[PATCH] crypto: chelsio: chtls: fix possible sleep-in-atomic-context bugs in abort_syn_rcv()

[PATCH] crypto: chelsio: chtls: fix possible sleep-in-atomic-context bugs in abort_syn_rcv()

[Jia-Ju Bai]

The driver may sleep while holding a spinlock.
The function call path (from bottom to top) in Linux 4.19 is:

drivers/crypto/chelsio/chtls/chtls_cm.c, 1806: 
	alloc_skb(GFP_KERNEL) in send_abort_rpl
drivers/crypto/chelsio/chtls/chtls_cm.c, 1925: 
	send_abort_rpl in abort_syn_rcv
drivers/crypto/chelsio/chtls/chtls_cm.c, 1920: 
	spin_lock in abort_syn_rcv

drivers/crypto/chelsio/chtls/chtls_cm.c, 1787: 
	alloc_skb(GFP_KERNEL) in send_defer_abort_rpl
drivers/crypto/chelsio/chtls/chtls_cm.c, 1811: 
	send_defer_abort_rpl in send_abort_rpl
drivers/crypto/chelsio/chtls/chtls_cm.c, 1925: 
    send_abort_rpl in abort_syn_rcv
drivers/crypto/chelsio/chtls/chtls_cm.c, 1920: 
    spin_lock in abort_syn_rcv

alloc_skb(GFP_KERNEL) can sleep at runtime.

To fix these possible bugs, GFP_KERNEL is replaced with GFP_ATOMIC.
Besides, in send_defer_abort_rpl(), error handling code is added to 
handle the failure of alloc_skb().

These bugs are found by a static analysis tool STCheck written by myself.

Signed-off-by: Jia-Ju Bai <baijiaju1990@gmail.com>
---
 drivers/crypto/chelsio/chtls/chtls_cm.c | 9 ++++++---
 1 file changed, 6 insertions(+), 3 deletions(-)

diff --git a/drivers/crypto/chelsio/chtls/chtls_cm.c b/drivers/crypto/chelsio/chtls/chtls_cm.c
index aca75237bbcf..e6e4c3ddc368 100644
--- a/drivers/crypto/chelsio/chtls/chtls_cm.c
+++ b/drivers/crypto/chelsio/chtls/chtls_cm.c
@@ -1805,8 +1805,11 @@ static void send_defer_abort_rpl(struct chtls_dev *cdev, struct sk_buff *skb)
 	struct cpl_abort_req_rss *req = cplhdr(skb);
 	struct sk_buff *reply_skb;
 
-	reply_skb = alloc_skb(sizeof(struct cpl_abort_rpl),
-			      GFP_KERNEL | __GFP_NOFAIL);
+	reply_skb = alloc_skb(sizeof(struct cpl_abort_rpl), GFP_ATOMIC);
+	if (!reply_skb) {
+		kfree_skb(skb);
+		return;
+	}
 	__skb_put(reply_skb, sizeof(struct cpl_abort_rpl));
 	set_abort_rpl_wr(reply_skb, GET_TID(req),
 			 (req->status & CPL_ABORT_NO_RST));
@@ -1825,7 +1828,7 @@ static void send_abort_rpl(struct sock *sk, struct sk_buff *skb,
 	csk = rcu_dereference_sk_user_data(sk);
 
 	reply_skb = alloc_skb(sizeof(struct cpl_abort_rpl),
-			      GFP_KERNEL);
+			      GFP_ATOMIC);
 
 	if (!reply_skb) {
 		req->status = (queue << 1);
-- 
2.17.1

Re: [PATCH] crypto: chelsio: chtls: fix possible sleep-in-atomic-context bugs in abort_syn_rcv()

[Herbert Xu]

On Wed, Dec 18, 2019 at 11:34:22AM +0800, Jia-Ju Bai wrote:
> The driver may sleep while holding a spinlock.
> The function call path (from bottom to top) in Linux 4.19 is:
> 
> drivers/crypto/chelsio/chtls/chtls_cm.c, 1806: 
> 	alloc_skb(GFP_KERNEL) in send_abort_rpl
> drivers/crypto/chelsio/chtls/chtls_cm.c, 1925: 
> 	send_abort_rpl in abort_syn_rcv
> drivers/crypto/chelsio/chtls/chtls_cm.c, 1920: 
> 	spin_lock in abort_syn_rcv
> 
> drivers/crypto/chelsio/chtls/chtls_cm.c, 1787: 
> 	alloc_skb(GFP_KERNEL) in send_defer_abort_rpl
> drivers/crypto/chelsio/chtls/chtls_cm.c, 1811: 
> 	send_defer_abort_rpl in send_abort_rpl
> drivers/crypto/chelsio/chtls/chtls_cm.c, 1925: 
>     send_abort_rpl in abort_syn_rcv
> drivers/crypto/chelsio/chtls/chtls_cm.c, 1920: 
>     spin_lock in abort_syn_rcv
> 
> alloc_skb(GFP_KERNEL) can sleep at runtime.
> 
> To fix these possible bugs, GFP_KERNEL is replaced with GFP_ATOMIC.
> Besides, in send_defer_abort_rpl(), error handling code is added to 
> handle the failure of alloc_skb().
> 
> These bugs are found by a static analysis tool STCheck written by myself.
> 
> Signed-off-by: Jia-Ju Bai <baijiaju1990@gmail.com>
> ---
>  drivers/crypto/chelsio/chtls/chtls_cm.c | 9 ++++++---
>  1 file changed, 6 insertions(+), 3 deletions(-)
> 
> diff --git a/drivers/crypto/chelsio/chtls/chtls_cm.c b/drivers/crypto/chelsio/chtls/chtls_cm.c
> index aca75237bbcf..e6e4c3ddc368 100644
> --- a/drivers/crypto/chelsio/chtls/chtls_cm.c
> +++ b/drivers/crypto/chelsio/chtls/chtls_cm.c
> @@ -1805,8 +1805,11 @@ static void send_defer_abort_rpl(struct chtls_dev *cdev, struct sk_buff *skb)
>  	struct cpl_abort_req_rss *req = cplhdr(skb);
>  	struct sk_buff *reply_skb;
>  
> -	reply_skb = alloc_skb(sizeof(struct cpl_abort_rpl),
> -			      GFP_KERNEL | __GFP_NOFAIL);
> +	reply_skb = alloc_skb(sizeof(struct cpl_abort_rpl), GFP_ATOMIC);
> +	if (!reply_skb) {
> +		kfree_skb(skb);
> +		return;
> +	}

This code now makes no sense because it's just trying the same
kmalloc twice.  Perhaps it really needs to be deferred as its
name suggests?

Thanks,
-- 
Email: Herbert Xu <herbert@gondor.apana.org.au>
Home Page: http://gondor.apana.org.au/~herbert/
PGP Key: http://gondor.apana.org.au/~herbert/pubkey.txt