GCC 7 checks for fall through comments, add the two missing ones.
Signed-off-by: NJakub Kicinski <jakub.kicinski@netronome.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Forgetting to disable preemption around tcf_action_stats_update()
seems to be a common mistake.  Add a helper function for updating
stats on all actions of a filter.
Signed-off-by: NJakub Kicinski <jakub.kicinski@netronome.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

We only kick RX free buffer queue controller every NFP_NET_FL_BATCH
(currently 16) entries.  This means that we will always kick the QC
when write ring index is divisable by NFP_NET_FL_BATCH.  There is
no need to keep counts.
Signed-off-by: NJakub Kicinski <jakub.kicinski@netronome.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Adding ring size to index calculation is pointless, since index
will be masked with ring size - 1.
Suggested-by: NDavid Laight <David.Laight@ACULAB.COM>
Signed-off-by: NJakub Kicinski <jakub.kicinski@netronome.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Ring pointers are unsigned.  Fix the print formats to avoid
showing users negative values.
Signed-off-by: NJakub Kicinski <jakub.kicinski@netronome.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

There is currently no timeout to the resource and lock acquiring
loops.  We printed warnings and depended on user sending a signal
to the waiting process to stop the waiting.  This doesn't work
very well when wait happens out of a work queue.  The simplest
example of that is PCI probe.  When user loads the module and card
is in a broken state modprobe will wait forever and signals sent
to it will not actually reach the probing thread.

Make sure all wait loops have a time out.  Set the upper wait time
to 60 seconds to stay on the safe side.
Signed-off-by: NJakub Kicinski <jakub.kicinski@netronome.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Add support for retrieving temperature and power sensor and limits via NSP.
Signed-off-by: NDavid Brunecz <david.brunecz@netronome.com>
Signed-off-by: NJakub Kicinski <jakub.kicinski@netronome.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

We want to support extendable commands, where newer versions
of the management FW may provide more information.  Zero out
the communication buffer before passing control to NSP.  This
way if management FW is old and only fills in first N bytes,
the remaining ones will be zeros which extended ABI fields
should reserve as not supported/not available.
Signed-off-by: NJakub Kicinski <jakub.kicinski@netronome.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

We currently print reserved BAR mappings info as we create them.
This makes the probe logs longer than necessary.  Print into a
buffer instead and log all the info as a single line.
Signed-off-by: NJakub Kicinski <jakub.kicinski@netronome.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

nfp_cpp_{read,write}() helpers perform device memory mapping (setting
the PCIe -> NOC translation BARs) and accessing it.  They, however,
currently implicitly expect that the length of entire operation will
fit in one BAR translation window.  There is a number of 16MB windows
available, and we don't really need to access such large areas today.

If the user, however, manages to trick the driver into making a big
mapping (e.g. by providing a huge fake FW file), the driver will
print a warning saying "No suitable BAR found for request" and a
stack trace - which most users find concerning.

To be future-proof and not scare users with warnings, make the
nfp_cpp_{read,write}() helpers do accesses chunk by chunk if the area
size is large.  Set the notion of "large" to 2MB, which is the size
of the smallest BAR window.
Signed-off-by: NJakub Kicinski <jakub.kicinski@netronome.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

For accessing PCIe ctrl memory we depend on the BAR aperture being
large enough to reach all registers.  Since the BAR aperture can
be set in the flash make sure the driver won't oops the kernel
when the PCIe configuration is unusual.
Signed-off-by: NJakub Kicinski <jakub.kicinski@netronome.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

If ioremap of PCIe ctrl memory failed we can still get to it through
PCI config space, therefore we allow ioremap() to fail.  When if fails,
however, we must leave all the IOMEM pointers as NULL.  Currently we
would calculate csr and em pointers, adding offsets to the potential
NULL value and therefore making the NULL-checks throughout the code
ineffective.
Signed-off-by: NJakub Kicinski <jakub.kicinski@netronome.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

PCI subsystem has support for drivers limiting the number of VFs
available below what the IOV capability claims.  Make use of it.

While at it remove the #ifdef/#endif on CONFIG_PCI_IOV, it was
there to avoid unnecessary warnings in case device read failed
but kernel doesn't have SR-IOV support anyway.  Device reads
should not fail.

Note that we still need the driver-internal check for the case
where max VFs is 0 since PCI subsystem treats 0 as limit not set.
Signed-off-by: NJakub Kicinski <jakub.kicinski@netronome.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Expose FW app ability to change MAC address at runtime.  Make sure
we only depend on it if FW app advertised the right capability.
Signed-off-by: NPablo Cascón <pablo.cascon@netronome.com>
Reviewed-by: NJakub Kicinski <jakub.kicinski@netronome.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Fix build with DEVLINK=m and NFP=y.

Fixes: 1851f93f ("nfp: add devlink support")
Reported-by: Nkbuild test robot <fengguang.wu@intel.com>
Signed-off-by: NJakub Kicinski <jakub.kicinski@netronome.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Add support for configuring port split with devlink.  Add devlink
callbacks to validate requested config and call NSP helpers.
Getting the right nfp_port structure can be done with simple iteration.
Signed-off-by: NJakub Kicinski <jakub.kicinski@netronome.com>
Reviewed-by: NSimon Horman <simon.horman@netronome.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

For port splitting we will need to know the total number of lanes
in a port.  Calculate that based on eth_table information.
Signed-off-by: NJakub Kicinski <jakub.kicinski@netronome.com>
Reviewed-by: NSimon Horman <simon.horman@netronome.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Extend nfp_port to contain devlink_port structures.  Register the
ports to allow users inspecting device ports.
Signed-off-by: NJakub Kicinski <jakub.kicinski@netronome.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

We will soon have to invoke more clean up for vNICs.
Move the cleanup callbacks into a helper.
Signed-off-by: NJakub Kicinski <jakub.kicinski@netronome.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Add initial devlink support.  This patch simply switches allocation
of per-adapter structure to devlink's priv and register devlink
with empty ops table.  See following patches for implementation
of particular ops.

We should now clear the app pointer on exit, this is how devlink
callbacks will know app is not initialized.
Signed-off-by: NSimon Horman <simon.horman@netronome.com>
Signed-off-by: NJakub Kicinski <jakub.kicinski@netronome.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Move mutex init to main file close to structure allocation.
This will allow mutex to be taken before net code runs (e.g.
from devlink callbacks).  While at it remember to destroy
the mutex.
Signed-off-by: NJakub Kicinski <jakub.kicinski@netronome.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

State of autonegotiation may have changed but is not yet refreshed.
Make sure ethtool respects the NFP_PORT_CHANGED flag when looking
at autoneg.
Signed-off-by: NJakub Kicinski <jakub.kicinski@netronome.com>
Reviewed-by: NSimon Horman <simon.horman@netronome.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

If reading new state of the port failed, mark the port back as CHANGED.
This way next user state request will trigger refresh, which will
hopefully succeed.
Signed-off-by: NJakub Kicinski <jakub.kicinski@netronome.com>
Reviewed-by: NSimon Horman <simon.horman@netronome.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

After port configuration is performed mark it as changed. This
will close a window of time between configuration and async
state refresh which runs from a workqueue where old port state
would be reported.
Signed-off-by: NJakub Kicinski <jakub.kicinski@netronome.com>
Reviewed-by: NSimon Horman <simon.horman@netronome.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Add link to nfp_ports to make it possible to iterate over all ports.
This will come in handy when some ports may be representors.
Signed-off-by: NJakub Kicinski <jakub.kicinski@netronome.com>
Reviewed-by: NSimon Horman <simon.horman@netronome.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Track whether physical port's state have changed since last refresh
inside the nfp_port structure instead of the vNIC structure.
Signed-off-by: NJakub Kicinski <jakub.kicinski@netronome.com>
Reviewed-by: NSimon Horman <simon.horman@netronome.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Always updating port state in place by overriding values in exiting
pf->eth_tbl makes things easier to manage and allows us to have a
common helper for both full and per-port refresh.
Signed-off-by: NJakub Kicinski <jakub.kicinski@netronome.com>
Reviewed-by: NSimon Horman <simon.horman@netronome.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Encapsulate port information into struct nfp_port.  nfp_port will
soon be extended to contain devlink_port information.  It also makes
it easier to reuse port-related code between vNICs and representors.
Signed-off-by: NJakub Kicinski <jakub.kicinski@netronome.com>
Reviewed-by: NSimon Horman <simon.horman@netronome.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

We only support core NIC apps which have vNICs for each physical port/
split and no representors right now.  Enforce that either each vNIC has
a NSP eth_table entry or if NSP port table is not available none do.

One scenario this will prevent from happening is user force-loading
wrong firmware file if FW app requires different firmwares per media
config.

While at it move some code to nfp_net_pf_alloc_vnic() to make it
counter-match nfp_net_pf_free_vnic() better.
Signed-off-by: NJakub Kicinski <jakub.kicinski@netronome.com>
Reviewed-by: NSimon Horman <simon.horman@netronome.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Introduce a concept of an application.  For now it's just grouping
pointers and serving as a layer of indirection.  It will help us
weaken the dependency on nfp_net in ethtool code.  Later series
will flesh out support for different apps in the driver.
Signed-off-by: NJakub Kicinski <jakub.kicinski@netronome.com>
Reviewed-by: NSimon Horman <simon.horman@netronome.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Soon a third place will need to free a struct nfp_net.  Add a free
counterpart to nfp_net_pf_alloc_vnic().
Signed-off-by: NJakub Kicinski <jakub.kicinski@netronome.com>
Reviewed-by: NSimon Horman <simon.horman@netronome.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

vNIC is a PCIe-side abstraction NFP firmwares supported by this
driver use.  It was initially meant to represent a device port
and therefore a netdev but today should be thought of as a way
of grouping descriptor rings and associated state.  Advanced apps
will have vNICs without netdevs and ports without a vNIC (using
representors instead).

Make sure code refers to vNICs as vNICs and not ports or netdevs.
No functional changes.
Signed-off-by: NJakub Kicinski <jakub.kicinski@netronome.com>
Reviewed-by: NSimon Horman <simon.horman@netronome.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

struct nfp_net represents a vNIC, we will be moving away from the
requirement for every vNIC to have a netdev associated with it.
Remove "netdev" from some function names and prefer passing
struct nfp_net pointer as argument instead of struct net_device *.
Signed-off-by: NJakub Kicinski <jakub.kicinski@netronome.com>
Reviewed-by: NSimon Horman <simon.horman@netronome.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Add nfp_cppcore_pcie_unit() helper to retrieve the PCIE unit of a CPP
handle and use the new helper as appropriate.
Signed-off-by: NSimon Horman <simon.horman@netronome.com>
Reviewed-by: NJakub Kicinski <jakub.kicinski@netronome.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Given that our rings are always a power of 2, we can simplify the
calculation of number of completed TX descriptors by using masking
instead of if statement based on whether the index have wrapped
or not.
Signed-off-by: NJakub Kicinski <jakub.kicinski@netronome.com>
Reviewed-by: NSimon Horman <simon.horman@netronome.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

We have a number of places where we calculate the descriptor
index based on a value which may have overflown.  Create a
macro for masking with the ring size.
Signed-off-by: NJakub Kicinski <jakub.kicinski@netronome.com>
Reviewed-by: NSimon Horman <simon.horman@netronome.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Since XDP TX ring holds "spare" RX buffers anyway, we don't have to
rush the completion.  We can wait until ring fills up completely
before trying to reclaim buffers.  If RX poll has ended an no
buffer has been queued for XDP TX we have no guarantee we will see
another interrupt, so run the reclaim there as well, to make sure
TX statistics won't become stale.

This should help us reclaim more buffers per single queue controller
register read.

Note that the XDP completion is very trivial, it only adds up
the sizes of transmitted frames for statistics so the latency
spike should be acceptable.  In case user sets the ring sizes
to something crazy, limit the completion to 2k entries.

The check if the ring is empty at the beginning of xdp_complete()
is no longer needed - the callers will perform it.
Signed-off-by: NJakub Kicinski <jakub.kicinski@netronome.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Introduce NFP_NET_CFG_CTRL_CSUM_COMPLETE capability and implement parsing
of CHECKSUM_COMPLETE metadata.
Signed-off-by: NJakub Kicinski <jakub.kicinski@netronome.com>
Signed-off-by: NEdwin Peer <edwin.peer@netronome.com>
Reviewed-by: NSimon Horman <simon.horman@netronome.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

ABI version 4 introduced metadata chaining. Using the ABI version to signal
metadata chaining precludes firmware that advertises new capabilities which
rely on prepended metadata from working on older kernels.

Capability bits are thus better suited to signalling the chained metadata
format. A new version of the RSS capability is introduced to distinguish
between the differing metadata formats for ABI versions other than 4.
Signed-off-by: NEdwin Peer <edwin.peer@netronome.com>
Reviewed-by: NJakub Kicinski <jakub.kicinski@netronome.com>
Reviewed-by: NSimon Horman <simon.horman@netronome.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Even if capability for RSS and IRQ moderation are present we may
have not initialized them for control vNIC.  Depend on selected
features mask (ctrl) rather than capabilities (cap) to determine
which features should be enabled.
Signed-off-by: NJakub Kicinski <jakub.kicinski@netronome.com>
Reviewed-by: NSimon Horman <simon.horman@netronome.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>