
>  	return 0;
>  }
>  
> +static int ionic_validate_qdesc(struct ionic_qdesc *q)
> +{
> +	if (!q->addr || (q->addr & (PAGE_SIZE - 1)))
> +		return -EINVAL;
> +
> +	return ionic_validate_qdesc_geom(q);
> +}

[Severity: Critical]
This problem wasn't introduced by this patch, but can a user bypass the queue
size check in ionic_validate_qdesc_geom() [1] when q->depth_log2 +
q->stride_log2 is 64 or greater?

drivers/infiniband/hw/ionic/ionic_controlpath.c:ionic_validate_qdesc_geom() {
	...
	if (q->size < BIT_ULL(q->depth_log2 + q->stride_log2))
		return -EINVAL;

	return 0;
}

Neither q->depth_log2 nor q->stride_log2 is bounded before the shift, so a
sum of 64 or more triggers shift undefined behavior (wrapping the shift count
modulo 64 on x86_64) and lets a small q->size pass validation before the
out-of-bounds queue parameters are sent to firmware.

> @@ -1993,10 +1997,57 @@ static void ionic_qp_sq_destroy_cmb(struct ionic_ibdev *dev,
>  	ionic_put_cmb(dev->lif_cfg.lif, qp->sq_cmb_pgid, qp->sq_cmb_order);
>  }
>  
> +static int ionic_get_user_q_umem(struct ionic_ibdev *dev, struct ionic_qdesc *q,
> +				 const struct uverbs_attr_bundle *attrs,
> +				 u16 attr_id, struct ib_umem **umem_out)
> +{
> +	struct ib_uverbs_buffer_desc desc;
> +	struct ib_umem *umem;
> +	int rc;
> +
> +	if (attrs && uverbs_attr_is_valid(attrs, attr_id)) {
> +		rc = uverbs_get_buffer_desc(attrs, attr_id, &desc);
> +		if (rc)
> +			return rc;
> +
> +		if (desc.length != q->size)
> +			return -EINVAL;
> +
> +		if (desc.type == IB_UVERBS_BUFFER_TYPE_DMABUF) {
> +			if (desc.addr & (PAGE_SIZE - 1))
> +				return -EINVAL;
> +
> +			rc = ionic_validate_qdesc_geom(q);

[Severity: Medium]
Should the IB_UVERBS_BUFFER_TYPE_DMABUF branch also reject non-zero q->addr,
and does CQ creation need the same qdesc handling when req_cq->addr is zero?

The commit message notes that qdesc.addr is zero for a dma-buf descriptor,
which is why ionic_validate_qdesc_geom() is split out here, but this branch
does not check that q->addr is zero.  Meanwhile, ionic_create_cq_common() in
commit "RDMA/ionic: Accept CQ ring buffers as ioctl umem attributes" still
calls ionic_validate_qdesc(req_cq) unconditionally before
ionic_cq_get_user_umem():

drivers/infiniband/hw/ionic/ionic_controlpath.c:ionic_create_cq_common() {
	...
	if (udata) {
		rc = ionic_validate_qdesc(req_cq);
		if (rc)
			goto err_qdesc;

		rc = ionic_cq_get_user_umem(dev, cq, req_cq, attrs, udma_idx);
	...
}

Because ionic_validate_qdesc() rejects !q->addr, creating a dma-buf CQ with
req_cq->addr set to 0 always fails with -EINVAL.

> +		} else if (desc.type == IB_UVERBS_BUFFER_TYPE_VA &&
> +			   desc.addr == q->addr) {
> +			rc = ionic_validate_qdesc(q);
> +		} else {
> +			rc = -EINVAL;
> +		}
> +		if (rc)
> +			return rc;
> +
> +		umem = ib_umem_get_attr_or_va(&dev->ibdev, attrs, attr_id,
> +					      q->addr, q->size, 0);

[Severity: Medium]
Does calling ib_umem_get_attr_or_va() here re-fetch struct
ib_uverbs_buffer_desc from user memory after desc has already been validated?

UVERBS_ATTR_UMEM is defined without UA_ALLOC_AND_COPY and struct
ib_uverbs_buffer_desc is 32 bytes (not inline), so ib_umem_get_attr_or_va()
-> ib_umem_resolve_desc() -> uverbs_get_buffer_desc() performs a second
copy_from_user() of the descriptor and only checks that umem->length >=
q->size.  A concurrent userspace thread modifying the descriptor between the
two reads can bypass the checks on desc.addr page alignment, desc.length ==
q->size, and desc.type, programming firmware with a mismatched or unaligned
ring mapping.  Pinning the validated local descriptor via
ib_umem_get_desc(&dev->ibdev, &desc, 0) would avoid the second copy.

[1] https://sashiko.dev/#/bug/linux-f6f7b3eb-a66a-4c37-9e7c-001f6bd900bf

-- 
Sashiko AI review · https://sashiko.dev/#/patchset/20261118152313.113714-1-abhijit.gangurde@amd.com?part=2

      reply	other threads:[~2026-11-18 15:51 UTC|newest]

Thread overview: 5+ messages / expand[flat|nested]  mbox.gz  Atom feed  top
2026-11-18 15:23 [for-next v1 0/2] RDMA/ionic: Accept CQ, SQ, and RQ rings as umem attributes Abhijit Gangurde
2026-11-18 15:23 ` [for-next v1 1/2] RDMA/ionic: Accept CQ ring buffers as ioctl " Abhijit Gangurde
2026-11-18 15:54   ` sashiko-bot
2026-11-18 15:23 ` [for-next v1 2/2] RDMA/ionic: Accept QP SQ and RQ rings " Abhijit Gangurde
2026-11-18 15:51   ` sashiko-bot [this message]

Reply instructions:

You may reply publicly to this message via plain-text email
using any one of the following methods:

* Save the following mbox file, import it into your mail client,
  and reply-to-all from there: mbox

  Avoid top-posting and favor interleaved quoting:
  https://en.wikipedia.org/wiki/Posting_style#Interleaved_style

* Reply using the --to, --cc, and --in-reply-to
  switches of git-send-email(1):

  git send-email \
    --in-reply-to=sashiko-outbox-221381@kernel.org \
    --to=sashiko-bot@kernel.org \
    --cc=abhijit.gangurde@amd.com \
    --cc=linux-rdma@vger.kernel.org \
    --cc=sashiko-reviews@lists.linux.dev \
    /path/to/YOUR_REPLY

  https://kernel.org/pub/software/scm/git/docs/git-send-email.html

* If your mail client supports setting the In-Reply-To header
  via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line before the message body.
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox