This patch updates the lowest limit for adaptive interrupt interrupt
moderation to roughly 12K interrupts per second.
The way I came about reaching 12K as the desired interrupt rate is by
testing with UDP flows. Specifically I had a simple test that ran a
netperf UDP_STREAM test at varying sizes. What I found was as the packet
sizes increased the performance fell steadily behind until we were only
able to receive at ~4Gb/s with a message size of 65507. A bit of digging
found that we were dropping packets for the socket in the network stack,
and looking at things further what I found was I could solve it by increasing
the interrupt rate, or increasing the rmem_default/rmem_max. What I found was
that when the interrupt coalescing resulted in more data being processed
per interrupt than could be stored in the socket buffer we started losing
packets and the performance dropped. So I reached 12K based on the
following math.
rmem_default = 212992
skb->truesize = 2994
212992 / 2994 = 71.14 packets to fill the buffer
packet rate at 1514 packet size is 812744pps
71.14 / 812744 = 87.9us to fill socket buffer
From there it was just a matter of choosing the interrupt rate and
providing a bit of wiggle room which is why I decided to go with 12K
interrupts per second as that uses a value of 84us.
The data below is based on VM to VM over a direct assigned ixgbe interface.
The test run was:
netperf -H <ip> -t UDP_STREAM"
Socket Message Elapsed Messages CPU Service
Size Size Time Okay Errors Throughput Util Demand
bytes bytes secs # # 10^6bits/sec % SS us/KB
Before:
212992 65507 60.00 1100662 0 9613.4 10.89 0.557
212992 60.00 473474 4135.4 11.27 0.576
After:
212992 65507 60.00 1100413 0 9611.2 10.73 0.549
212992 60.00 974132 8508.3 11.69 0.598
Using bare metal the data is similar but not as dramatic as the throughput
increases from about 8.5Gb/s to 9.5Gb/s.
Signed-off-by: Alexander Duyck <alexander.h.duyck@redhat.com>
Tested-by: Krishneil Singh <krishneil.k.singh@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
This patch will add in the new MAC defines and fit it into the switch
cases throughout the driver. New functionality and enablement support will
be added in following patches.
Signed-off-by: Don Skidmore <donald.c.skidmore@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
When we can't get MSI-X vectors, we disable a few features which require
MSI-X vectors. Print warnings just like we do when disabling DCB.
Signed-off-by: Jacob Keller <jacob.e.keller@intel.com>
Tested-by: Phil Schmitt <phillip.j.schmitt@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
Again, we should not be directly using netif_printk, as we have our own
error print routines that we generate. In addition, instead of using an
early return we can just use the else block of this one line if
statement.
Signed-off-by: Jacob Keller <jacob.e.keller@intel.com>
Tested-by: Phil Schmitt <phillip.j.schmitt@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
In this case, disabling DCB is not an error. We can still function, but
we just have to let the user know. In addition, since we call this
during probe before allocating our netdevice structure, we should use
e_dev_warn instead of e_warn.
Signed-off-by: Jacob Keller <jacob.e.keller@intel.com>
Tested-by: Phil Schmitt <phillip.j.schmitt@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
Our calculated v_budget doesn't matter except if we allocate MSI-X
vectors. We shouldn't need to calculate this outside of the function, so
don't. Instead, only calculate it once we attempt to acquire MSI-X
vectors. This helps collocate all of the MSI-X vector code together.
Signed-off-by: Jacob Keller <jacob.e.keller@intel.com>
Tested-by: Phil Schmitt <phillip.j.schmitt@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
We already have to kfree this value if we fail, and this is only part of
MSI-X mode, so we should simply allocate the value where we need it.
This is cleaner, and makes it a lot more obvious why we are freeing it
inside of ixgbe_acquire_msix_vectors.
Signed-off-by: Jacob Keller <jacob.e.keller@intel.com>
Tested-by: Phil Schmitt <phillip.j.schmitt@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
Similar to how ixgbevf handles acquiring MSI-X vectors, we can return an
error code instead of relying on the flag being set. This makes it more
clear that we have failed to setup MSI-X mode, and also will make it
easier to consolidate MSI-X related code all into the single function.
Signed-off-by: Jacob Keller <jacob.e.keller@intel.com>
Tested-by: Phil Schmitt <phillip.j.schmitt@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
The netif_printk relies on our netdevice structure to be registered
already. We may call ixgbe_acquire_msix_vectors prior to registering our
netdevice, so we should not use the netdevice specific printk.
Signed-off-by: Jacob Keller <jacob.e.keller@intel.com>
Tested-by: Phil Schmitt <phillip.j.schmitt@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
This change addresses several issues in the current ixgbe implementation of
busy poll sockets.
First was the fact that it was possible for frames to be delivered out of
order if they were held in GRO. This is addressed by flushing the GRO buffers
before releasing the q_vector back to the idle state.
The other issue was the fact that we were having to take a spinlock on
changing the state to and from idle. To resolve this I have replaced the
state value with an atomic and use atomic_cmpxchg to change the value from
idle, and a simple atomic set to restore it back to idle after we have
acquired it. This allows us to only use a locked operation on acquiring the
vector without a need for a locked operation to release it.
Signed-off-by: Alexander Duyck <alexander.h.duyck@intel.com>
Tested-by: Phil Schmitt <phillip.j.schmitt@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
Since we previously called ixgbe_set_num_queues just prior to attempting
to set our interrupt scheme, it may be non obvious why we have to call
it again inside the function. Add a comment which helps make it more
obvious that we are resetting features based on the fact that we do not
have MSI-X enabled, and cannot use the previous settings.
Signed-off-by: Jacob Keller <jacob.e.keller@intel.com>
Tested-by: Phil Schmitt <phillip.j.schmitt@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
This patch fixes various log strings that are split over multiple lines
in the ixgbe driver. This cleans up checkpatch.pl warnings, and makes it
easier to search the code for warning strings displayed to the kernel
log.
Signed-off-by: Jacob Keller <jacob.e.keller@intel.com>
Tested-by: Phil Schmitt <phillip.j.schmitt@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
This patch updates the contact information on the ixgbe driver files so
that every file includes the Linux NICS address, as it is still used,
but only a few of the files mentioned it.
Signed-off-by: Jacob Keller <jacob.e.keller@intel.com>
Tested-by: Phil Schmitt <phillip.j.schmitt@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
As result of deprecation of MSI-X/MSI enablement functions
pci_enable_msix() and pci_enable_msi_block() all drivers
using these two interfaces need to be updated to use the
new pci_enable_msi_range() and pci_enable_msix_range()
interfaces.
Signed-off-by: Alexander Gordeev <agordeev@redhat.com>
Cc: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
Cc: Jesse Brandeburg <jesse.brandeburg@intel.com>
Cc: Bruce Allan <bruce.w.allan@intel.com>
Cc: e1000-devel@lists.sourceforge.net
Cc: netdev@vger.kernel.org
Cc: linux-pci@vger.kernel.org
Signed-off-by: David S. Miller <davem@davemloft.net>
Now that l2 acceleration ops are in place from the prior patch,
enable ixgbe to take advantage of these operations. Allow it to
allocate queues for a macvlan so that when we transmit a frame,
we can do the switching in hardware inside the ixgbe card, rather
than in software.
Signed-off-by: John Fastabend <john.r.fastabend@intel.com>
Signed-off-by: Neil Horman <nhorman@tuxdriver.com>
CC: Andy Gospodarek <andy@greyhouse.net>
CC: "David S. Miller" <davem@davemloft.net>
Signed-off-by: David S. Miller <davem@davemloft.net>
Add the ixgbe driver code implementing ndo_ll_poll.
Adds ndo_ll_poll method and locking between it and the napi poll.
When receiving a packet we use skb_mark_ll to record the napi it came from.
Add each napi to the napi_hash right after netif_napi_add().
Signed-off-by: Alexander Duyck <alexander.h.duyck@intel.com>
Signed-off-by: Jesse Brandeburg <jesse.brandeburg@intel.com>
Signed-off-by: Eliezer Tamir <eliezer.tamir@linux.intel.com>
Reviewed-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
This change adds support for ixgbe to configure the XPS queue mapping on
load. The result of this change is that on open we will now be resetting
the number of Tx queues, and then setting the default configuration for XPS
based on if ATR is enabled or disabled.
Signed-off-by: Alexander Duyck <alexander.h.duyck@intel.com>
Reviewed-by: John Fastabend <john.r.fastabend@intel.com>
Tested-by: Phil Schmitt <phillip.j.schmitt@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
Instead of adjusting the FCoE and Flow director limits based on the number
of CPUs we can define them much sooner. This allows the user to come
through later and adjust them once we have updated the code to support the
set_channels ethtool operation.
I am still allowing for FCoE and RSS queues to be separated if the number
queues is less than the number of CPUs. This essentially treats the two
groupings like they are two separate traffic classes.
In addition I am changing the initialization to use the MAX_TX/RX_QUEUES
defines instead of trying to compute the value as it will be possible in
upcoming patches for the user to request the maximum number of queues.
I have also updated things so that the upper limit on queues is exactly 63
instead of allowing it to go up to 64. The reason for this change is to
address the fact thqt the driver only supports up to 63 queue vectors since
the hardware supports 64 MSI-X vectors, but one must be reserved for "other"
causes.
Signed-off-by: Alexander Duyck <alexander.h.duyck@intel.com>
Tested-by: Phil Schmitt <phillip.j.schmitt@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
Signed-off-by: Don Skidmore <donald.c.skidmore@intel.com>
Tested-by: Phil Schmitt <phillip.j.schmitt@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
The q_vector->itr check in ixgbe_configure_tx_ring() was done prior to it
being set, which resulted in TXDCTL.WTHRESH always being set to 1 on driver
load, while consequent resets would set it to 8.
This patch moves the setting of q_vector->itr in ixgbe_alloc_q_vector() to
make sure that TXDCTL.WTHRESH is set to 8 by default.
Signed-off-by: Emil Tantilov <emil.s.tantilov@intel.com>
Tested-by: Phil Schmitt <phillip.j.schmitt@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
When enabling DCB the rings belonging to a q_vector on CPU 0 were not
reinitializing their DCA registers. Upon closer inspection the issue was
that the q_vector CPU variable was left at 0 resulting in the driver not
updating the DCA registers.
In order to guarantee the DCA registers will be updated I am adding a
couple line change so that we initialize the CPU variable to -1 which will
force a DCA update the first time an interrupt fires on that q_vector.
In addition we were setting the CPU affinity hint to all CPUs when we were
not specifying a CPU. Instead we should leave it as all zeros to avoid any
possible confusion about the fact that we shouldn't be giving a hint.
Signed-off-by: Alexander Duyck <alexander.h.duyck@intel.com>
Tested-by: Phil Schmitt <phillip.j.schmitt@intel.com>
Tested-by: Ross Brattain <ross.b.brattain@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
This change makes it so that we can use 1TC DCB in the case of MSI and
legacy interrupts. The advantage to this is that it allows us to fully
support FCoE w/ DCB instead of having to drop to link flow control only
when using these interrupt modes.
Cc: John Fastabend <john.r.fastabend@intel.com>
Signed-off-by: Alexander Duyck <alexander.h.duyck@intel.com>
Tested-by: Ross Brattain <ross.b.brattain@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
This change makes it so that we can use the atr_sample_rate to determine if
we are capable of supporting ATR. The advantage to this approach is that it
allows us to now determine the setting of the IXGBE_FLAG_FDIR_HASH_CAPABLE
based on the queueing scheme, instead of the queueing scheme being based on
the flag.
Using this approach there are essentially 5 conditions that must be checked
prior to trying to enable ATR:
1. Is SR-IOV disabled?
2. Are the number of TCs <= 1?
3. Is RSS queueing limit greater than 1?
4. Is atr_sample_rate set?
5. Is Flow Director perfect filtering disabled?
If any of these conditions are enabled they should disable ATR filtering.
Note that in the case of conditions 1 through 4 being met we will set
things up for ATR queueing, however if test 5 fails we will still leave the
queues allocated for use by perfect filters. The reason for this is to
allow for us to switch back and forth between ntuple and ATR without
needing to reallocate the descriptor rings.
Signed-off-by: Alexander Duyck <alexander.h.duyck@intel.com>
Tested-by: Phil Schmitt <phillip.j.schmitt@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
This is meant to fix a bug in which we were not checking for pre-existing
VFs if we were not setting the max_vfs value at driver load. What happens
now is that we always call ixgbe_enable_sriov and this checks for
pre-existing VFs ore requested VFs prior to deciding on no SR-IOV.
Signed-off-by: Alexander Duyck <alexander.h.duyck@intel.com>
Tested-by: Phil Schmitt <phillip.j.schmitt@intel.com>
Tested-by: Sibai Li <sibai.li@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
All of our hardware supports RSS even if it is only for a single queue. So
instead of toting around the RSS enable flag I am updating the code so that
all devices are enabled and if we want to disable RSS it is indicated via
the RSS mask.
Signed-off-by: Alexander Duyck <alexander.h.duyck@intel.com>
Tested-by: Phil Schmitt <phillip.j.schmitt@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
This change essentially makes it so that we can enable almost all of the
features all at once. This patch allows for the combination of SR-IOV,
DCB, and FCoE in the case of the x540. It also beefs up the SR-IOV by
adding support for RSS to the PF.
The testing matrix gets to be very complex for this patch as there are a
number of different features and subsets for queueing options. I tried to
narrow these down a bit by restricting the PF to only supporting 4TC DCB
when it is enabled in addition to SR-IOV.
Cc: Greg Rose <gregory.v.rose@intel.com>
Cc: John Fastabend <john.r.fastabend@intel.com>
Signed-off-by: Alexander Duyck <alexander.h.duyck@intel.com>
Tested-by: Phil Schmitt <phillip.j.schmitt@intel.com>
Tested-by: Ross Brattain <ross.b.brattain@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
This change cleans up some of the logic in an attempt to try and simplify
things for how we are configuring DCB w/ RSS.
In this patch I basically did 3 things. I updated the logic for getting
the first register index. I applied the fact that all TCs get the same
number of queues to simplify the looping logic in caching the DCB ring
register. Finally I updated how we configure the RQTC register to match
the fact that all TCs are assigned the same number of queues.
Cc: John Fastabend <john.r.fastabend@intel.com>
Signed-off-by: Alexander Duyck <alexander.h.duyck@intel.com>
Tested-by: Phil Schmitt <phillip.j.schmitt@intel.com>
Tested-by: Ross Brattain <ross.b.brattain@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
It makes much more sense for us to configure the real number of Tx and Rx
queues in the ixgbe_open call than it does in ixgbe_set_num_queues. By
setting the number in ixgbe_open we can avoid a number of unecessary
updates and only have to make the calls once.
Signed-off-by: Alexander Duyck <alexander.h.duyck@intel.com>
Tested-by: Phil Schmitt <phillip.j.schmitt@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
This change merges the ixgbe_cache_ring_fcoe and ixgbe_set_fcoe_queues
logic into the DCB and RSS initialization calls.
Cc: John Fastabend <john.r.fastabend@intel.com>
Signed-off-by: Alexander Duyck <alexander.h.duyck@intel.com>
Tested-by: Phil Schmitt <phillip.j.schmitt@intel.com>
Tested-by: Ross Brattain <ross.b.brattain@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
In upcoming patches it will become increasingly common to need to determine
the FCoE traffic class in order to determine the correct queues for FCoE.
In order to make this easier I am adding a function for obtaining the FCoE
traffic class based on the user priority.
Cc: John Fastabend <john.r.fastabend@intel.com>
Signed-off-by: Alexander Duyck <alexander.h.duyck@intel.com>
Tested-by: Phil Schmitt <phillip.j.schmitt@intel.com>
Tested-by: Ross Brattain <ross.b.brattain@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
There are really only 3 modes that can control the number of queues. Those
are RSS, DCB, and VMDq/SR-IOV. Currently we have things much more broken
up than they need to be for how we are configuring the rings. In order to
try and straiten some of this out I am going to start merging similar
functionality into single functions. To start with I am merging the Flow
Director ring configuration into the RSS ring configuration since Flow
Director cannot function with DCB or SR-IOV.
Signed-off-by: Alexander Duyck <alexander.h.duyck@intel.com>
Tested-by: Phil Schmitt <phillip.j.schmitt@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
The mask value for ring features was overloaded for FCoE which can lead to
some confusion. In order to avoid any confusion I am splitting the mask
value and adding an offset value. This can be used for the start of the
FCoE rings, and in the future I hope to use it to store the start of the
registers for SR-IOV.
Signed-off-by: Alexander Duyck <alexander.h.duyck@intel.com>
Tested-by: Phil Schmitt <phillip.j.schmitt@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
We are currently using indices to indicate the upper limit on a ring
feature. However since we can switch back and forth on features such as
DCB and that has effects on other features such as RSS it is preferable to
instead store the upper limit separate from the current value for the
number of rings related to the feature.
Signed-off-by: Alexander Duyck <alexander.h.duyck@intel.com>
Tested-by: Phil Schmitt <phillip.j.schmitt@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
It makes much more sense for us to count q_vectors instead of MSI-X
vectors. We were using num_msix_vectors to find the number of q_vectors in
multiple places. This was wasteful since we only had one place that
actually needs the number of MSI-X vectors and that is in slow path.
Signed-off-by: Alexander Duyck <alexander.h.duyck@intel.com>
Tested-by: Phil Schmitt <phillip.j.schmitt@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
Fix incorrect start markers, wrapped summary lines, missing section
breaks, incorrect separators, and some name mismatches. Delete
a few that are content-free.
Signed-off-by: Ben Hutchings <bhutchings@solarflare.com>
Acked-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
FCoE target mode was experiencing issues due to the fact that we were
sending up data frames that were padded to 60 bytes after the DDP logic had
already stripped the frame down to 52 or 56 depending on the use of VLANs.
This was resulting in the FCoE DDP logic having issues since it thought the
frame still had data in it due to the padding.
To resolve this, adding code so that we do not pad FCoE frames prior to
handling them to the stack.
CC: <stable@vger.kernel.org>
Signed-off-by: Alexander Duyck <alexander.h.duyck@intel.com>
Tested-by: Phil Schmitt <phillip.j.schmitt@intel.com>
Tested-by: Ross Brattain <ross.b.brattain@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
This change reorders the mapping of rings to q_vectors in the case that the
number of rings exceeds the number of q_vectors. Previously we would
allocate the first R/N queues to the first q_vector where R is the number
of rings and N is the number of q_vectors. Instead of doing this we can do
a better job of interleaving the rings to the CPUs by assigning every Nth
ring to the q_vector.
The below tables illustrate this change for the R = 16 N = 4 case.
Before patch After patch
q_vector: 0 1 2 3 0 1 2 3
Rings: 0 4 8 12 0 1 2 3
1 5 9 13 4 5 6 7
3 6 10 14 8 9 10 11
4 7 11 15 12 13 14 15
This should improve the performance for both DCB or ATR when the number of
rings exceeds the number of q_vectors allocated by the adapter.
Signed-off-by: Alexander Duyck <alexander.h.duyck@intel.com>
Tested-by: Ross Brattain <ross.b.brattain@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
This patch makes it so that we identify FCoE rings earlier than
ixgbe_set_rx_buffer_len. Instead we identify the Rx FCoE rings at
allocation time in ixgbe_alloc_q_vector.
The motivation behind this change is to avoid memory corruption when FCoE
is enabled. Without this change we were initializing the rings at 0, and
2K on systems with 4K pages, then when we bumped the buffer size to 4K with
order 1 pages we were accessing offsets 2K and 6K instead of 0 and 4K.
This was resulting in memory corruptions.
Signed-off-by: Alexander Duyck <alexander.h.duyck@intel.com>
Acked-by: Yi Zou <yi.zou@intel.com>
Tested-by: Ross Brattain <ross.b.brattain@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
Resolve namespace issues when FCoE or DCB is not enabled.
The issue is with certain configurations we end up with namespace
problems. A simple example:
ixgbe_main.c
- defines func A()
- uses func A()
ixgbe_fcoe.c
- uses func A()
ixgbe.h
- has prototype for func A()
For default (FCoE included) all is good. But when it isn't the namespace
checker complains about how func A() could be static.
To resolve this, created a ixgbe_lib file to contain functions used
by DCB/FCoE and their helper functions so that they are always in
namespace whether or not DCB/FCoE is enabled.
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
Tested-by: Phil Schmitt <phillip.j.schmitt@intel.com>
