diff mbox series

[RFC] ath10k: Fix skb->len (properly) in ath10k_sdio_mbox_rx_packet

Message ID 20190827003326.147452-1-drinkcat@chromium.org (mailing list archive)
State RFC
Delegated to: Kalle Valo
Headers show
Series [RFC] ath10k: Fix skb->len (properly) in ath10k_sdio_mbox_rx_packet | expand

Commit Message

Nicolas Boichat Aug. 27, 2019, 12:33 a.m. UTC
(not a formal patch, take this as a bug report for now, I can clean
up depending on the feedback I get here)

There's at least 3 issues here, and the patch fixes 2/3 only, I'm not sure
how/if 1 should be handled.
 1. ath10k_sdio_mbox_rx_alloc allocating skb of a incorrect size (too
    small)
 2. ath10k_sdio_mbox_rx_packet calling skb_put with that incorrect size.
 3. ath10k_sdio_mbox_rx_process_packet attempts to fixup the size, but
    does not use proper skb_put commands to do so, so we end up with
    a mismatch between skb->head + skb->tail and skb->data + skb->len.

Let's start with 3, this is quite serious as this and causes corruptions
in the TCP stack, as the stack tries to coalesce packets, and relies on
skb->tail being correct (that is, skb_tail_pointer must point to the
first byte _after_ the data): one must never manipulate skb->len
directly.

Instead, we need to use skb_put to allocate more space (which updates
skb->len and skb->tail). But it seems odd to do that in
ath10k_sdio_mbox_rx_process_packet, so I move the code to
ath10k_sdio_mbox_rx_packet (point 2 above).

However, there is still something strange (point 1 above), why is
ath10k_sdio_mbox_rx_alloc allocating packets of the incorrect
(too small?) size? What happens if the packet is bigger than alloc_len?
Does this lead to corruption/lost data?

Fixes: 8530b4e7b22bc3b ("ath10k: sdio: set skb len for all rx packets")
Signed-off-by: Nicolas Boichat <drinkcat@chromium.org>

---

One simple way to test this is this scriplet, that sends a lot of
small packets over SSH:
(for i in `seq 1 300`; do echo $i; sleep 0.1; done) | ssh $IP cat

In my testing it rarely ever reach 300 without failure.

 drivers/net/wireless/ath/ath10k/sdio.c | 18 ++++++++++++------
 1 file changed, 12 insertions(+), 6 deletions(-)

Comments

Wen Gong Aug. 27, 2019, 3:33 a.m. UTC | #1
> -----Original Message-----
> From: ath10k <ath10k-bounces@lists.infradead.org> On Behalf Of Nicolas
> Boichat
> Sent: Tuesday, August 27, 2019 8:33 AM
> To: kvalo@codeaurora.org
> Cc: Alagu Sankar <alagusankar@silex-india.com>; netdev@vger.kernel.org;
> briannorris@chromium.org; linux-wireless@vger.kernel.org; linux-
> kernel@vger.kernel.org; ath10k@lists.infradead.org;
> wgong@codeaurora.org; niklas.cassel@linaro.org; tientzu@chromium.org;
> David S . Miller <davem@davemloft.net>
> Subject: [EXT] [PATCH, RFC] ath10k: Fix skb->len (properly) in
> ath10k_sdio_mbox_rx_packet
> 
> (not a formal patch, take this as a bug report for now, I can clean
> up depending on the feedback I get here)
> 
> There's at least 3 issues here, and the patch fixes 2/3 only, I'm not sure
> how/if 1 should be handled.
>  1. ath10k_sdio_mbox_rx_alloc allocating skb of a incorrect size (too
>     small)
>  2. ath10k_sdio_mbox_rx_packet calling skb_put with that incorrect size.
>  3. ath10k_sdio_mbox_rx_process_packet attempts to fixup the size, but
>     does not use proper skb_put commands to do so, so we end up with
>     a mismatch between skb->head + skb->tail and skb->data + skb->len.
> 
> Let's start with 3, this is quite serious as this and causes corruptions
> in the TCP stack, as the stack tries to coalesce packets, and relies on
> skb->tail being correct (that is, skb_tail_pointer must point to the
> first byte _after_ the data): one must never manipulate skb->len
> directly.
> 
> Instead, we need to use skb_put to allocate more space (which updates
> skb->len and skb->tail). But it seems odd to do that in
> ath10k_sdio_mbox_rx_process_packet, so I move the code to
> ath10k_sdio_mbox_rx_packet (point 2 above).
> 
> However, there is still something strange (point 1 above), why is
> ath10k_sdio_mbox_rx_alloc allocating packets of the incorrect
> (too small?) size? What happens if the packet is bigger than alloc_len?
> Does this lead to corruption/lost data?
> 
> Fixes: 8530b4e7b22bc3b ("ath10k: sdio: set skb len for all rx packets")
> Signed-off-by: Nicolas Boichat <drinkcat@chromium.org>
> 
> ---
> 
> One simple way to test this is this scriplet, that sends a lot of
> small packets over SSH:
> (for i in `seq 1 300`; do echo $i; sleep 0.1; done) | ssh $IP cat
> 
> In my testing it rarely ever reach 300 without failure.
> 
>  drivers/net/wireless/ath/ath10k/sdio.c | 18 ++++++++++++------
>  1 file changed, 12 insertions(+), 6 deletions(-)
> 
> diff --git a/drivers/net/wireless/ath/ath10k/sdio.c
> b/drivers/net/wireless/ath/ath10k/sdio.c
> index 8ed4fbd8d6c3888..a9f5002863ee7bb 100644
> --- a/drivers/net/wireless/ath/ath10k/sdio.c
> +++ b/drivers/net/wireless/ath/ath10k/sdio.c
> @@ -381,16 +381,14 @@ static int
> ath10k_sdio_mbox_rx_process_packet(struct ath10k *ar,
>  	struct ath10k_htc_hdr *htc_hdr = (struct ath10k_htc_hdr *)skb->data;
>  	bool trailer_present = htc_hdr->flags &
> ATH10K_HTC_FLAG_TRAILER_PRESENT;
>  	enum ath10k_htc_ep_id eid;
> -	u16 payload_len;
>  	u8 *trailer;
>  	int ret;
> 
> -	payload_len = le16_to_cpu(htc_hdr->len);
> -	skb->len = payload_len + sizeof(struct ath10k_htc_hdr);
> +	/* TODO: Remove this? */
If the pkt->act_len has set again in ath10k_sdio_mbox_rx_packet, seems not needed.
> +	WARN_ON(skb->len != le16_to_cpu(htc_hdr->len) + sizeof(*htc_hdr));
> 
>  	if (trailer_present) {
> -		trailer = skb->data + sizeof(*htc_hdr) +
> -			  payload_len - htc_hdr->trailer_len;
> +		trailer = skb->data + skb->len - htc_hdr->trailer_len;
> 
>  		eid = pipe_id_to_eid(htc_hdr->eid);
> 
> @@ -637,8 +635,16 @@ static int ath10k_sdio_mbox_rx_packet(struct
> ath10k *ar,
>  	ret = ath10k_sdio_readsb(ar, ar_sdio->mbox_info.htc_addr,
>  				 skb->data, pkt->alloc_len);
>  	pkt->status = ret;
> -	if (!ret)
> +	if (!ret) {
> +		/* Update actual length. */
> +		/* FIXME: This looks quite wrong, why is pkt->act_len not
> +		 * correct in the first place?
> +		 */
Firmware will do bundle for rx packet, and the aligned length by block size(256) of each packet's len is same 
in a bundle.

Eg. 
packet 1 len: 300, aligned length:512
packet 2 len: 400, aligned length:512
packet 3 len: 200, aligned length:256
packet 4 len: 100, aligned length:256
packet 5 len: 700, aligned length:768
packet 6 len: 600, aligned length:768

then packet 1,2 will in bundle 1, packet 3,4 in a bundle 2, packet 5,6 in a bundle 3.

For bundle 1, packet 1,2 will both allocate with len 512, and act_len is 300 first,
then packet 2's len will be overwrite to 400.

For bundle 2, packet 3,4 will both allocate with len 256, and act_len is 200 first,
then packet 4's len will be overwrite to 100.

For bundle 3, packet 5,6 will both allocate with len 768, and act_len is 700 first,
then packet 6's len will be overwrite to 600.

> +		struct ath10k_htc_hdr *htc_hdr =
> +			(struct ath10k_htc_hdr *)skb->data;
> +		pkt->act_len = le16_to_cpu(htc_hdr->len) + sizeof(*htc_hdr);
>  		skb_put(skb, pkt->act_len);
> +	}
> 
>  	return ret;
>  }
> --
> 2.23.0.187.g17f5b7556c-goog
> 
> 
> _______________________________________________
> ath10k mailing list
> ath10k@lists.infradead.org
> http://lists.infradead.org/mailman/listinfo/ath10k
Nicolas Boichat Aug. 27, 2019, 5:23 a.m. UTC | #2
On Tue, Aug 27, 2019 at 11:34 AM Wen Gong <wgong@qti.qualcomm.com> wrote:
>
> > -----Original Message-----
> > From: ath10k <ath10k-bounces@lists.infradead.org> On Behalf Of Nicolas
> > Boichat
> > Sent: Tuesday, August 27, 2019 8:33 AM
> > To: kvalo@codeaurora.org
> > Cc: Alagu Sankar <alagusankar@silex-india.com>; netdev@vger.kernel.org;
> > briannorris@chromium.org; linux-wireless@vger.kernel.org; linux-
> > kernel@vger.kernel.org; ath10k@lists.infradead.org;
> > wgong@codeaurora.org; niklas.cassel@linaro.org; tientzu@chromium.org;
> > David S . Miller <davem@davemloft.net>
> > Subject: [EXT] [PATCH, RFC] ath10k: Fix skb->len (properly) in
> > ath10k_sdio_mbox_rx_packet
> >
> > (not a formal patch, take this as a bug report for now, I can clean
> > up depending on the feedback I get here)
> >
> > There's at least 3 issues here, and the patch fixes 2/3 only, I'm not sure
> > how/if 1 should be handled.
> >  1. ath10k_sdio_mbox_rx_alloc allocating skb of a incorrect size (too
> >     small)
> >  2. ath10k_sdio_mbox_rx_packet calling skb_put with that incorrect size.
> >  3. ath10k_sdio_mbox_rx_process_packet attempts to fixup the size, but
> >     does not use proper skb_put commands to do so, so we end up with
> >     a mismatch between skb->head + skb->tail and skb->data + skb->len.
> >
> > Let's start with 3, this is quite serious as this and causes corruptions
> > in the TCP stack, as the stack tries to coalesce packets, and relies on
> > skb->tail being correct (that is, skb_tail_pointer must point to the
> > first byte _after_ the data): one must never manipulate skb->len
> > directly.
> >
> > Instead, we need to use skb_put to allocate more space (which updates
> > skb->len and skb->tail). But it seems odd to do that in
> > ath10k_sdio_mbox_rx_process_packet, so I move the code to
> > ath10k_sdio_mbox_rx_packet (point 2 above).
> >
> > However, there is still something strange (point 1 above), why is
> > ath10k_sdio_mbox_rx_alloc allocating packets of the incorrect
> > (too small?) size? What happens if the packet is bigger than alloc_len?
> > Does this lead to corruption/lost data?
> >
> > Fixes: 8530b4e7b22bc3b ("ath10k: sdio: set skb len for all rx packets")
> > Signed-off-by: Nicolas Boichat <drinkcat@chromium.org>
> >
> > ---
> >
> > One simple way to test this is this scriplet, that sends a lot of
> > small packets over SSH:
> > (for i in `seq 1 300`; do echo $i; sleep 0.1; done) | ssh $IP cat
> >
> > In my testing it rarely ever reach 300 without failure.
> >
> >  drivers/net/wireless/ath/ath10k/sdio.c | 18 ++++++++++++------
> >  1 file changed, 12 insertions(+), 6 deletions(-)
> >
> > diff --git a/drivers/net/wireless/ath/ath10k/sdio.c
> > b/drivers/net/wireless/ath/ath10k/sdio.c
> > index 8ed4fbd8d6c3888..a9f5002863ee7bb 100644
> > --- a/drivers/net/wireless/ath/ath10k/sdio.c
> > +++ b/drivers/net/wireless/ath/ath10k/sdio.c
> > @@ -381,16 +381,14 @@ static int
> > ath10k_sdio_mbox_rx_process_packet(struct ath10k *ar,
> >       struct ath10k_htc_hdr *htc_hdr = (struct ath10k_htc_hdr *)skb->data;
> >       bool trailer_present = htc_hdr->flags &
> > ATH10K_HTC_FLAG_TRAILER_PRESENT;
> >       enum ath10k_htc_ep_id eid;
> > -     u16 payload_len;
> >       u8 *trailer;
> >       int ret;
> >
> > -     payload_len = le16_to_cpu(htc_hdr->len);
> > -     skb->len = payload_len + sizeof(struct ath10k_htc_hdr);
> > +     /* TODO: Remove this? */
> If the pkt->act_len has set again in ath10k_sdio_mbox_rx_packet, seems not needed.

Sure, will drop.

> > +     WARN_ON(skb->len != le16_to_cpu(htc_hdr->len) + sizeof(*htc_hdr));
> >
> >       if (trailer_present) {
> > -             trailer = skb->data + sizeof(*htc_hdr) +
> > -                       payload_len - htc_hdr->trailer_len;
> > +             trailer = skb->data + skb->len - htc_hdr->trailer_len;
> >
> >               eid = pipe_id_to_eid(htc_hdr->eid);
> >
> > @@ -637,8 +635,16 @@ static int ath10k_sdio_mbox_rx_packet(struct
> > ath10k *ar,
> >       ret = ath10k_sdio_readsb(ar, ar_sdio->mbox_info.htc_addr,
> >                                skb->data, pkt->alloc_len);
> >       pkt->status = ret;
> > -     if (!ret)
> > +     if (!ret) {
> > +             /* Update actual length. */
> > +             /* FIXME: This looks quite wrong, why is pkt->act_len not
> > +              * correct in the first place?
> > +              */
> Firmware will do bundle for rx packet, and the aligned length by block size(256) of each packet's len is same
> in a bundle.
>
> Eg.
> packet 1 len: 300, aligned length:512
> packet 2 len: 400, aligned length:512
> packet 3 len: 200, aligned length:256
> packet 4 len: 100, aligned length:256
> packet 5 len: 700, aligned length:768
> packet 6 len: 600, aligned length:768
>
> then packet 1,2 will in bundle 1, packet 3,4 in a bundle 2, packet 5,6 in a bundle 3.
>
> For bundle 1, packet 1,2 will both allocate with len 512, and act_len is 300 first,
> then packet 2's len will be overwrite to 400.
>
> For bundle 2, packet 3,4 will both allocate with len 256, and act_len is 200 first,
> then packet 4's len will be overwrite to 100.
>
> For bundle 3, packet 5,6 will both allocate with len 768, and act_len is 700 first,
> then packet 6's len will be overwrite to 600.

Ok thanks, I'll send a v2 with an improved description.

> > +             struct ath10k_htc_hdr *htc_hdr =
> > +                     (struct ath10k_htc_hdr *)skb->data;
> > +             pkt->act_len = le16_to_cpu(htc_hdr->len) + sizeof(*htc_hdr);
> >               skb_put(skb, pkt->act_len);
> > +     }
> >
> >       return ret;
> >  }
> > --
> > 2.23.0.187.g17f5b7556c-goog
> >
> >
> > _______________________________________________
> > ath10k mailing list
> > ath10k@lists.infradead.org
> > http://lists.infradead.org/mailman/listinfo/ath10k
diff mbox series

Patch

diff --git a/drivers/net/wireless/ath/ath10k/sdio.c b/drivers/net/wireless/ath/ath10k/sdio.c
index 8ed4fbd8d6c3888..a9f5002863ee7bb 100644
--- a/drivers/net/wireless/ath/ath10k/sdio.c
+++ b/drivers/net/wireless/ath/ath10k/sdio.c
@@ -381,16 +381,14 @@  static int ath10k_sdio_mbox_rx_process_packet(struct ath10k *ar,
 	struct ath10k_htc_hdr *htc_hdr = (struct ath10k_htc_hdr *)skb->data;
 	bool trailer_present = htc_hdr->flags & ATH10K_HTC_FLAG_TRAILER_PRESENT;
 	enum ath10k_htc_ep_id eid;
-	u16 payload_len;
 	u8 *trailer;
 	int ret;
 
-	payload_len = le16_to_cpu(htc_hdr->len);
-	skb->len = payload_len + sizeof(struct ath10k_htc_hdr);
+	/* TODO: Remove this? */
+	WARN_ON(skb->len != le16_to_cpu(htc_hdr->len) + sizeof(*htc_hdr));
 
 	if (trailer_present) {
-		trailer = skb->data + sizeof(*htc_hdr) +
-			  payload_len - htc_hdr->trailer_len;
+		trailer = skb->data + skb->len - htc_hdr->trailer_len;
 
 		eid = pipe_id_to_eid(htc_hdr->eid);
 
@@ -637,8 +635,16 @@  static int ath10k_sdio_mbox_rx_packet(struct ath10k *ar,
 	ret = ath10k_sdio_readsb(ar, ar_sdio->mbox_info.htc_addr,
 				 skb->data, pkt->alloc_len);
 	pkt->status = ret;
-	if (!ret)
+	if (!ret) {
+		/* Update actual length. */
+		/* FIXME: This looks quite wrong, why is pkt->act_len not
+		 * correct in the first place?
+		 */
+		struct ath10k_htc_hdr *htc_hdr =
+			(struct ath10k_htc_hdr *)skb->data;
+		pkt->act_len = le16_to_cpu(htc_hdr->len) + sizeof(*htc_hdr);
 		skb_put(skb, pkt->act_len);
+	}
 
 	return ret;
 }