Jeff Kirsher says:

====================
1GbE Intel Wired LAN Driver Updates 2017-03-17

This series contains updates to mainly igb, with one fix for ixgbe.

Alex does all the changes in the series, starting with adding support
for DMA_ATTR_WEAK_ORDERING to improve performance on some platforms.
Modified igb to use the length of the packet instead of the DD status
bit to determine if a new descriptor is ready to be processed.  Modified
the driver to only go through the region in the receive ring that was
designated to be cleaned up, instead of going through the entire ring
on cleanup.  Cleaned up the transmit side, by clearing the transmit
buffer_info only when resetting the rings.  Added a new upper limit for
receive, which is based on the size of a 2K buffer minus padding, which
will allow us to support build_skb going forward.  Fixed ethtool testing
to only sync on the size of the frame that is being tested, instead of
the entire receive buffer.  Updated the handling of page addresses to
always use a void pointer with the consistent name of "va" to indicate
that we are working with a virtual address.  Added a "chicken bit" so
that we can turn off the new receive allocation feature, in the case
where we need to fallback to the legacy receive path.  Added support for
using 3K buffers in order 1 pages the same way we were using 2K buffers
in 4K pages.  Added support for padding packet, since we limit the size
of the frame, we are able to write to an offset within the buffer instead
of having to write at the very start of the buffer.  This allows us to
leaving padding room for things like supporting XDP in the future.
Refactored the receive buffer page management, since there are 2-3 paths
that can be taken depending on what receive modes are enabled, so to
improve maintainability, break out the common bits into their own
functions.  Add support for build_skb, again.  Lastly, fixed a typo in
igb and ixgbe code comments.
====================
Signed-off-by: David S. Miller <davem@davemloft.net>

There was a typo that I had left in the code comments for the igb and ixgbe
functions that enabled build_skb support.
Signed-off-by: Alexander Duyck <alexander.h.duyck@intel.com>
Tested-by: Aaron Brown <aaron.f.brown@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>

This reverts commit f9d40f6a ("igb: Revert support for build_skb in
igb") and adds a few changes to update it to work with the latest version
of igb. We are now able to revert the removal of this due to the fact
that with the recent changes to the page count and the use of
DMA_ATTR_SKIP_CPU_SYNC we can make the pages writable so we should not be
invalidating the additional data added when we call build_skb.
Signed-off-by: Alexander Duyck <alexander.h.duyck@intel.com>
Tested-by: Aaron Brown <aaron.f.brown@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>

At this point we have 2 to 3 paths that can be taken depending on what Rx
modes are enabled.  In order to better support that and improve the
maintainability I am breaking out the common bits from those paths and
making them into their own functions.
Signed-off-by: Alexander Duyck <alexander.h.duyck@intel.com>
Tested-by: Aaron Brown <aaron.f.brown@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>

With the size of the frame limited we can now write to an offset within the
buffer instead of having to write at the very start of the buffer.  The
advantage to this is that it allows us to leave padding room for things
like supporting XDP in the future.
Signed-off-by: Alexander Duyck <alexander.h.duyck@intel.com>
Tested-by: Aaron Brown <aaron.f.brown@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>

This patch adds support for using 3K buffers in order 1 pages the same way
we were using 2K buffers in 4K pages.  We are reserving 1K of room for now
to have space available for future headroom and tailroom when we enable
build_skb support.

One side effect of this patch is that we can end up using a larger buffer
if jumbo frames is enabled.  The impact shouldn't be too great, but it
could hurt small packet performance for UDP workloads if jumbo frames is
enabled as the truesize of frames will be larger.
Signed-off-by: Alexander Duyck <alexander.h.duyck@intel.com>
Tested-by: Aaron Brown <aaron.f.brown@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>

Since there are potential drawbacks to the new Rx allocation approach I
thought it best to add a "chicken bit" so that we can turn the feature off
if in the event that a problem is found.

It also provides a means of validating the legacy Rx path in the event that
we are forced to fall back.  At some point in the future when we are
convinced we don't need it anymore we might be able to drop the legacy-rx
flag.
Signed-off-by: Alexander Duyck <alexander.h.duyck@intel.com>
Tested-by: Aaron Brown <aaron.f.brown@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>

Update the handling of page addresses so that we always refer to them using
a void pointer, and try to use the consistent name of va indicating we are
working with a virtual address.
Signed-off-by: Alexander Duyck <alexander.h.duyck@intel.com>
Tested-by: Aaron Brown <aaron.f.brown@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>

We only need to sync the size of the frame that is read to test.  We don't
need to sync the entire Rx buffer.  This way the testing is more consistent
with how we handle things in the receive path.
Signed-off-by: Alexander Duyck <alexander.h.duyck@intel.com>
Tested-by: Aaron Brown <aaron.f.brown@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>

In order to support the use of build_skb going forward it will be necessary
to place a maximum limit on the amount of data we can receive when jumbo
frames is not enabled.  In order to do this I am adding a new upper limit
for receive based on the size of a 2K buffer minus padding.
Signed-off-by: Alexander Duyck <alexander.h.duyck@intel.com>
Tested-by: Aaron Brown <aaron.f.brown@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>

In the case of the Tx rings we need to only clear the Tx buffer_info when
we are resetting the rings.  Ideally we do this when we configure the ring
to bring it back up instead of when we are taking it down in order to avoid
dirtying pages we don't need to.

In addition we don't need to clear the Tx descriptor ring since we will
fully repopulate it when we begin transmitting frames and next_to_watch can
be cleared to prevent the ring from being cleaned beyond that point instead
of needing to touch anything in the Tx descriptor ring.

Finally with these changes we can avoid having to reset the skb member of
the Tx buffer_info structure in the cleanup path since the skb will always
be associated with the first buffer which has next_to_watch set.
Signed-off-by: Alexander Duyck <alexander.h.duyck@intel.com>
Tested-by: Aaron Brown <aaron.f.brown@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>

This change makes it so that instead of going through the entire ring on Rx
cleanup we only go through the region that was designated to be cleaned up
and stop when we reach the region where new allocations should start.

In addition we can avoid having to perform a memset on the Rx buffer_info
structures until we are about to start using the ring again.  By deferring
this we can avoid dirtying the cache any more than we have to which can
help to improve the time needed to bring the interface down and then back
up again in a reset or suspend/resume cycle.
Signed-off-by: Alexander Duyck <alexander.h.duyck@intel.com>
Tested-by: Aaron Brown <aaron.f.brown@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>

This change makes it so that we use the length of the packet instead of the
DD status bit to determine if a new descriptor is ready to be processed.
The obvious advantage is that it cuts down on reads as we don't really even
need the DD bit if going from a 0 to a non-zero value on size is enough to
inform us that the packet has been completed.

In addition I have updated the code so that we only reset the Rx descriptor
length for descriptor zero when resetting a ring instead of having to do a
memset with 0 over the entire ring.  By doing this we can save some time on
initialization.
Signed-off-by: Alexander Duyck <alexander.h.duyck@intel.com>
Tested-by: Aaron Brown <aaron.f.brown@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>

Since we are already using DMA attributes in igb for Rx there is no reason
why we can't also apply DMA_ATTR_WEAK_ORDERING which is needed on some
platforms to improve performance.
Signed-off-by: Alexander Duyck <alexander.h.duyck@intel.com>
Tested-by: Aaron Brown <aaron.f.brown@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>

Information reported to ethtool about link modes is wrong for 25G NIC.  Fix
it by checking for presence of 25G NIC, checking the link speed reported by
NIC firmware, and then assigning proper values to the
ethtool_link_ksettings struct.
Signed-off-by: Manish Awasthi <manish.awasthi@cavium.com>
Signed-off-by: Felix Manlunas <felix.manlunas@cavium.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

Stephen Hemminger says:

====================
netvsc: small changes for net-next

One bugfix, and two non-code patches
====================
Signed-off-by: David S. Miller <davem@davemloft.net>

Not used anywhere.
Signed-off-by: Stephen Hemminger <sthemmin@microsoft.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

Add some short description of how callback's and NAPI interoperate.
Signed-off-by: Stephen Hemminger <sthemmin@microsoft.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

Change the argument to channel callback from the channel pointer
to the internal data structure containing per-channel info.
This avoids any possible races when callback happens during
initialization and makes IRQ code simpler.
Signed-off-by: Stephen Hemminger <sthemmin@microsoft.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

Alexei Starovoitov says:

====================
bpf: inline bpf_map_lookup_elem()

bpf_map_lookup_elem() is one of the most frequently used helper functions.
Improve JITed program performance by inlining this helper.

bpf_map_type	before  after
hash		58M	74M
array		174M	280M

The values are number of lookups per second in ideal conditions
measured by micro-benchmark in patch 6.

The 'perf report' for HASH map type:
before:
    54.23%  map_perf_test  [kernel.kallsyms]  [k] __htab_map_lookup_elem
    14.24%  map_perf_test  [kernel.kallsyms]  [k] lookup_elem_raw
     8.84%  map_perf_test  [kernel.kallsyms]  [k] htab_map_lookup_elem
     5.93%  map_perf_test  [kernel.kallsyms]  [k] bpf_map_lookup_elem
     2.30%  map_perf_test  [kernel.kallsyms]  [k] bpf_prog_da4fc6a3f41761a2
     1.49%  map_perf_test  [kernel.kallsyms]  [k] kprobe_ftrace_handler

after:
    60.03%  map_perf_test  [kernel.kallsyms]  [k] __htab_map_lookup_elem
    18.07%  map_perf_test  [kernel.kallsyms]  [k] lookup_elem_raw
     2.91%  map_perf_test  [kernel.kallsyms]  [k] bpf_prog_da4fc6a3f41761a2
     1.94%  map_perf_test  [kernel.kallsyms]  [k] _einittext
     1.90%  map_perf_test  [kernel.kallsyms]  [k] __audit_syscall_exit
     1.72%  map_perf_test  [kernel.kallsyms]  [k] kprobe_ftrace_handler

so the cost of htab_map_lookup_elem() and bpf_map_lookup_elem()
is gone after inlining.

'per-cpu' and 'lru' map types can be optimized similarly in the future.

Note the sparse will complain that bpf is addictive ;)
kernel/bpf/hashtab.c:438:19: sparse: subtraction of functions? Share your drugs
kernel/bpf/verifier.c:3342:38: sparse: subtraction of functions? Share your drugs
it's not a new warning, just in new places.
====================
Signed-off-by: David S. Miller <davem@davemloft.net>

$ map_perf_test 128
speed of HASH bpf_map_lookup_elem() in lookups per second
	w/o JIT		w/JIT
before	46M		58M
after	42M		74M

perf report
before:
    54.23%  map_perf_test  [kernel.kallsyms]  [k] __htab_map_lookup_elem
    14.24%  map_perf_test  [kernel.kallsyms]  [k] lookup_elem_raw
     8.84%  map_perf_test  [kernel.kallsyms]  [k] htab_map_lookup_elem
     5.93%  map_perf_test  [kernel.kallsyms]  [k] bpf_map_lookup_elem
     2.30%  map_perf_test  [kernel.kallsyms]  [k] bpf_prog_da4fc6a3f41761a2
     1.49%  map_perf_test  [kernel.kallsyms]  [k] kprobe_ftrace_handler

after:
    60.03%  map_perf_test  [kernel.kallsyms]  [k] __htab_map_lookup_elem
    18.07%  map_perf_test  [kernel.kallsyms]  [k] lookup_elem_raw
     2.91%  map_perf_test  [kernel.kallsyms]  [k] bpf_prog_da4fc6a3f41761a2
     1.94%  map_perf_test  [kernel.kallsyms]  [k] _einittext
     1.90%  map_perf_test  [kernel.kallsyms]  [k] __audit_syscall_exit
     1.72%  map_perf_test  [kernel.kallsyms]  [k] kprobe_ftrace_handler

Notice that bpf_map_lookup_elem() and htab_map_lookup_elem() are trivial
functions, yet they take sizeable amount of cpu time.
htab_map_gen_lookup() removes bpf_map_lookup_elem() and converts
htab_map_lookup_elem() into three BPF insns which causing cpu time
for bpf_prog_da4fc6a3f41761a2() slightly increase.

$ map_perf_test 256
speed of ARRAY bpf_map_lookup_elem() in lookups per second
	w/o JIT		w/JIT
before	97M		174M
after	64M		280M

before:
    37.33%  map_perf_test  [kernel.kallsyms]  [k] array_map_lookup_elem
    13.95%  map_perf_test  [kernel.kallsyms]  [k] bpf_map_lookup_elem
     6.54%  map_perf_test  [kernel.kallsyms]  [k] bpf_prog_da4fc6a3f41761a2
     4.57%  map_perf_test  [kernel.kallsyms]  [k] kprobe_ftrace_handler

after:
    32.86%  map_perf_test  [kernel.kallsyms]  [k] bpf_prog_da4fc6a3f41761a2
     6.54%  map_perf_test  [kernel.kallsyms]  [k] kprobe_ftrace_handler

array_map_gen_lookup() removes calls to array_map_lookup_elem()
and bpf_map_lookup_elem() and replaces them with 7 bpf insns.

The performance without JIT is slower, since executing extra insns
in the interpreter is slower than running native C code,
but with JIT the performance gains are obvious,
since native C->x86 code is replaced with fewer bpf->x86 instructions.
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Acked-by: Daniel Borkmann <daniel@iogearbox.net>
Signed-off-by: David S. Miller <davem@davemloft.net>

Optimize:
bpf_call
  bpf_map_lookup_elem
    map->ops->map_lookup_elem
      htab_map_lookup_elem
        __htab_map_lookup_elem
into:
bpf_call
  __htab_map_lookup_elem

to improve performance of JITed programs.
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Acked-by: Daniel Borkmann <daniel@iogearbox.net>
Signed-off-by: David S. Miller <davem@davemloft.net>

Optimize bpf_call -> bpf_map_lookup_elem() -> array_map_lookup_elem()
into a sequence of bpf instructions.
When JIT is on the sequence of bpf instructions is the sequence
of native cpu instructions with significantly faster performance
than indirect call and two function's prologue/epilogue.
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Acked-by: Daniel Borkmann <daniel@iogearbox.net>
Signed-off-by: David S. Miller <davem@davemloft.net>

convert_ctx_accesses() replaces single bpf instruction with a set of
instructions. Adjust corresponding insn_aux_data while patching.
It's needed to make sure subsequent 'for(all insn)' loops
have matching insn and insn_aux_data.
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Acked-by: Daniel Borkmann <daniel@iogearbox.net>
Signed-off-by: David S. Miller <davem@davemloft.net>

reduce indent and make it iterate over instructions similar to
convert_ctx_accesses(). Also convert hard BUG_ON into soft verifier error.
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Acked-by: Daniel Borkmann <daniel@iogearbox.net>
Signed-off-by: David S. Miller <davem@davemloft.net>

no functional change.
move fixup_bpf_calls() to verifier.c
it's being refactored in the next patch
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Acked-by: Daniel Borkmann <daniel@iogearbox.net>
Signed-off-by: David S. Miller <davem@davemloft.net>

The tcp_tw_recycle was already broken for connections
behind NAT, since the per-destination timestamp is not
monotonically increasing for multiple machines behind
a single destination address.

After the randomization of TCP timestamp offsets
in commit 8a5bd45f6616 (tcp: randomize tcp timestamp offsets
for each connection), the tcp_tw_recycle is broken for all
types of connections for the same reason: the timestamps
received from a single machine is not monotonically increasing,
anymore.

Remove tcp_tw_recycle, since it is not functional. Also, remove
the PAWSPassive SNMP counter since it is only used for
tcp_tw_recycle, and simplify tcp_v4_route_req and tcp_v6_route_req
since the strict argument is only set when tcp_tw_recycle is
enabled.
Signed-off-by: Soheil Hassas Yeganeh <soheil@google.com>
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: Neal Cardwell <ncardwell@google.com>
Signed-off-by: Yuchung Cheng <ycheng@google.com>
Cc: Lutz Vieweg <lvml@5t9.de>
Cc: Florian Westphal <fw@strlen.de>
Signed-off-by: David S. Miller <davem@davemloft.net>

Commit 8a5bd45f6616 (tcp: randomize tcp timestamp offsets for each connection)
randomizes TCP timestamps per connection. After this commit,
there is no guarantee that the timestamps received from the
same destination are monotonically increasing. As a result,
the per-destination timestamp cache in TCP metrics (i.e., tcpm_ts
in struct tcp_metrics_block) is broken and cannot be relied upon.

Remove the per-destination timestamp cache and all related code
paths.

Note that this cache was already broken for caching timestamps of
multiple machines behind a NAT sharing the same address.
Signed-off-by: Soheil Hassas Yeganeh <soheil@google.com>
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: Neal Cardwell <ncardwell@google.com>
Signed-off-by: Yuchung Cheng <ycheng@google.com>
Cc: Lutz Vieweg <lvml@5t9.de>
Cc: Florian Westphal <fw@strlen.de>
Signed-off-by: David S. Miller <davem@davemloft.net>

Shannon Nelson says:

====================
sunvnet: better connection management

These patches remove some problems in handling of carrier state
with the ldmvsw vswitch, remove  an xoff misuse in sunvnet, and
add stats for debug and tracking of point-to-point connections
between the ldom VMs.

v2:
 - added ldmvsw ndo_open to reset the LDC channel
 - updated copyrights
====================
Signed-off-by: David S. Miller <davem@davemloft.net>

The sunvnet netdev is connected to the controlling ldom's vswitch
for network bridging.  However, for higher performance between ldoms,
there also is a channel between each client ldom.  These connections are
represented in the sunvnet driver by a queue for each ldom.  The driver
uses select_queue to tell the stack which queue to use by tracking the mac
addresses on the other end of each port.  When a connected ldom shuts down,
the driver receives an LDC_EVENT_RESET and the port is removed from the
driver, thus a queue with no ldom on the other end will never be selected
for Tx.

The driver was trying to reinforce the "don't use this queue" notion with
netif_tx_stop_queue() and netif_tx_wake_queue(), which really should only
be used to signal a Tx queue is full (aka XOFF).  This misuse of queue
state resulted in NETDEV WATCHDOG messages and lots of unnecessary calls
into the driver's tx_timeout handler.  Simply removing these takes care
of the problem.

Orabug: 25190537
Signed-off-by: Shannon Nelson <shannon.nelson@oracle.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

Make sure multicast packets get counted in the device.

Orabug: 25190537
Signed-off-by: Shannon Nelson <shannon.nelson@oracle.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

Track our used and unused queue indexies correctly.  Otherwise, as ports
dropped out and returned, they all eventually ended up with the same
queue index.

Orabug: 25190537
Signed-off-by: Shannon Nelson <shannon.nelson@oracle.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

In this driver, there is a "port" created for the connection to each of
the other ldoms; a netdev queue is mapped to each port, and they are
collected under a single netdev.  The generic netdev statistics show
us all the traffic in and out of our network device, but don't show
individual queue/port stats.  This patch breaks out the traffic counts
for the individual ports and gives us a little view into the state of
those connections.

Orabug: 25190537
Signed-off-by: Shannon Nelson <shannon.nelson@oracle.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

When an ldom VM is bound, the network vswitch infrastructure is set up for
it, but was being forced 'UP' by the userland switch configuration script.
When 'UP' but not actually connected to a running VM, the ipv6 neighbor
probes fail (not a horrible thing) and start cluttering up the kernel logs.
Funny thing: these are debug messages that never actually show up, but
we do see the net_ratelimited messages that say N callbacks were
suppressed.

This patch defers the netif_carrier_on() until an actual link has been
established with the VM, as indicated by receiving an LDC_EVENT_UP from
the underlying LDC protocol.  Similarly, we take the link down when we
see the LDC_EVENT_RESET.  Now when we see the ndo_open(), we reset the
link to get things talking again.

Orabug: 25525312
Signed-off-by: Shannon Nelson <shannon.nelson@oracle.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

Cut-n-paste enablement of 802.3ad bonding on 25G NICs, which currently
report 0 as their bandwidth.

CC: Jay Vosburgh <j.vosburgh@gmail.com>
CC: Veaceslav Falico <vfalico@gmail.com>
CC: Andy Gospodarek <andy@greyhouse.net>
CC: netdev@vger.kernel.org
Signed-off-by: Jarod Wilson <jarod@redhat.com>
Acked-by: Andy Gospodarek <andy@greyhouse.net>
Signed-off-by: David S. Miller <davem@davemloft.net>

replace comma to semi colons in tcp_westwood_info().
Acked-by: Stephen Hemminger <stephen@networkplumber.org>
Signed-off-by: David S. Miller <davem@davemloft.net>

All IRQs owned by the PF and VF drivers share the same nondescript name
"octeon"; this makes it difficult to setup interrupt affinity.

Change the IRQ names to reflect their specific purpose:

    LiquidIO<id>-<func>-<type>-<queue pair num>

Examples:
    LiquidIO0-pf0-rxtx-3
    LiquidIO1-vf1-rxtx-0
    LiquidIO0-pf0-aux

We cannot use netdev->name for naming the IRQs because:

    1.  Early during init, the PF and VF drivers require interrupts to
        send/receive control data from the NIC firmware; so the PF and VF
        must request IRQs long before the netdev struct is registered.

    2.  The IRQ name can only be specified at the time it is requested.
        It cannot be changed after that.
Signed-off-by: Rick Farrington <ricardo.farrington@cavium.com>
Signed-off-by: Felix Manlunas <felix.manlunas@cavium.com>
Signed-off-by: Satanand Burla <satananda.burla@cavium.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

Remove invalid call to dma_sync_single_for_cpu() because previous DMA
allocation was coherent--not streaming.  Remove code that references fields
in struct list_head; replace it with calls to list_empty() and
list_first_entry().  Also, add comment to clarify complicated if statement.
Signed-off-by: Rick Farrington <ricardo.farrington@cavium.com>
Signed-off-by: Felix Manlunas <felix.manlunas@cavium.com>
Signed-off-by: Derek Chickles <derek.chickles@cavium.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

I recently reported on the netem list that iperf network benchmarks
show unexpected results when a bandwidth throttling rate has been
configured for netem. Specifically:

1) The measured link bandwidth *increases* when a higher delay is added
2) The measured link bandwidth appears higher than the specified limit
3) The measured link bandwidth for the same very slow settings varies significantly across
  machines

The issue can be reproduced by using tc to configure netem with a
512kbit rate and various (none, 1us, 50ms, 100ms, 200ms) delays on a
veth pair between network namespaces, and then using iperf (or any
other network benchmarking tool) to test throughput. Complete detailed
instructions are in the original email chain here:
https://lists.linuxfoundation.org/pipermail/netem/2017-February/001672.html

There appear to be two underlying bugs causing these effects:

- The first issue causes long delays when the rate is slow and no
  delay is configured (e.g., "rate 512kbit"). This is because SKBs are
  not orphaned when no delay is configured, so orphaning does not
  occur until *after* the rate-induced delay has been applied. For
  this reason, adding a tiny delay (e.g., "rate 512kbit delay 1us")
  dramatically increases the measured bandwidth.

- The second issue is that rate-induced delays are not correctly
  applied, allowing SKB delays to occur in parallel. The indended
  approach is to compute the delay for an SKB and to add this delay to
  the end of the current queue. However, the code does not detect
  existing SKBs in the queue due to improperly testing sch->q.qlen,
  which is nonzero even when packets exist only in the
  rbtree. Consequently, new SKBs do not wait for the current queue to
  empty. When packet delays vary significantly (e.g., if packet sizes
  are different), then this also causes unintended reordering.

I modified the code to expect a delay (and orphan the SKB) when a rate
is configured. I also added some defensive tests that correctly find
the latest scheduled delivery time, even if it is (unexpectedly) for a
packet in sch->q. I have tested these changes on the latest kernel
(4.11.0-rc1+) and the iperf / ping test results are as expected.
Signed-off-by: Nik Unger <njunger@uwaterloo.ca>
Signed-off-by: Stephen Hemminger <stephen@networkplumber.org>
Signed-off-by: David S. Miller <davem@davemloft.net>

Or Gerlitz says:

====================
small set of sched cleanups

Just two cleanups -- but for the 2nd one I think we need ack from
Cong Wang to make sure this isn't actually a bug report..

changes from V1:
  - addressed comment from Sergei to use 12 hex digits etc
====================
Signed-off-by: David S. Miller <davem@davemloft.net>