Doing the segmentation in the forward path has one major drawback:
When using virtio, we may process gso udp packets coming
from host network stack. In that case, netfilter POSTROUTING
will see one packet with udp header followed by multiple ip
fragments.
Delay the segmentation and do it after POSTROUTING invocation
to avoid this.
Fixes: fe6cc55f3a ("net: ip, ipv6: handle gso skbs in forwarding path")
Signed-off-by: Florian Westphal <fw@strlen.de>
Signed-off-by: David S. Miller <davem@davemloft.net>
If conntrack defragments incoming ipv6 frags it stores largest original
frag size in ip6cb and sets ->local_df.
We must thus first test the largest original frag size vs. mtu, and not
vice versa.
Without this patch PKTTOOBIG is still generated in ip6_fragment() later
in the stack, but
1) IPSTATS_MIB_INTOOBIGERRORS won't increment
2) packet did (needlessly) traverse netfilter postrouting hook.
Fixes: fe6cc55f3a ("net: ip, ipv6: handle gso skbs in forwarding path")
Signed-off-by: Florian Westphal <fw@strlen.de>
Signed-off-by: David S. Miller <davem@davemloft.net>
local_df means 'ignore DF bit if set', so if its set we're
allowed to perform ip fragmentation.
This wasn't noticed earlier because the output path also drops such skbs
(and emits needed icmp error) and because netfilter ip defrag did not
set local_df until couple of days ago.
Only difference is that DF-packets-larger-than MTU now discarded
earlier (f.e. we avoid pointless netfilter postrouting trip).
While at it, drop the repeated test ip_exceeds_mtu, checking it once
is enough...
Fixes: fe6cc55f3a ("net: ip, ipv6: handle gso skbs in forwarding path")
Signed-off-by: Florian Westphal <fw@strlen.de>
Signed-off-by: David S. Miller <davem@davemloft.net>
Pull networking fixes from David Miller:
1) e1000e computes header length incorrectly wrt vlans, fix from Vlad
Yasevich.
2) ns_capable() check in sock_diag netlink code, from Andrew
Lutomirski.
3) Fix invalid queue pairs handling in virtio_net, from Amos Kong.
4) Checksum offloading busted in sxgbe driver due to incorrect
descriptor layout, fix from Byungho An.
5) Fix build failure with SMC_DEBUG set to 2 or larger, from Zi Shen
Lim.
6) Fix uninitialized A and X registers in BPF interpreter, from Alexei
Starovoitov.
7) Fix arch dependencies of candence driver.
8) Fix netlink capabilities checking tree-wide, from Eric W Biederman.
9) Don't dump IFLA_VF_PORTS if netlink request didn't ask for it in
IFLA_EXT_MASK, from David Gibson.
10) IPV6 FIB dump restart doesn't handle table changes that happen
meanwhile, causing the code to loop forever or emit dups, fix from
Kumar Sandararajan.
11) Memory leak on VF removal in bnx2x, from Yuval Mintz.
12) Bug fixes for new Altera TSE driver from Vince Bridgers.
13) Fix route lookup key in SCTP, from Xugeng Zhang.
14) Use BH blocking spinlocks in SLIP, as per a similar fix to CAN/SLCAN
driver. From Oliver Hartkopp.
15) TCP doesn't bump retransmit counters in some code paths, fix from
Eric Dumazet.
16) Clamp delayed_ack in tcp_cubic to prevent theoretical divides by
zero. Fix from Liu Yu.
17) Fix locking imbalance in error paths of HHF packet scheduler, from
John Fastabend.
18) Properly reference the transport module when vsock_core_init() runs,
from Andy King.
19) Fix buffer overflow in cdc_ncm driver, from Bjørn Mork.
20) IP_ECN_decapsulate() doesn't see a correct SKB network header in
ip_tunnel_rcv(), fix from Ying Cai.
* git://git.kernel.org/pub/scm/linux/kernel/git/davem/net: (132 commits)
net: macb: Fix race between HW and driver
net: macb: Remove 'unlikely' optimization
net: macb: Re-enable RX interrupt only when RX is done
net: macb: Clear interrupt flags
net: macb: Pass same size to DMA_UNMAP as used for DMA_MAP
ip_tunnel: Set network header properly for IP_ECN_decapsulate()
e1000e: Restrict MDIO Slow Mode workaround to relevant parts
e1000e: Fix issue with link flap on 82579
e1000e: Expand workaround for 10Mb HD throughput bug
e1000e: Workaround for dropped packets in Gig/100 speeds on 82579
net/mlx4_core: Don't issue PCIe speed/width checks for VFs
net/mlx4_core: Load the Eth driver first
net/mlx4_core: Fix slave id computation for single port VF
net/mlx4_core: Adjust port number in qp_attach wrapper when detaching
net: cdc_ncm: fix buffer overflow
Altera TSE: ALTERA_TSE should depend on HAS_DMA
vsock: Make transport the proto owner
net: sched: lock imbalance in hhf qdisc
net: mvmdio: Check for a valid interrupt instead of an error
net phy: Check for aneg completion before setting state to PHY_RUNNING
...
Pull USB fixes from Greg KH:
"Here are some small fixes and device ids for 3.15-rc4.
All have been in linux-next just fine"
* tag 'usb-3.15-rc4' of git://git.kernel.org/pub/scm/linux/kernel/git/gregkh/usb:
USB: Nokia 5300 should be treated as unusual dev
USB: Nokia 305 should be treated as unusual dev
fsl-usb: do not test for PHY_CLK_VALID bit on controller version 1.6
usb: storage: shuttle_usbat: fix discs being detected twice
usb: qcserial: add a number of Dell devices
USB: OHCI: fix problem with global suspend on ATI controllers
usb: gadget: at91-udc: fix irq and iomem resource retrieval
usb: phy: fsm: change "|" to "||" for condition OTG_STATE_A_WAIT_BCON at statemachine
usb: phy: fsm: update OTG HNP state transition
Pull tty/serial fixes from Greg KH:
"Here are some tty and serial driver fixes for things reported
recently"
* tag 'tty-3.15-rc4' of git://git.kernel.org/pub/scm/linux/kernel/git/gregkh/tty:
tty: Fix lockless tty buffer race
Revert "tty: Fix race condition between __tty_buffer_request_room and flush_to_ldisc"
drivers/tty/hvc: don't free hvc_console_setup after init
n_tty: Fix n_tty_write crash when echoing in raw mode
tty: serial: 8250_core.c Bug fix for Exar chips.
Pull staging / iio fixes from Greg KH:
"Here are some small IIO driver fixes for 3.15-rc4 that resolve some
reported issues"
* tag 'staging-3.15-rc4' of git://git.kernel.org/pub/scm/linux/kernel/git/gregkh/staging:
iio: adc: Nothing in ADC should be a bool CONFIG
iio: exynos_adc: use indio_dev->dev structure to handle child nodes
iio:imu:mpu6050: Fixed segfault in Invensens MPU driver due to null dereference
staging:iio:ad2s1200 fix missing parenthesis in a for statment.
Pull Ceph fixes from Sage Weil:
"First, there is a critical fix for the new primary-affinity function
that went into -rc1.
The second batch of patches from Zheng fix a range of problems with
directory fragmentation, readdir, and a few odds and ends for cephfs"
* 'for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/sage/ceph-client:
ceph: reserve caps for file layout/lock MDS requests
ceph: avoid releasing caps that are being used
ceph: clear directory's completeness when creating file
libceph: fix non-default values check in apply_primary_affinity()
ceph: use fpos_cmp() to compare dentry positions
ceph: check directory's completeness before emitting directory entry
Under "heavy" RX load, the driver cannot handle the descriptors fast
enough. In detail, when a descriptor is consumed, its used flag is
cleared and once the RX budget is consumed all descriptors with a
cleared used flag are prepared to receive more data. Under load though,
the HW may constantly receive more data and use those descriptors with a
cleared used flag before they are actually prepared for next usage.
The head and tail pointers into the RX-ring should always be valid and
we can omit clearing and checking of the used flag.
Signed-off-by: Soren Brinkmann <soren.brinkmann@xilinx.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Coverage data suggests that the unlikely case of receiving data while
the receive handler is running may not be that unlikely.
Coverage data after running iperf for a while:
91320: 891: work_done = bp->macbgem_ops.mog_rx(bp, budget);
91320: 892: if (work_done < budget) {
2362: 893: napi_complete(napi);
-: 894:
-: 895: /* Packets received while interrupts were disabled */
4724: 896: status = macb_readl(bp, RSR);
2362: 897: if (unlikely(status)) {
762: 898: if (bp->caps & MACB_CAPS_ISR_CLEAR_ON_WRITE)
762: 899: macb_writel(bp, ISR, MACB_BIT(RCOMP));
-: 900: napi_reschedule(napi);
-: 901: } else {
1600: 902: macb_writel(bp, IER, MACB_RX_INT_FLAGS);
-: 903: }
-: 904: }
Signed-off-by: Soren Brinkmann <soren.brinkmann@xilinx.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
When data is received during the driver processing received data the
NAPI is re-scheduled. In that case the RX interrupt should not be
re-enabled.
Signed-off-by: Soren Brinkmann <soren.brinkmann@xilinx.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
A few interrupt flags were not cleared in the ISR, resulting in a sytem
trapped in the ISR in cases one of those interrupts occurred. Clear all
flags to avoid such situations.
Signed-off-by: Soren Brinkmann <soren.brinkmann@xilinx.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Just as commit "net: macb: DMA-unmap full rx-buffer"
(48330e08fa), pass the size that
was used for mapping the memory also to the unmap routine to
avoid warnings from the DMA_API.
Signed-off-by: Soren Brinkmann <soren.brinkmann@xilinx.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
In ip_tunnel_rcv(), set skb->network_header to inner IP header
before IP_ECN_decapsulate().
Without the fix, IP_ECN_decapsulate() takes outer IP header as
inner IP header, possibly causing error messages or packet drops.
Note that this skb_reset_network_header() call was in this spot when
the original feature for checking consistency of ECN bits through
tunnels was added in eccc1bb8d4 ("tunnel: drop packet if ECN present
with not-ECT"). It was only removed from this spot in 3d7b46cd20
("ip_tunnel: push generic protocol handling to ip_tunnel module.").
Fixes: 3d7b46cd20 ("ip_tunnel: push generic protocol handling to ip_tunnel module.")
Reported-by: Neal Cardwell <ncardwell@google.com>
Signed-off-by: Ying Cai <ycai@google.com>
Acked-by: Neal Cardwell <ncardwell@google.com>
Acked-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Jeff Kirsher says:
====================
Intel Wired LAN Driver Updates
This series contains updates to e1000e only.
David provides four fixes for e1000e, first is a workaround for a hardware
erratum on 82579 devices which experienced packet loss in gigabit and 100
speeds when interconnect between the PHY and MAC is exiting K1 power saving
state. Second expands the scope of a workaround to include i217 and i218
parts as well to address over aggressive transmit behavior when connecting
at 10Mbs half-duplex. Next is to resolve a reported link flap issue on
82579 parts which was root caused as an interoperability problem between
82579 and at least some Broadcom PHYs in the Energy Efficient Ethernet wake
mechanism. Lastly, restricts the workaround of putting the PHY into MDIO
slow mode to access the PHY id to relevant parts since this issue has been
fixed on the newer hardware.
====================
Signed-off-by: David S. Miller <davem@davemloft.net>
It has been determined that the workaround of putting the PHY into MDIO
slow mode to access the PHY id is not necessary with Lynx Point and newer
parts. The issue that necessitated the workaround has been fixed on the
newer hardware.
We will maintains, as a last ditch attempt, the conversion to MDIO Slow
Mode in the failure branch when attempting to access the PHY id so as to
cover all contingencies.
Signed-off-by: Dave Ertman <davidx.m.ertman@intel.com>
Tested-by: Aaron Brown <aaron.f.brown@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
Several customers have reported a link flap issue on 82579. The symptoms
are random and intermittent link losses when 82579 is connected to specific
link partners. Issue has been root caused as interoperability problem
between 82579 and at least some Broadcom PHYs in the Energy Efficient
Ethernet wake mechanism.
To fix the issue, we are disabling the Phase Locked Loop shutdown in 100M
Low Power Idle. This solution will cause an increase of power in 100M EEE
link. It will cost additional 28mW in this specific mode.
Cc: Lukasz Adamczuk <lukasz.adamczuk@intel.com>
Signed-off-by: Dave Ertman <davidx.m.ertman@intel.com>
Tested-by: Aaron Brown <aaron.f.brown@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
In commit 772d05c51c "e1000e: slow performance
between two 82579 connected via 10Mbit hub", a workaround was put into place
to address the overaggressive transmit behavior of 82579 parts when connecting
at 10Mbs half-duplex.
This same behavior is seen on i217 and i218 parts as well. This patch expands
the original workaround to encompass these parts.
Signed-off-by: Dave Ertman <davidx.m.ertman@intel.com>
Tested-by: Aaron Brown <aaron.f.brown@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
This is a workaround for a HW erratum on 82579 devices.
Erratum is #23 in Intel 6 Series Chipset and Intel C200 Series Chipset
specification Update June 2013.
Problem: 82579 parts experience packet loss in Gig and 100 speeds
when interconnect between PHY and MAC is exiting K1 power saving state.
This was previously believed to only affect 1Gig speed, but has been observed
at 100Mbs also.
Workaround: Disable K1 for 82579 devices at Gig and 100 speeds.
Signed-off-by: Dave Ertman <davidx.m.ertman@intel.com>
Tested-by: Aaron Brown <aaron.f.brown@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
Or Gerlitz says:
====================
This series contains fixes for 3.15-rc, mostly around SRIOV. The patches by Jack,
Matan and myself fix few issues related to mlx4 SRIOV support for RoCE and single
port VFs, and the patch from Eyal eliminates checking PCI caps for VFs which is misleading.
Patches done against the net tree, commit 014f1b2 "net: bonding: Fix format string
mismatch in bond_sysfs.c"
We'd be happy to get Eyal's patch queued in your -stable list for 3.14.y
====================
Signed-off-by: David S. Miller <davem@davemloft.net>
Carrying out PCI speed/width checks through pcie_get_minimum_link()
on VFs yield wrong results, so remove them.
Fixes: b912b2f ('net/mlx4_core: Warn if device doesn't have enough PCI bandwidth')
Signed-off-by: Eyal Perry <eyalpe@mellanox.com>
Signed-off-by: Or Gerlitz <ogerlitz@mellanox.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
When running in SRIOV mode, VM that is assigned with a non-provisioned
Ethernet VFs get themselves a random mac when the Eth driver starts. In
this case, if the IB driver startup code that deals with RoCE runs first,
it will use a zero mac as the source mac for the Para-Virtual CM MADs
which is buggy. To handle that, we change the order of loading.
Signed-off-by: Or Gerlitz <ogerlitz@mellanox.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Florian Westphal
Bjørn Mork
Linus Torvalds
Soren Brinkmann
Ying Cai
David S. Miller
David Ertman
Eyal Perry
Or Gerlitz