The Linux Kernel Mailing List
 help / color / mirror / Atom feed
From: Zhu Yanjun <yanjun.zhu@linux.dev>
To: Yiran Qiu <eritque-arcus@ikuyo.dev>,
	zyjzyj2000@gmail.com, jgg@ziepe.ca, leon@kernel.org,
	"yanjun.zhu@linux.dev" <yanjun.zhu@linux.dev>
Cc: linux-rdma@vger.kernel.org, linux-kernel@vger.kernel.org,
	stable@vger.kernel.org
Subject: Re: [PATCH] RDMA/rxe: reject ODP accesses that wrap the iova space
Date: Sat, 21 Nov 2026 13:27:01 -0800	[thread overview]
Message-ID: <6462f655-303b-4df1-9ef0-b82bc482b55f@linux.dev> (raw)
In-Reply-To: <20261120060000.302440-1-eritque-arcus@ikuyo.dev>


在 2026/11/19 22:00, Yiran Qiu 写道:
> rxe_check_pagefault() decides whether an ODP access needs a page fault
> by walking its pages, and bounds the page index against the umem inside
> that walk. When iova + length wraps past 2^64 the loop does not run, so
> neither that bound nor the HMM_PFN_VALID and HMM_PFN_WRITE checks apply,
> and no fault is taken. __rxe_odp_mr_copy() and the atomic write and
> flush helpers then index map.pfn_list[] with (iova - ib_umem_start()) >>
> page_shift, which the registrant controls: ib_uverbs_reg_mr() only
> requires the MR's iova to share the page offset of the user address. The
> entry read past pfn_list[] goes to hmm_pfn_to_page(), and that page is
> copied to or from the peer.
>
> rxe_check_pagefault() also keeps the index in an int, so an index of
> 2^32 or more can be truncated to a value that passes the bound, while
> the flush and atomic write helpers use the full unsigned long.
>
> Reject a wrapping range in rxe_odp_map_range_and_lock(), before the
> fault path, whose umem range check wraps the same way, and keep the
> index in an unsigned long.
>
> Found with afb-afl, an LLM-driven kernel audit/repro pipeline; the fix
> and this changelog were also drafted with LLM assistance. Reproduced on
> an unmodified KASAN build of mainline by an unprivileged user with a
> self-connected rxe QP. An RDMA READ at the end of a two-page ODP MR
> whose iova range ends at 2^64:
>
>    BUG: KASAN: slab-out-of-bounds in rxe_odp_mr_copy+0x35f/0x410
>    Read of size 8 at addr ffff8881037a2210 by task kworker/u16:0/13
>    The buggy address is located 0 bytes to the right of
>     allocated 16-byte region [ffff8881037a2200, ffff8881037a2210)
>    Oops: general protection fault, maybe for address 0xffff888103d0c04a: 0000 [#1] SMP KASAN NOPTI
>    RIP: 0010:memcpy+0xc/0x30
>
> An ATOMIC WRITE to an MR whose iova is 2^44 above the umem start:
>
>    BUG: unable to handle page fault for address: ffffed11202b3194
>    RIP: 0010:rxe_odp_do_atomic_write+0x131/0x2f0
>
> With this patch both are NAKed with no report, and an in-range READ and
> ATOMIC WRITE return the same results as before. The wrap was not tried
> on the flush path. The rdma-core v65.0 tests, ODP included, pass on rxe
> with and without this patch.
>
> Fixes: 2fae67ab63db ("RDMA/rxe: Add support for Send/Recv/Write/Read with ODP")
> Cc: stable@vger.kernel.org # 6.15.x: 7797bbb0e86f: RDMA/rxe: bound the ODP page index against the umem in rxe_check_pagefault()
> Signed-off-by: Yiran Qiu <eritque-arcus@ikuyo.dev>
> ---
> Tested on mainline 4405c18c95d2 (v7.4-rc3+) in QEMU. Applies to
> rdma.git for-rc and for-next.
>
> This completes commit 7797bbb0e86f, whose bound sits in the loop that
> a wrapping range skips; stable trees need that commit first.
>
> Without KASAN, an RDMA READ or WRITE at the end of the range reads or
> overwrites a physical page the registrant chooses. On a lab kernel with
> CONFIG_RANDOMIZE_BASE=n this was taken from an unprivileged user to
> uid 0; with KASLR the chain also needs an address leak. The rxe link
> has to be created by an administrator; the trigger needs only the
> uverbs device. A reproducer and the test logs are available on request.
>
>   drivers/infiniband/sw/rxe/rxe_odp.c | 5 ++++-
>   1 file changed, 4 insertions(+), 1 deletion(-)
>
> diff --git a/drivers/infiniband/sw/rxe/rxe_odp.c b/drivers/infiniband/sw/rxe/rxe_odp.c
> index 9d27244f89bc..2ab64f128e57 100644
> --- a/drivers/infiniband/sw/rxe/rxe_odp.c
> +++ b/drivers/infiniband/sw/rxe/rxe_odp.c
> @@ -125,7 +125,7 @@ static inline bool rxe_check_pagefault(struct ib_umem_odp *umem_odp, u64 iova,
>   	bool need_fault = false;
>   	u64 access = HMM_PFN_VALID;
>   	u64 addr;
> -	int idx;
> +	unsigned long idx;
>   
>   	if (write)
>   		access |= HMM_PFN_WRITE;
> @@ -171,6 +171,9 @@ static int rxe_odp_map_range_and_lock(struct rxe_mr *mr, u64 iova, int length, u
>   	if (unlikely(length < 1))
>   		return -EINVAL;
>   
> +	if (unlikely(iova + length < iova))
> +		return -EFAULT;

Using the followings instead of the above.

"

u64 end;

if (check_add_overflow(iova, (u64)length, &end))

     return -EFAULT;

"

makes the 64-bit wrap-around check explicit and avoids ambiguity from 
implicit integer promotions.

Except that, I am fine with this commit.

Reviewed-by: Zhu Yanjun <yanjun.zhu@linux.dev>

Zhu Yanjun

> +
>   	mutex_lock(&umem_odp->umem_mutex);
>   
>   	need_fault = rxe_check_pagefault(umem_odp, iova, length, write);

-- 
Best Regards,
Yanjun.Zhu


      reply	other threads:[~2026-11-21 21:27 UTC|newest]

Thread overview: 2+ messages / expand[flat|nested]  mbox.gz  Atom feed  top
2026-11-20  6:00 [PATCH] RDMA/rxe: reject ODP accesses that wrap the iova space Yiran Qiu
2026-11-21 21:27 ` Zhu Yanjun [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=6462f655-303b-4df1-9ef0-b82bc482b55f@linux.dev \
    --to=yanjun.zhu@linux.dev \
    --cc=eritque-arcus@ikuyo.dev \
    --cc=jgg@ziepe.ca \
    --cc=leon@kernel.org \
    --cc=linux-kernel@vger.kernel.org \
    --cc=linux-rdma@vger.kernel.org \
    --cc=stable@vger.kernel.org \
    --cc=zyjzyj2000@gmail.com \
    /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