Due to a misplaced parenthesis, the expression

  (unlikely(offset) < 0),

which expands to

  (__builtin_expect(!!(offset), 0) < 0),

never evaluates to true. Therefore, when sending packets with
PF_PACKET/SOCK_DGRAM, packet_snd() does not abort as intended
if the creation of the layer 2 header fails.

Spotted by Coverity - CID 1259975 ("Operands don't affect result").

Fixes: 9c707762 ("packet: make packet_snd fail on len smaller than l2 header")
Signed-off-by: NChristoph Jaeger <cj@linux.com>
Acked-by: NEric Dumazet <edumazet@google.com>
Acked-by: NWillem de Bruijn <willemb@google.com>
Acked-by: NDaniel Borkmann <dborkman@redhat.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Commit d75b1ade ("net: less interrupt masking in NAPI") uncovered
wrong alx_poll() behavior.

A NAPI poll() handler is supposed to return exactly the budget when/if
napi_complete() has not been called.

It is also supposed to return number of frames that were received, so
that netdev_budget can have a meaning.

Also, in case of TX pressure, we still have to dequeue received
packets : alx_clean_rx_irq() has to be called even if
alx_clean_tx_irq(alx) returns false, otherwise device is half duplex.
Signed-off-by: NEric Dumazet <edumazet@google.com>
Fixes: d75b1ade ("net: less interrupt masking in NAPI")
Reported-by: NOded Gabbay <oded.gabbay@amd.com>
Bisected-by: NOded Gabbay <oded.gabbay@amd.com>
Tested-by: NOded Gabbay <oded.gabbay@amd.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

A NAPI poll() handler is supposed to return exactly the budget when/if
napi_complete() has not been called.

It is also supposed to return number of frames that were received, so
that netdev_budget can have a meaning.
Signed-off-by: NEric Dumazet <edumazet@google.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

This patch initialises the fep->netdev pointer. This pointer was not
initialised at all, but is used in fec_enet_timeout_work and in some
error paths.
Signed-off-by: NHubert Feurstein <h.feurstein@gmail.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

TRSCER register is configured differently by SoCs. TRSCER of R-Car Gen2 is
RINT8 bit only valid, other bits are reserved bits. This removes access to
TRSCER register reserve bit by adding variable trscer_err_mask to
sh_eth_cpu_data structure, set the register information to each SoCs.
Signed-off-by: NNobuhiro Iwamatsu <nobuhiro.iwamatsu.yj@renesas.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

FDR register of R-Car set in fdr_value can have the original settings.
This sets the value that is suitable for each SoCs to fdr_value of R8A777x
and R8A779x.
Signed-off-by: NNobuhiro Iwamatsu <nobuhiro.iwamatsu.yj@renesas.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Jeff Kirsher says:

====================
Intel Wired LAN Driver Updates 2015-01-06

This series contains fixes to i40e only.

Jesse provides a fix for when the driver was polling with interrupts
disabled the hardware would occasionally not write back descriptors.
His fix causes the driver to detect this situation and force an interrupt
to fire which will flush the stuck descriptor.

Anjali provides a couple of fixes, the first corrects an issue where
the receive port checksum error counter was incrementing incorrectly with
UDP encapsulated tunneled traffic.  The second fix resolves an issue where
the driver was examining the outer protocol layer to set the inner protocol
layer checksum offload.  In the case of TCP over IPv6 over an IPv4 based
VXLAN, the inner checksum offloads would be set to look for IPv4/UDP
instead of IPv6/TCP, so fixed the issue so that the driver will look at
the proper layer for encapsulation offload settings.

v2: fixed a bug in patch 01 of the series, where the interrupt rate impacted
    4 port workloads by reducing throughput.
====================
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Pull networking fixes from David Miller:
 "Just a pile of random fixes, including:

   1) Do not apply TSO limits to non-TSO packets, fix from Herbert Xu.

   2) MDI{,X} eeprom check in e100 driver is reversed, from John W.
      Linville.

   3) Missing error return assignments in several ethernet drivers, from
      Julia Lawall.

   4) Altera TSE device doesn't come back up after ifconfig down/up
      sequence, fix from Kostya Belezko.

   5) Add more cases to the check for whether the qmi_wwan device has a
      bogus MAC address and needs to be assigned a random one.  From
      Kristian Evensen.

   6) Fix interrupt hangs in CPSW, from Felipe Balbi.

   7) Implement ndo_features_check in r8152 so that the stack doesn't
      feed GSO packets which are outside of the chip's capabilities.
      From Hayes Wang"

* git://git.kernel.org/pub/scm/linux/kernel/git/davem/net: (26 commits)
  qla3xxx: don't allow never end busy loop
  xen-netback: fixing the propagation of the transmit shaper timeout
  r8152: support ndo_features_check
  batman-adv: fix potential TT client + orig-node memory leak
  batman-adv: fix multicast counter when purging originators
  batman-adv: fix counter for multicast supporting nodes
  batman-adv: fix lock class for decoding hash in network-coding.c
  batman-adv: fix delayed foreign originator recognition
  batman-adv: fix and simplify condition when bonding should be used
  Revert "mac80211: Fix accounting of the tailroom-needed counter"
  net: ethernet: cpsw: fix hangs with interrupts
  enic: free all rq buffs when allocation fails
  qmi_wwan: Set random MAC on devices with buggy fw
  openvswitch: Consistently include VLAN header in flow and port stats.
  tcp: Do not apply TSO segment limit to non-TSO packets
  Altera TSE: Add missing phydev
  net/mlx4_core: Fix error flow in mlx4_init_hca()
  net/mlx4_core: Correcly update the mtt's offset in the MR re-reg flow
  qlcnic: Fix return value in qlcnic_probe()
  net: axienet: fix error return code
  ...

Pull IPMI fixlet from Corey Minyard:
 "Fix a compile warning"

* tag 'for-linus-3' of git://git.code.sf.net/p/openipmi/linux-ipmi:
  ipmi: Fix compile warning with tv_usec

The driver was examining the outer protocol layer to set the inner protocol
layer checksum offload.  In the case of TCP over IPV6 over an IPv4 based
VXLAN the inner checksum offloads would be set to look for IPv4/UDP instead
of IPv6/TCP.  This code fixes that so that the driver will look at the
proper layer for encapsulation offload settings.
Signed-off-by: NAnjali Singhai <anjali.singhai@intel.com>
Signed-off-by: NGreg Rose <gregory.v.rose@intel.com>
Signed-off-by: NJeff Kirsher <jeffrey.t.kirsher@intel.com>

The Rx port checksum error counter was incrementing incorrectly with
UDP encapsulated tunneled traffic.  This patch fixes the problem so that
the port_rx_csum counter will show accurate statistics.
Signed-off-by: NAnjali Singhai <anjali.singhai@intel.com>
Signed-off-by: NGreg Rose <gregory.v.rose@intel.com>
Tested-by: NJim Young <james.m.young@intel.com>
Signed-off-by: NJeff Kirsher <jeffrey.t.kirsher@intel.com>

When the driver was polling with interrupts disabled the hardware
will occasionally not write back descriptors.  This patch causes
the driver to detect this situation and force an interrupt to
fire which will flush the stuck descriptor.  Does not conflict
with napi because if we are already polling the napi_schedule is
ignored.  Additionally the extra interrupts are rate limited, so
don't cause a burden to the CPU.

Change-ID: Iba4616d2a71288672a5f08e4512e2704b97335e8
Signed-off-by: NJesse Brandeburg <jesse.brandeburg@intel.com>
Signed-off-by: NJeff Kirsher <jeffrey.t.kirsher@intel.com>

The counter variable wasn't increased at all which may stuck under
certain circumstances.
Signed-off-by: NAndy Shevchenko <andy.shevchenko@gmail.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Pull ext4 bugfixes from Ted Ts'o:
 "Revert a potential seek_data/hole regression which shows up when using
  ext4 to handle ext3 file systems, plus two minor bug fixes"

* tag 'ext4_for_linus_stable' of git://git.kernel.org/pub/scm/linux/kernel/git/tytso/ext4:
  ext4: remove spurious KERN_INFO from ext4_warning call
  Revert "ext4: fix suboptimal seek_{data,hole} extents traversial"
  ext4: prevent online resize with backup superblock

Jay Foad reports that the address sanitizer test (asan) sometimes gets
confused by a stack pointer that ends up being outside the stack vma
that is reported by /proc/maps.

This happens due to an interaction between RLIMIT_STACK and the guard
page: when we do the guard page check, we ignore the potential error
from the stack expansion, which effectively results in a missing guard
page, since the expected stack expansion won't have been done.

And since /proc/maps explicitly ignores the guard page (commit
d7824370: "mm: fix up some user-visible effects of the stack guard
page"), the stack pointer ends up being outside the reported stack area.

This is the minimal patch: it just propagates the error.  It also
effectively makes the guard page part of the stack limit, which in turn
measn that the actual real stack is one page less than the stack limit.

Let's see if anybody notices.  We could teach acct_stack_growth() to
allow an extra page for a grow-up/grow-down stack in the rlimit test,
but I don't want to add more complexity if it isn't needed.
Reported-and-tested-by: NJay Foad <jay.foad@gmail.com>
Cc: stable@vger.kernel.org
Signed-off-by: NLinus Torvalds <torvalds@linux-foundation.org>

Included changes:
- ensure bonding is used (if enabled) for packets coming in the soft
  interface
- fix race condition to avoid orig_nodes to be deleted right after
  being added
- avoid false positive lockdep splats by assigning lockclass to
  the proper hashtable lock objects
- avoid miscounting of multicast 'disabled' nodes in the network
- fix memory leak in the Global Translation Table in case of
  originator interval change
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Since e9ce7cb6 ("xen-netback: Factor queue-specific data into queue struct"),
the transimt shaper timeout is always set to 0.  The value the user sets via
xenbus is never propagated to the transmit shaper.

This patch fixes the issue.

Cc: Anthony Liguori <aliguori@amazon.com>
Signed-off-by: NImre Palik <imrep@amazon.de>
Acked-by: NIan Campbell <ian.campbell@citrix.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Merge tag 'mac80211-for-davem-2015-01-06' of git://git.kernel.org/pub/scm/linux/kernel/git/jberg/mac80211

Here's just a single fix - a revert of a patch that broke the
p54 and cw2100 drivers (arguably due to bad assumptions there.)
Since this affects kernels since 3.17, I decided to revert for
now and we'll revisit this optimisation properly for -next.
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Support ndo_features_check to avoid:
 - the transport offset is more than the hw limitation when using hw checksum.
 - the skb->len of a GSO packet is more than the limitation.
Signed-off-by: NHayes Wang <hayeswang@realtek.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

This patch fixes a potential memory leak which can occur once an
originator times out. On timeout the according global translation table
entry might not get purged correctly. Furthermore, the non purged TT
entry will cause its orig-node to leak, too. Which additionally can lead
to the new multicast optimization feature not kicking in because of a
therefore bogus counter.

In detail: The batadv_tt_global_entry->orig_list holds the reference to
the orig-node. Usually this reference is released after
BATADV_PURGE_TIMEOUT through: _batadv_purge_orig()->
batadv_purge_orig_node()->batadv_update_route()->_batadv_update_route()->
batadv_tt_global_del_orig() which purges this global tt entry and
releases the reference to the orig-node.

However, if between two batadv_purge_orig_node() calls the orig-node
timeout grew to 2*BATADV_PURGE_TIMEOUT then this call path isn't
reached. Instead the according orig-node is removed from the
originator hash in _batadv_purge_orig(), the batadv_update_route()
part is skipped and won't be reached anymore.

Fixing the issue by moving batadv_tt_global_del_orig() out of the rcu
callback.
Signed-off-by: NLinus Lüssing <linus.luessing@c0d3.blue>
Acked-by: NAntonio Quartulli <antonio@meshcoding.com>
Signed-off-by: NMarek Lindner <mareklindner@neomailbox.ch>
Signed-off-by: NAntonio Quartulli <antonio@meshcoding.com>

When purging an orig_node we should only decrease counter tracking the
number of nodes without multicast optimizations support if it was
increased through this orig_node before.

A not yet quite initialized orig_node (meaning it did not have its turn
in the mcast-tvlv handler so far) which gets purged would not adhere to
this and will lead to a counter imbalance.

Fixing this by adding a check whether the orig_node is mcast-initalized
before decreasing the counter in the mcast-orig_node-purging routine.

Introduced by 60432d75
("batman-adv: Announce new capability via multicast TVLV")
Reported-by: NTobias Hachmer <tobias@hachmer.de>
Signed-off-by: NLinus Lüssing <linus.luessing@c0d3.blue>
Signed-off-by: NMarek Lindner <mareklindner@neomailbox.ch>
Signed-off-by: NAntonio Quartulli <antonio@meshcoding.com>

A miscounting of nodes having multicast optimizations enabled can lead
to multicast packet loss in the following scenario:

If the first OGM a node receives from another one has no multicast
optimizations support (no multicast tvlv) then we are missing to
increase the counter. This potentially leads to the wrong assumption
that we could safely use multicast optimizations.

Fixings this by increasing the counter if the initial OGM has the
multicast TVLV unset, too.

Introduced by 60432d75
("batman-adv: Announce new capability via multicast TVLV")
Reported-by: NTobias Hachmer <tobias@hachmer.de>
Signed-off-by: NLinus Lüssing <linus.luessing@c0d3.blue>
Signed-off-by: NMarek Lindner <mareklindner@neomailbox.ch>
Signed-off-by: NAntonio Quartulli <antonio@meshcoding.com>

batadv_has_set_lock_class() is called with the wrong hash table as first
argument (probably due to a copy-paste error), which leads to false
positives when running with lockdep.

Introduced-by: 612d2b4f
("batman-adv: network coding - save overheard and tx packets for decoding")
Signed-off-by: NMartin Hundebøll <martin@hundeboll.net>
Signed-off-by: NMarek Lindner <mareklindner@neomailbox.ch>
Signed-off-by: NAntonio Quartulli <antonio@meshcoding.com>

Currently it can happen that the reception of an OGM from a new
originator is not being accepted. More precisely it can happen that
an originator struct gets allocated and initialized
(batadv_orig_node_new()), even the TQ gets calculated and set correctly
(batadv_iv_ogm_calc_tq()) but still the periodic orig_node purging
thread will decide to delete it if it has a chance to jump between
these two function calls.

This is because batadv_orig_node_new() initializes the last_seen value
to zero and its caller (batadv_iv_ogm_orig_get()) makes it visible to
other threads by adding it to the hash table already.
batadv_iv_ogm_calc_tq() will set the last_seen variable to the correct,
current time a few lines later but if the purging thread jumps in between
that it will think that the orig_node timed out and will wrongly
schedule it for deletion already.

If the purging interval is the same as the originator interval (which is
the default: 1 second), then this game can continue for several rounds
until the random OGM jitter added enough difference between these
two (in tests, two to about four rounds seemed common).

Fixing this by initializing the last_seen variable of an orig_node
to the current time before adding it to the hash table.
Signed-off-by: NLinus Lüssing <linus.luessing@c0d3.blue>
Signed-off-by: NMarek Lindner <mareklindner@neomailbox.ch>
Signed-off-by: NAntonio Quartulli <antonio@meshcoding.com>

The current condition actually does NOT consider bonding when the
interface the packet came in from is the soft interface, which is the
opposite of what it should do (and the comment describes). Fix that and
slightly simplify the condition.
Reported-by: NRay Gibson <booray@gmail.com>
Signed-off-by: NSimon Wunderlich <sw@simonwunderlich.de>
Signed-off-by: NMarek Lindner <mareklindner@neomailbox.ch>
Signed-off-by: NAntonio Quartulli <antonio@meshcoding.com>

Pull powerpc fixes from Michael Ellerman:

 - Wire up sys_execveat(). Tested on 32 & 64 bit.

 - Fix for kdump on LE systems with cpus hot unplugged.

 - Revert Anton's fix for "kernel BUG at kernel/smpboot.c:134!", this
   broke other platforms, we'll do a proper fix for 3.20.

* tag 'powerpc-3.19-3' of git://git.kernel.org/pub/scm/linux/kernel/git/mpe/linux:
  Revert "powerpc: Secondary CPUs must set cpu_callin_map after setting active and online"
  powerpc/kdump: Ignore failure in enabling big endian exception during crash
  powerpc: Wire up sys_execveat() syscall

Pull ia64 fixlet from Tony Luck:
 "Add execveat syscall"

* tag 'please-pull-syscall' of git://git.kernel.org/pub/scm/linux/kernel/git/aegl/linux:
  [IA64] Enable execveat syscall for ia64

See commit 51f39a1f
    syscalls: implement execveat() system call
Signed-off-by: NTony Luck <tony.luck@intel.com>

This reverts commit ca34e3b5.

It turns out that the p54 and cw2100 drivers assume that there's
tailroom even when they don't say they really need it. However,
there's currently no way for them to explicitly say they do need
it, so for now revert this.

This fixes https://bugzilla.kernel.org/show_bug.cgi?id=90331.

Cc: stable@vger.kernel.org
Fixes: ca34e3b5 ("mac80211: Fix accounting of the tailroom-needed counter")
Reported-by: NChristopher Chavez <chrischavez@gmx.us>
Bisected-by: NLarry Finger <Larry.Finger@lwfinger.net>
Debugged-by: NChristian Lamparter <chunkeey@googlemail.com>
Signed-off-by: NJohannes Berg <johannes.berg@intel.com>

The CPSW IP implements pulse-signaled interrupts. Due to
that we must write a correct, pre-defined value to the
CPDMA_MACEOIVECTOR register so the controller generates
a pulse on the correct IRQ line to signal the End Of
Interrupt.

The way the driver is written today, all four IRQ lines
are requested using the same IRQ handler and, because of
that, we could fall into situations where a TX IRQ fires
but we tell the controller that we ended an RX IRQ (or
vice-versa). This situation triggers an IRQ storm on the
reserved IRQ 127 of INTC which will in turn call ack_bad_irq()
which will, then, print a ton of:

	unexpected IRQ trap at vector 00

In order to fix the problem, we are moving all calls to
cpdma_ctlr_eoi() inside the IRQ handler and making sure
we *always* write the correct value to the CPDMA_MACEOIVECTOR
register. Note that the algorithm assumes that IRQ numbers and
value-to-be-written-to-EOI are proportional, meaning that a
write of value 0 would trigger an EOI pulse for the RX_THRESHOLD
Interrupt and that's the IRQ number sitting in the 0-th index
of our irqs_table array.

This, however, is safe at least for current implementations of
CPSW so we will refrain from making the check smarter (and, as
a side-effect, slower) until we actually have a platform where
IRQ lines are swapped.

This patch has been tested for several days with AM335x- and
AM437x-based platforms. AM57x was left out because there are
still pending patches to enable ethernet in mainline for that
platform. A read of the TRM confirms the statement on previous
paragraph.
Reported-by: NYegor Yefremov <yegorslists@googlemail.com>
Fixes: 510a1e72 (drivers: net: davinci_cpdma: acknowledge interrupt properly)
Cc: <stable@vger.kernel.org> # v3.9+
Signed-off-by: NFelipe Balbi <balbi@ti.com>
Acked-by: NTony Lindgren <tony@atomide.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Pull UML fixes from Richard Weinberger:
 "Two fixes for UML regressions. Nothing exciting"

* 'for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/rw/uml:
  x86, um: actually mark system call tables readonly
  um: Skip futex_atomic_cmpxchg_inatomic() test

Commit 9fc2105a ("ARM: 7830/1: delay: don't bother reporting
bogomips in /proc/cpuinfo") breaks audio in python, and probably
elsewhere, with message

  FATAL: cannot locate cpu MHz in /proc/cpuinfo

I'm not the first one to hit it, see for example

  https://theredblacktree.wordpress.com/2014/08/10/fatal-cannot-locate-cpu-mhz-in-proccpuinfo/
  https://devtalk.nvidia.com/default/topic/765800/workaround-for-fatal-cannot-locate-cpu-mhz-in-proc-cpuinf/?offset=1

Reading original changelog, I have to say "Stop breaking working setups.
You know who you are!".
Signed-off-by: NPavel Machek <pavel@ucw.cz>
Signed-off-by: NLinus Torvalds <torvalds@linux-foundation.org>

Commit a074335a ("x86, um: Mark system call tables readonly") was
supposed to mark the sys_call_table in UML as RO by adding the const,
but it doesn't have the desired effect as it's nevertheless being placed
into the data section since __cacheline_aligned enforces sys_call_table
being placed into .data..cacheline_aligned instead. We need to use
the ____cacheline_aligned version instead to fix this issue.

Before:

$ nm -v arch/x86/um/sys_call_table_64.o | grep -1 "sys_call_table"
                 U sys_writev
0000000000000000 D sys_call_table
0000000000000000 D syscall_table_size

After:

$ nm -v arch/x86/um/sys_call_table_64.o | grep -1 "sys_call_table"
                 U sys_writev
0000000000000000 R sys_call_table
0000000000000000 D syscall_table_size

Fixes: a074335a ("x86, um: Mark system call tables readonly")
Cc: H. Peter Anvin <hpa@zytor.com>
Cc: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: NDaniel Borkmann <dborkman@redhat.com>
Signed-off-by: NRichard Weinberger <richard@nod.at>

futex_atomic_cmpxchg_inatomic() does not work on UML because
it triggers a copy_from_user() in kernel context.
On UML copy_from_user() can only be used if the kernel was called
by a real user space process such that UML can use ptrace()
to fetch the value.
Reported-by: NMiklos Szeredi <miklos@szeredi.hu>
Suggested-by: NGeert Uytterhoeven <geert@linux-m68k.org>
Signed-off-by: NRichard Weinberger <richard@nod.at>
Tested-by: NDaniel Walter <d.walter@0x90.at>

When allocation of all RQs fail, we do not free previously allocated buffers,
before returning error. This causes memory leak.

This patch fixes this by calling vnic_rq_clean(), which frees all the rq
buffers.
Signed-off-by: NGovindarajulu Varadarajan <_govind@gmx.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Some buggy firmwares export an incorrect MAC address (00:a0:c6:00:00:00). This
makes for example checking devices for random MAC addresses tricky, and you
might end up with multiple network interfaces with the same address.

This patch tries to fix, or at least improve, the situation by setting the MAC
address of devices with this firmware bug to a random address. I tested the
patch with two devices that has this firmware bug (Huawei E398 and E392), and
network traffic worked fine after changing the address.
Signed-off-by: NKristian Evensen <kristian.evensen@gmail.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Pull SCSI fixes from James Bottomley:
 "This is a set of three fixes: one to correct an abort path thinko
  causing failures (and a panic) in USB on device misbehaviour, One to
  fix an out of order issue in the fnic driver and one to match discard
  expectations to qemu which otherwise cause Linux to behave badly as a
  guest"

* tag 'scsi-fixes' of git://git.kernel.org/pub/scm/linux/kernel/git/jejb/scsi:
  SCSI: fix regression in scsi_send_eh_cmnd()
  fnic: IOMMU Fault occurs when IO and abort IO is out of order
  sd: tweak discard heuristics to work around QEMU SCSI issue

Until now, when VLAN acceleration was in use, the bytes of the VLAN header
were not included in port or flow byte counters.  They were however
included when VLAN acceleration was not used.  This commit corrects the
inconsistency, by always including the VLAN header in byte counters.

Previous discussion at
http://openvswitch.org/pipermail/dev/2014-December/049521.htmlReported-by: NMotonori Shindo <mshindo@vmware.com>
Signed-off-by: NBen Pfaff <blp@nicira.com>
Reviewed-by: NFlavio Leitner <fbl@sysclose.org>
Acked-by: NPravin B Shelar <pshelar@nicira.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Thomas Jarosch reported IPsec TCP stalls when a PMTU event occurs.

In fact the problem was completely unrelated to IPsec.  The bug is
also reproducible if you just disable TSO/GSO.

The problem is that when the MSS goes down, existing queued packet
on the TX queue that have not been transmitted yet all look like
TSO packets and get treated as such.

This then triggers a bug where tcp_mss_split_point tells us to
generate a zero-sized packet on the TX queue.  Once that happens
we're screwed because the zero-sized packet can never be removed
by ACKs.

Fixes: 1485348d ("tcp: Apply device TSO segment limit earlier")
Reported-by: NThomas Jarosch <thomas.jarosch@intra2net.com>
Signed-off-by: NHerbert Xu <herbert@gondor.apana.org.au>

Cheers,
Signed-off-by: NDavid S. Miller <davem@davemloft.net>