| Message ID | 1392249836-27151-1-git-send-email-emilgoode@gmail.com |
|---|---|
| State | Superseded, archived |
| Delegated to: | David Miller |
| Headers | show |
Emil Goode <emilgoode@gmail.com> writes: > This patch removes a generic hard_header_len check from the usbnet > module that is causing dropped packages under certain circumstances > for devices that send rx packets that cross urb boundaries. > > One example is the AX88772B which occasionally send rx packets that > cross urb boundaries where the remaining partial packet is sent with > no hardware header. When the buffer with a partial packet is of less > number of octets than the value of hard_header_len the buffer is > discarded by the usbnet module. > > With AX88772B this can be reproduced by using ping with a packet > size between 1965-1976. > > The bug has been reported here: > > https://bugzilla.kernel.org/show_bug.cgi?id=29082 > > This patch introduces the following changes: > - Removes the generic hard_header_len check in the rx_complete > function in the usbnet module. > - Introduces a ETH_HLEN check for skbs that are not cloned from > within a rx_fixup callback. > - For safety a hard_header_len check is added to each rx_fixup > callback function that could be affected by this change. > These extra checks could possibly be removed by someone > who has the hardware to test. > > The changes place full responsibility on the rx_fixup callback > functions that clone skbs to only pass valid skbs to the > usbnet_skb_return function. > > Signed-off-by: Emil Goode <emilgoode@gmail.com> > Reported-by: Igor Gnatenko <i.gnatenko.brain@gmail.com> Great work! Looks good to me. Just a couple of questions: > diff --git a/drivers/net/usb/qmi_wwan.c b/drivers/net/usb/qmi_wwan.c > index ff5c871..b2e2aee 100644 > --- a/drivers/net/usb/qmi_wwan.c > +++ b/drivers/net/usb/qmi_wwan.c > @@ -80,10 +80,10 @@ static int qmi_wwan_rx_fixup(struct usbnet *dev, struct sk_buff *skb) > { > __be16 proto; > > - /* usbnet rx_complete guarantees that skb->len is at least > - * hard_header_len, so we can inspect the dest address without > - * checking skb->len > - */ > + /* This check is no longer done by usbnet */ > + if (skb->len < dev->net->hard_header_len) > + return 0; > + > switch (skb->data[0] & 0xf0) { > case 0x40: > proto = htons(ETH_P_IP); > diff --git a/drivers/net/usb/rndis_host.c b/drivers/net/usb/rndis_host.c > index a48bc0f..524a47a 100644 > --- a/drivers/net/usb/rndis_host.c > +++ b/drivers/net/usb/rndis_host.c > @@ -492,6 +492,10 @@ EXPORT_SYMBOL_GPL(rndis_unbind); > */ > int rndis_rx_fixup(struct usbnet *dev, struct sk_buff *skb) > { > + /* This check is no longer done by usbnet */ > + if (skb->len < dev->net->hard_header_len) > + return 0; > + Wouldn't it be better to test against ETH_HLEN, since that is a constant and "obviously correct" in this case? > diff --git a/drivers/net/usb/usbnet.c b/drivers/net/usb/usbnet.c > index 4671da7..dd10d58 100644 > --- a/drivers/net/usb/usbnet.c > +++ b/drivers/net/usb/usbnet.c > @@ -542,17 +542,19 @@ static inline void rx_process (struct usbnet *dev, struct sk_buff *skb) > } > // else network stack removes extra byte if we forced a short packet > > - if (skb->len) { > - /* all data was already cloned from skb inside the driver */ > - if (dev->driver_info->flags & FLAG_MULTI_PACKET) > - dev_kfree_skb_any(skb); > - else > - usbnet_skb_return(dev, skb); > + /* all data was already cloned from skb inside the driver */ > + if (dev->driver_info->flags & FLAG_MULTI_PACKET) > + goto done; > + > + if (skb->len < ETH_HLEN) { > + dev->net->stats.rx_errors++; > + dev->net->stats.rx_length_errors++; > + netif_dbg(dev, rx_err, dev->net, "rx length %d\n", skb->len); > + } else { > + usbnet_skb_return(dev, skb); > return; > } > > - netif_dbg(dev, rx_err, dev->net, "drop\n"); > - dev->net->stats.rx_errors++; > done: > skb_queue_tail(&dev->done, skb); > } At first glance I wondered where the dev_kfree_skb_any(skb) went. But then I realized that you leave that to the common rx_cleanup path, using the dev->done queue. Is that right? If so, then I guess it could use a bit of explaining in the commit message? If I'm wrong, then I'm still looking for the explanation of where the dev_kfree_skb_any went :-) Bjørn -- To unsubscribe from this list: send the line "unsubscribe netdev" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
On Thu, Feb 13, 2014 at 10:05:39AM +0100, Bjørn Mork wrote: > Emil Goode <emilgoode@gmail.com> writes: > > > This patch removes a generic hard_header_len check from the usbnet > > module that is causing dropped packages under certain circumstances > > for devices that send rx packets that cross urb boundaries. > > > > One example is the AX88772B which occasionally send rx packets that > > cross urb boundaries where the remaining partial packet is sent with > > no hardware header. When the buffer with a partial packet is of less > > number of octets than the value of hard_header_len the buffer is > > discarded by the usbnet module. > > > > With AX88772B this can be reproduced by using ping with a packet > > size between 1965-1976. > > > > The bug has been reported here: > > > > https://bugzilla.kernel.org/show_bug.cgi?id=29082 > > > > This patch introduces the following changes: > > - Removes the generic hard_header_len check in the rx_complete > > function in the usbnet module. > > - Introduces a ETH_HLEN check for skbs that are not cloned from > > within a rx_fixup callback. > > - For safety a hard_header_len check is added to each rx_fixup > > callback function that could be affected by this change. > > These extra checks could possibly be removed by someone > > who has the hardware to test. > > > > The changes place full responsibility on the rx_fixup callback > > functions that clone skbs to only pass valid skbs to the > > usbnet_skb_return function. > > > > Signed-off-by: Emil Goode <emilgoode@gmail.com> > > Reported-by: Igor Gnatenko <i.gnatenko.brain@gmail.com> > > > Great work! Looks good to me. Thank you :) > Just a couple of questions: > > > diff --git a/drivers/net/usb/qmi_wwan.c b/drivers/net/usb/qmi_wwan.c > > index ff5c871..b2e2aee 100644 > > --- a/drivers/net/usb/qmi_wwan.c > > +++ b/drivers/net/usb/qmi_wwan.c > > @@ -80,10 +80,10 @@ static int qmi_wwan_rx_fixup(struct usbnet *dev, struct sk_buff *skb) > > { > > __be16 proto; > > > > - /* usbnet rx_complete guarantees that skb->len is at least > > - * hard_header_len, so we can inspect the dest address without > > - * checking skb->len > > - */ > > + /* This check is no longer done by usbnet */ > > + if (skb->len < dev->net->hard_header_len) > > + return 0; > > + > > switch (skb->data[0] & 0xf0) { > > case 0x40: > > proto = htons(ETH_P_IP); > > diff --git a/drivers/net/usb/rndis_host.c b/drivers/net/usb/rndis_host.c > > index a48bc0f..524a47a 100644 > > --- a/drivers/net/usb/rndis_host.c > > +++ b/drivers/net/usb/rndis_host.c > > @@ -492,6 +492,10 @@ EXPORT_SYMBOL_GPL(rndis_unbind); > > */ > > int rndis_rx_fixup(struct usbnet *dev, struct sk_buff *skb) > > { > > + /* This check is no longer done by usbnet */ > > + if (skb->len < dev->net->hard_header_len) > > + return 0; > > + > > Wouldn't it be better to test against ETH_HLEN, since that is a constant > and "obviously correct" in this case? Some minidrivers change the default hard_header_len value so using it guarantees that the patch will not make any change to how the code is currently working. Using ETH_HLEN could be more informative about what the minidriver should check before passing skbs to usbnet_skb_return(). Then I think the comment should be changed as well. My intention was to not make any changes that affect how the code works for devices I cannot test, but I think either way is fine and if you insist on changing it let me know. > > diff --git a/drivers/net/usb/usbnet.c b/drivers/net/usb/usbnet.c > > index 4671da7..dd10d58 100644 > > --- a/drivers/net/usb/usbnet.c > > +++ b/drivers/net/usb/usbnet.c > > @@ -542,17 +542,19 @@ static inline void rx_process (struct usbnet *dev, struct sk_buff *skb) > > } > > // else network stack removes extra byte if we forced a short packet > > > > - if (skb->len) { > > - /* all data was already cloned from skb inside the driver */ > > - if (dev->driver_info->flags & FLAG_MULTI_PACKET) > > - dev_kfree_skb_any(skb); > > - else > > - usbnet_skb_return(dev, skb); > > + /* all data was already cloned from skb inside the driver */ > > + if (dev->driver_info->flags & FLAG_MULTI_PACKET) > > + goto done; > > + > > + if (skb->len < ETH_HLEN) { > > + dev->net->stats.rx_errors++; > > + dev->net->stats.rx_length_errors++; > > + netif_dbg(dev, rx_err, dev->net, "rx length %d\n", skb->len); > > + } else { > > + usbnet_skb_return(dev, skb); > > return; > > } > > > > - netif_dbg(dev, rx_err, dev->net, "drop\n"); > > - dev->net->stats.rx_errors++; > > done: > > skb_queue_tail(&dev->done, skb); > > } > > > At first glance I wondered where the dev_kfree_skb_any(skb) went. But > then I realized that you leave that to the common rx_cleanup path, using > the dev->done queue. Is that right? If so, then I guess it could use a > bit of explaining in the commit message? Yes I should have put a comment in the changelog about this. All skbs that are passed to rx_process have their state set to rx_cleanup and just because the skb was cloned doesn't mean that we should free the original in a different way. As it is I think we are acctually missing a call to usb_free_urb that is called on the common rx_cleanup path. I will resend and add a comment about this. Best regards, Emil Goode -- To unsubscribe from this list: send the line "unsubscribe netdev" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
Emil Goode <emilgoode@gmail.com> writes: > Yes I should have put a comment in the changelog about this. All skbs that > are passed to rx_process have their state set to rx_cleanup and just because > the skb was cloned doesn't mean that we should free the original in a > different way. As it is I think we are acctually missing a call to > usb_free_urb that is called on the common rx_cleanup path. Yes, I wondered about that as well, thinking that this part actually should be a separate patch for net+stable. But then I wondered how we could possibly have had a bug like that living here for so long, and the answer is we haven't. You don't want to free the URB. It is already resubmitted with a different skb as its buffer. The call to usb_free_urb in the rx_cleanup path is there only for the cases where the URB is not resubmitted. The rx_complete callback is controlling this by updating entry->urb as appropriate. So it will always be NULL if rx_process is called, and the call to usb_free_urb has no effect. Rearranging this code to always take the same cleanup path is still a very nice cleanup IMHO. Bjørn -- To unsubscribe from this list: send the line "unsubscribe netdev" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
From: Emil Goode > > > int rndis_rx_fixup(struct usbnet *dev, struct sk_buff *skb) > > > { > > > + /* This check is no longer done by usbnet */ > > > + if (skb->len < dev->net->hard_header_len) > > > + return 0; > > > + > > > > Wouldn't it be better to test against ETH_HLEN, since that is a constant > > and "obviously correct" in this case? > > Some minidrivers change the default hard_header_len value so using it > guarantees that the patch will not make any change to how the code is > currently working. Using ETH_HLEN could be more informative about what > the minidriver should check before passing skbs to usbnet_skb_return(). > Then I think the comment should be changed as well. My intention was to > not make any changes that affect how the code works for devices I cannot > test, but I think either way is fine and if you insist on changing it > let me know. I think that test is to ensure that the data passed to the mini-driver contains the ethernet frame encapsulation header (this typically contains the actual frame length and some flags) so that the minidriver won't read off the end of the usb data. Any check for stupidly short ethernet frames would be later on. IIRC the absolute minimum 802.3 ethernet frame is 17 bytes (after frame padding has been stripped). David
diff --git a/drivers/net/usb/ax88179_178a.c b/drivers/net/usb/ax88179_178a.c index d6f64da..955df81 100644 --- a/drivers/net/usb/ax88179_178a.c +++ b/drivers/net/usb/ax88179_178a.c @@ -1118,6 +1118,10 @@ static int ax88179_rx_fixup(struct usbnet *dev, struct sk_buff *skb) u16 hdr_off; u32 *pkt_hdr; + /* This check is no longer done by usbnet */ + if (skb->len < dev->net->hard_header_len) + return 0; + skb_trim(skb, skb->len - 4); memcpy(&rx_hdr, skb_tail_pointer(skb), 4); le32_to_cpus(&rx_hdr); diff --git a/drivers/net/usb/gl620a.c b/drivers/net/usb/gl620a.c index e4a8a93..1cc24e6 100644 --- a/drivers/net/usb/gl620a.c +++ b/drivers/net/usb/gl620a.c @@ -84,6 +84,10 @@ static int genelink_rx_fixup(struct usbnet *dev, struct sk_buff *skb) u32 size; u32 count; + /* This check is no longer done by usbnet */ + if (skb->len < dev->net->hard_header_len) + return 0; + header = (struct gl_header *) skb->data; // get the packet count of the received skb diff --git a/drivers/net/usb/mcs7830.c b/drivers/net/usb/mcs7830.c index a305a7b..82d844a 100644 --- a/drivers/net/usb/mcs7830.c +++ b/drivers/net/usb/mcs7830.c @@ -526,8 +526,9 @@ static int mcs7830_rx_fixup(struct usbnet *dev, struct sk_buff *skb) { u8 status; - if (skb->len == 0) { - dev_err(&dev->udev->dev, "unexpected empty rx frame\n"); + /* This check is no longer done by usbnet */ + if (skb->len < dev->net->hard_header_len) { + dev_err(&dev->udev->dev, "unexpected tiny rx frame\n"); return 0; } diff --git a/drivers/net/usb/net1080.c b/drivers/net/usb/net1080.c index 0a85d92..4cbdb13 100644 --- a/drivers/net/usb/net1080.c +++ b/drivers/net/usb/net1080.c @@ -364,6 +364,10 @@ static int net1080_rx_fixup(struct usbnet *dev, struct sk_buff *skb) struct nc_trailer *trailer; u16 hdr_len, packet_len; + /* This check is no longer done by usbnet */ + if (skb->len < dev->net->hard_header_len) + return 0; + if (!(skb->len & 0x01)) { netdev_dbg(dev->net, "rx framesize %d range %d..%d mtu %d\n", skb->len, dev->net->hard_header_len, dev->hard_mtu, diff --git a/drivers/net/usb/qmi_wwan.c b/drivers/net/usb/qmi_wwan.c index ff5c871..b2e2aee 100644 --- a/drivers/net/usb/qmi_wwan.c +++ b/drivers/net/usb/qmi_wwan.c @@ -80,10 +80,10 @@ static int qmi_wwan_rx_fixup(struct usbnet *dev, struct sk_buff *skb) { __be16 proto; - /* usbnet rx_complete guarantees that skb->len is at least - * hard_header_len, so we can inspect the dest address without - * checking skb->len - */ + /* This check is no longer done by usbnet */ + if (skb->len < dev->net->hard_header_len) + return 0; + switch (skb->data[0] & 0xf0) { case 0x40: proto = htons(ETH_P_IP); diff --git a/drivers/net/usb/rndis_host.c b/drivers/net/usb/rndis_host.c index a48bc0f..524a47a 100644 --- a/drivers/net/usb/rndis_host.c +++ b/drivers/net/usb/rndis_host.c @@ -492,6 +492,10 @@ EXPORT_SYMBOL_GPL(rndis_unbind); */ int rndis_rx_fixup(struct usbnet *dev, struct sk_buff *skb) { + /* This check is no longer done by usbnet */ + if (skb->len < dev->net->hard_header_len) + return 0; + /* peripheral may have batched packets to us... */ while (likely(skb->len)) { struct rndis_data_hdr *hdr = (void *)skb->data; diff --git a/drivers/net/usb/smsc75xx.c b/drivers/net/usb/smsc75xx.c index f17b9e0..d9e7892 100644 --- a/drivers/net/usb/smsc75xx.c +++ b/drivers/net/usb/smsc75xx.c @@ -2106,6 +2106,10 @@ static void smsc75xx_rx_csum_offload(struct usbnet *dev, struct sk_buff *skb, static int smsc75xx_rx_fixup(struct usbnet *dev, struct sk_buff *skb) { + /* This check is no longer done by usbnet */ + if (skb->len < dev->net->hard_header_len) + return 0; + while (skb->len > 0) { u32 rx_cmd_a, rx_cmd_b, align_count, size; struct sk_buff *ax_skb; diff --git a/drivers/net/usb/smsc95xx.c b/drivers/net/usb/smsc95xx.c index 8dd54a0..424db65e 100644 --- a/drivers/net/usb/smsc95xx.c +++ b/drivers/net/usb/smsc95xx.c @@ -1723,6 +1723,10 @@ static void smsc95xx_rx_csum_offload(struct sk_buff *skb) static int smsc95xx_rx_fixup(struct usbnet *dev, struct sk_buff *skb) { + /* This check is no longer done by usbnet */ + if (skb->len < dev->net->hard_header_len) + return 0; + while (skb->len > 0) { u32 header, align_count; struct sk_buff *ax_skb; diff --git a/drivers/net/usb/sr9800.c b/drivers/net/usb/sr9800.c index 4175eb9..575be80 100644 --- a/drivers/net/usb/sr9800.c +++ b/drivers/net/usb/sr9800.c @@ -63,6 +63,10 @@ static int sr_rx_fixup(struct usbnet *dev, struct sk_buff *skb) { int offset = 0; + /* This check is no longer done by usbnet */ + if (skb->len < dev->net->hard_header_len) + return 0; + while (offset + sizeof(u32) < skb->len) { struct sk_buff *sr_skb; u16 size; diff --git a/drivers/net/usb/usbnet.c b/drivers/net/usb/usbnet.c index 4671da7..dd10d58 100644 --- a/drivers/net/usb/usbnet.c +++ b/drivers/net/usb/usbnet.c @@ -542,17 +542,19 @@ static inline void rx_process (struct usbnet *dev, struct sk_buff *skb) } // else network stack removes extra byte if we forced a short packet - if (skb->len) { - /* all data was already cloned from skb inside the driver */ - if (dev->driver_info->flags & FLAG_MULTI_PACKET) - dev_kfree_skb_any(skb); - else - usbnet_skb_return(dev, skb); + /* all data was already cloned from skb inside the driver */ + if (dev->driver_info->flags & FLAG_MULTI_PACKET) + goto done; + + if (skb->len < ETH_HLEN) { + dev->net->stats.rx_errors++; + dev->net->stats.rx_length_errors++; + netif_dbg(dev, rx_err, dev->net, "rx length %d\n", skb->len); + } else { + usbnet_skb_return(dev, skb); return; } - netif_dbg(dev, rx_err, dev->net, "drop\n"); - dev->net->stats.rx_errors++; done: skb_queue_tail(&dev->done, skb); } @@ -574,13 +576,6 @@ static void rx_complete (struct urb *urb) switch (urb_status) { /* success */ case 0: - if (skb->len < dev->net->hard_header_len) { - state = rx_cleanup; - dev->net->stats.rx_errors++; - dev->net->stats.rx_length_errors++; - netif_dbg(dev, rx_err, dev->net, - "rx length %d\n", skb->len); - } break; /* stalls need manual reset. this is rare ... except that
This patch removes a generic hard_header_len check from the usbnet module that is causing dropped packages under certain circumstances for devices that send rx packets that cross urb boundaries. One example is the AX88772B which occasionally send rx packets that cross urb boundaries where the remaining partial packet is sent with no hardware header. When the buffer with a partial packet is of less number of octets than the value of hard_header_len the buffer is discarded by the usbnet module. With AX88772B this can be reproduced by using ping with a packet size between 1965-1976. The bug has been reported here: https://bugzilla.kernel.org/show_bug.cgi?id=29082 This patch introduces the following changes: - Removes the generic hard_header_len check in the rx_complete function in the usbnet module. - Introduces a ETH_HLEN check for skbs that are not cloned from within a rx_fixup callback. - For safety a hard_header_len check is added to each rx_fixup callback function that could be affected by this change. These extra checks could possibly be removed by someone who has the hardware to test. The changes place full responsibility on the rx_fixup callback functions that clone skbs to only pass valid skbs to the usbnet_skb_return function. Signed-off-by: Emil Goode <emilgoode@gmail.com> Reported-by: Igor Gnatenko <i.gnatenko.brain@gmail.com> --- An alternative solution is to add the ETH_HLEN check to usbnet_skb_return() and add short skbs to the skb->done list to be cleaned up in rx_process(). drivers/net/usb/ax88179_178a.c | 4 ++++ drivers/net/usb/gl620a.c | 4 ++++ drivers/net/usb/mcs7830.c | 5 +++-- drivers/net/usb/net1080.c | 4 ++++ drivers/net/usb/qmi_wwan.c | 8 ++++---- drivers/net/usb/rndis_host.c | 4 ++++ drivers/net/usb/smsc75xx.c | 4 ++++ drivers/net/usb/smsc95xx.c | 4 ++++ drivers/net/usb/sr9800.c | 4 ++++ drivers/net/usb/usbnet.c | 25 ++++++++++--------------- 10 files changed, 45 insertions(+), 21 deletions(-)