During initialization the driver checks whether the flashed FW image
suits its requirements by checking that it's sufficiently new.
However, there's only a weak backward compatibility scheme that is
actually guaranteed by the FW, so driver must also upper bound the
version to prevent compatibility issues between current driver and some
possible future fw.
Signed-off-by: Yuval Mintz <yuvalm@mellanox.com>
Reviewed-by: Ido Schimmel <idosch@mellanox.com>
Signed-off-by: Jiri Pirko <jiri@mellanox.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Jakub Kicinski says:
====================
nfp: devlink, capabilities extensions and updates
This series starts with an improvement to the usability of the device
memory accessors (CPP transactions). Next few patches are devoted to
fixing the devlink locking. After recent patches for mlxsw the locking
scheme of devlink ops has to be reworked. Following patches improve
NFP code dealing with "representors", and expands the error message
printed when driver has no support for loaded FW.
Second part of the series is focused on vNIC capabilities read from
vNIC control memory (often referred to as "BAR0" for historical reasons).
TLV capability format is established and immediately made use of. The
next patches rework parsing of features for control vNIC which allows
apps to mask out features they don't want enabled.
====================
Signed-off-by: David S. Miller <davem@davemloft.net>
BPF firmware currently exposes IRQ moderation capability.
The driver will make use of it by default, inserting 50 usec
delay to every control message exchange. This cuts the number
of messages per second we can exchange by almost half.
None of the other capabilities make much sense for BPF control
vNIC, either. Disable them all.
Signed-off-by: Jakub Kicinski <jakub.kicinski@netronome.com>
Reviewed-by: Dirk van der Merwe <dirk.vandermerwe@netronome.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Most vNIC capabilities are netdev related. It makes no sense
to initialize them and waste FW resources. Some are even
counter-productive, like IRQ moderation, which will slow
down exchange of control messages.
Add to nfp_app a mask of enabled control vNIC capabilities
for apps to use. Make flower and BPF enable all capabilities
for now. No functional changes.
Signed-off-by: Jakub Kicinski <jakub.kicinski@netronome.com>
Reviewed-by: Dirk van der Merwe <dirk.vandermerwe@netronome.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
nfp_net_init() is a little long and we are about to add more
code to reading capabilties. Move the capability reading,
parsing and validating out. Only actual initialization
will stay in nfp_net_init().
No functional changes.
Signed-off-by: Jakub Kicinski <jakub.kicinski@netronome.com>
Reviewed-by: Dirk van der Merwe <dirk.vandermerwe@netronome.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Allow specifying alternative vNIC mailbox location in TLV caps.
This way we can size the mailbox to the needs and not necessarily
waste 512B of ctrl memory space.
Signed-off-by: Jakub Kicinski <jakub.kicinski@netronome.com>
Reviewed-by: Dirk van der Merwe <dirk.vandermerwe@netronome.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
PCIe island clock frequency is used when converting coalescing
parameters from usecs to NFP timestamps. Most chips don't run
at 1200MHz, allow FW to provide us with the real frequency.
Signed-off-by: Jakub Kicinski <jakub.kicinski@netronome.com>
Reviewed-by: Dirk van der Merwe <dirk.vandermerwe@netronome.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
NFP is entirely programmable, including the PCI data interface.
Using a fixed control BAR layout certainly makes implementations
easier, but require careful considerations when space is allocated.
Once BAR area is allocated to one feature nothing else can use it.
Allocating space statically also requires it to be sized upfront,
which leads to either unnecessary limitation or wastage.
We currently have a 32bit capability word defined which tells drivers
which application FW features are supported. Most of the bits
are exhausted. The same bits are also reused for enabling specific
features. Bulk of capabilities don't have a need for an enable bit,
however, leading to confusion and wastage.
TLVs seems like a better fit for expressing capabilities of applications
running on programmable hardware.
This patch leaves the front of the BAR as is, and declares a TLV
capability start at offset 0x58. Most of the space up to 0x0d90
is already allocated, but the used space can be wrapped with RESERVED
TLVs. E.g.:
Address Type Length
0x0058 RESERVED 0xe00 /* Wrap basic structures */
0x0e5c FEATURE_A 0x004
0x0e64 FEATURE_B 0x004
0x0e6c RESERVED 0x990 /* Wrap qeueue stats */
0x1800 FEATURE_C 0x100
Signed-off-by: Jakub Kicinski <jakub.kicinski@netronome.com>
Reviewed-by: Dirk van der Merwe <dirk.vandermerwe@netronome.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
When driver app matching loaded FW is not found users are faced with:
nfp: failed to find app with ID 0x%02x
This message does not properly explain that matching driver code is
either not built into the driver or the driver is too old.
Signed-off-by: Jakub Kicinski <jakub.kicinski@netronome.com>
Reviewed-by: Dirk van der Merwe <dirk.vandermerwe@netronome.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Representors are grouped in sets by type. Currently the whole
sets are under RCU protection, but individual representor pointers
are not. This causes some inconveniences when representors have
to be destroyed, because we have to allocate new sets to remove
any representors. Protect the individual pointers with RCU.
Signed-off-by: Jakub Kicinski <jakub.kicinski@netronome.com>
Reviewed-by: Dirk van der Merwe <dirk.vandermerwe@netronome.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
The write side of repr tables is always done under pf->lock.
Add a helper to dereference repr table pointers under protection
of that lock.
Signed-off-by: Jakub Kicinski <jakub.kicinski@netronome.com>
Reviewed-by: Dirk van der Merwe <dirk.vandermerwe@netronome.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Devlink used to have two global locks: devlink lock and port lock,
our lock ordering looked like this:
devlink lock -> driver's pf->lock -> devlink port lock
After recent changes port lock was replaced with per-instance
lock. Unfortunately, new per-instance lock is taken on most
operations now. This means we can only grab the pf->lock from
the port split/unsplit ops. Lock ordering looks like this:
devlink lock -> driver's pf->lock -> devlink instance lock
Since we can't take pf->lock from most devlink ops, make sure
nfp_apps are prepared to service them as soon as devlink is
registered. Locking the pf must be pushed down after
nfp_app_init() callback.
The init order looks like this:
nfp_app_init
devlink_register
nfp_app_start
netdev/port_register
As soon as app_init is done nfp_apps must be ready to service
devlink-related callbacks. apps can only register their own
devlink objects from nfp_app_start.
Fixes: 2406e7e546 ("devlink: Add per devlink instance lock")
Signed-off-by: Jakub Kicinski <jakub.kicinski@netronome.com>
Reviewed-by: Dirk van der Merwe <dirk.vandermerwe@netronome.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
NFP app is currently shut down as soon as all the vNICs are gone.
This means we can't depend on the app existing throughout the
lifetime of the device. Free the app only from PCI remove path.
Signed-off-by: Jakub Kicinski <jakub.kicinski@netronome.com>
Reviewed-by: Dirk van der Merwe <dirk.vandermerwe@netronome.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Currently the helpers for accessing 4 or 8 byte values over
the CPP bus return the length of IO on success. If the IO
was short caller has to deal with error handling. The short
IO for 4/8B values is completely impractical. Make the
helpers return an error if full access was not possible.
Fix the few places which are actually dealing with errors
correctly, most call sites already only deal with negative
return codes.
Signed-off-by: Jakub Kicinski <jakub.kicinski@netronome.com>
Reviewed-by: Dirk van der Merwe <dirk.vandermerwe@netronome.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Most of the packets return true for is_last_ethertype_ip, surround it
with likely compiler hint.
Signed-off-by: Gal Pressman <galp@mellanox.com>
Signed-off-by: Saeed Mahameed <saeedm@mellanox.com>
Extend the stats group API to have an update_stats() callback which
will be used to fetch the hardware or software counters data.
Signed-off-by: Kamal Heib <kamalh@mellanox.com>
Signed-off-by: Saeed Mahameed <saeedm@mellanox.com>
Merge the per priority traffic and pfc groups into one group, because
both groups share the same update_stats() callback which will be
introduced in the upcoming patch.
Signed-off-by: Kamal Heib <kamalh@mellanox.com>
Signed-off-by: Saeed Mahameed <saeedm@mellanox.com>
Add per-channel counter ch#_eq_rearm to monitor how many lost interrupt
recovery actions happened upon TX timeouts.
Signed-off-by: Eran Ben Elisha <eranbe@mellanox.com>
Reviewed-by: Tariq Toukan <tariqt@mellanox.com>
Signed-off-by: Saeed Mahameed <saeedm@mellanox.com>
Up until this patch, on every TX timeout we would try to do channels
recovery. However, in case of a lost interrupt for an EQ, the channel
associated to it cannot be recovered if reopened as it would never get
another interrupt on sent/received traffic, and eventually ends up with
another TX timeout (Restarting the EQ is not part of channel recovery).
This patch adds a mechanism for explicitly polling EQ in case of a TX
timeout in order to recover from a lost interrupt. If this is not the
case (no pending EQEs), perform a channels full recovery as usual.
Once a lost EQE is recovered, it triggers the NAPI to run and handle all
pending completions. This will free some budget in the bql (via calling
netdev_tx_completed_queue) or by clearing pending TXWQEs and waking up
the queue. One of the above actions will move the queue to be ready for
transmit again.
Signed-off-by: Eran Ben Elisha <eranbe@mellanox.com>
Reviewed-by: Tariq Toukan <tariqt@mellanox.com>
Signed-off-by: Saeed Mahameed <saeedm@mellanox.com>
When TX timeout occurs, EQ consumer index and irqn can help in debug for
understanding the SW state of EQ. Add them to the logger prints for the
relevant EQ only.
Signed-off-by: Eran Ben Elisha <eranbe@mellanox.com>
Reviewed-by: Tariq Toukan <tariqt@mellanox.com>
Signed-off-by: Saeed Mahameed <saeedm@mellanox.com>
When driver callback for TX timeout is being called, it handles all
stopped xmit queues (not only the ones which their timeout expired).
Add usecs since last transmit to TX timeout logs per send queue in order
to monitor if the queue timeout expired.
Signed-off-by: Eran Ben Elisha <eranbe@mellanox.com>
Reviewed-by: Tariq Toukan <tariqt@mellanox.com>
Signed-off-by: Saeed Mahameed <saeedm@mellanox.com>
For a given hairpin packet buffer size, different queue sizes
(values of log_hairpin_num_packets) determine how the data is broken
to strides on the RQ. Currently the chosen value is set to 64B strides.
Signed-off-by: Or Gerlitz <ogerlitz@mellanox.com>
Signed-off-by: Saeed Mahameed <saeedm@mellanox.com>
Allow to specify the size of the hairpin queues along with the
packet buffer data size from the core setup code.
If the driver doesn't provide this, the FW applies proper value that
matches the provided data size and a FW chosen RQ stride size.
Signed-off-by: Or Gerlitz <ogerlitz@mellanox.com>
Signed-off-by: Saeed Mahameed <saeedm@mellanox.com>
Support RSS for hairpin traffic. We create multiple hairpin RQ/SQ pairs
and RSS TTC table per hairpin instance and steer the related flows
through that table so they are spread between the pairs.
We open one pair per 50Gbs link speed, for all speeds <= 50Gbs, there
is one pair and no RSS while for 100Gbs ports two RSSed pairs.
Signed-off-by: Or Gerlitz <ogerlitz@mellanox.com>
Signed-off-by: Saeed Mahameed <saeedm@mellanox.com>
Enhance the hairpin setup code at the core to support a set of N
(RQ,SQ) pairs. This will be later used by the caller to set RSS
spreading among the different RQs.
Signed-off-by: Or Gerlitz <ogerlitz@mellanox.com>
Signed-off-by: Saeed Mahameed <saeedm@mellanox.com>
This will allow to be able and set TC rule whose steering dest is
RSS TTC steering table.
Signed-off-by: Or Gerlitz <ogerlitz@mellanox.com>
Signed-off-by: Saeed Mahameed <saeedm@mellanox.com>
In order to use RSS for hairpin, we refactor the code that deals with
setup of the TTC steering tables. This is done using an interim ttc
params object that has the flow table attributes, TIR numbers, etc.
Signed-off-by: Or Gerlitz <ogerlitz@mellanox.com>
Signed-off-by: Saeed Mahameed <saeedm@mellanox.com>
As part of the QoS model, on xmit, the HW mandates that all packets going
through a given SQ have the same priority. To align hairpin SQs with that,
we use the priority given as part of the matching for the hairpin hash key.
This ensures that flows/packets mapped to different HW priorities will
go through different hairpin instances. If no priority is given for
matching, we treat that as an 8th priority, this is in order not to
harm cases where priority is specified.
Only the PCP priority trust model is supported.
Signed-off-by: Or Gerlitz <ogerlitz@mellanox.com>
Signed-off-by: Saeed Mahameed <saeedm@mellanox.com>
The peer vhca id spans less bits vs the ifindex and can
well serve for the hairpin hash key, move to use that.
This is a pre-step to put more info into the hairpin hash
key in downstream patch while keeping it at 32 bits.
Signed-off-by: Or Gerlitz <ogerlitz@mellanox.com>
Signed-off-by: Saeed Mahameed <saeedm@mellanox.com>
Yuchung Cheng says:
====================
tcp: do not use RTT from delayed ACKs for min-RTT
This patch set prevents TCP sender from using RTT samples from
(suspected) delayed ACKs as the minimum RTT, to avoid unbounded
over-estimation of the network path delay. This issue is common
when a connection has extended periods of one packet chit-chat
beyond the min RTT filter window. The first patch does that for TCP
general min RTT estimation. The second patch addresses specifically
the BBR congestion control's min RTT filter.
====================
Signed-off-by: David S. Miller <davem@davemloft.net>
A persistent connection may send tiny amount of data (e.g. health-check)
for a long period of time. BBR's windowed min RTT filter may only see
RTT samples from delayed ACKs causing BBR to grossly over-estimate
the path delay depending how much the ACK was delayed at the receiver.
This patch skips RTT samples that are likely coming from delayed ACKs. Note
that it is possible the sender never obtains a valid measure to set the
min RTT. In this case BBR will continue to set cwnd to initial window
which seems fine because the connection is thin stream.
Signed-off-by: Yuchung Cheng <ycheng@google.com>
Acked-by: Neal Cardwell <ncardwell@google.com>
Acked-by: Soheil Hassas Yeganeh <soheil@google.com>
Acked-by: Priyaranjan Jha <priyarjha@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
This patch avoids having TCP sender or congestion control
overestimate the min RTT by orders of magnitude. This happens when
all the samples in the windowed filter are one-packet transfer
like small request and health-check like chit-chat, which is farily
common for applications using persistent connections. This patch
tries to conservatively labels and skip RTT samples obtained from
this type of workload.
Signed-off-by: Yuchung Cheng <ycheng@google.com>
Signed-off-by: Soheil Hassas Yeganeh <soheil@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>
Gen11 removes the Resource Streamer, which frees up a big chunk of
the context image. BSpec indicates 12544 DWORDs (13 pages), plus
one page for PPHWSP.
Please notice that, when looking at the BSpec context image table,
the right filter has to be applied as some rows are excluded for
specific GENs. Also, some rows apply per-subslice (for the
calculation above, we have supposed I915_MAX_SUBSLICES = 8).
v2: Rebase.
v3: Use the right size as per the BSpec.
v4:
- Rebased on top of the default context size (Rodrigo)
- Clarify in the commit message where the subslice calculation
comes from.
v5: s/12538/12544/ (Daniele)
BSpec: 18907
Cc: Rodrigo Vivi <rodrigo.vivi@intel.com>
Acked-by: Ben Widawsky <benjamin.widawsky@intel.com> (older version)
Reviewed-by: Daniele Ceraolo Spurio <daniele.ceraolospurio@intel.com>
Signed-off-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Signed-off-by: Oscar Mateo <oscar.mateo@intel.com>
Signed-off-by: Paulo Zanoni <paulo.r.zanoni@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/1515711307-28979-2-git-send-email-oscar.mateo@intel.com
Letting tipc_poll() dereference a socket's pointer to struct tipc_group
entails a race risk, as the group item may be deleted in a concurrent
tipc_sk_join() or tipc_sk_leave() thread.
We now move the 'open' flag in struct tipc_group to struct tipc_sock,
and let the former retain only a pointer to the moved field. This will
eliminate the race risk.
Reported-by: syzbot+799dafde0286795858ac@syzkaller.appspotmail.com
Signed-off-by: Jon Maloy <jon.maloy@ericsson.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Remove the switch block in l2tp_nl_cmd_session_create() that
checks pseudowire-specific parameters since just L2TP_PWTYPE_ETH and
L2TP_PWTYPE_PPP are currently supported and no actual checks are
performed. Moreover the L2TP_PWTYPE_IP/default case presents a harmless
issue in error handling (break instead of goto out_tunnel)
Signed-off-by: Lorenzo Bianconi <lorenzo.bianconi@redhat.com>
Acked-by: Guillaume Nault <g.nault@alphalink.fr>
Signed-off-by: David S. Miller <davem@davemloft.net>
The kmalloc() failure to allocate memory generates enough information
and doesn't need to be accompanied by another driver print.
Fixes: d69a24e036 ("IB/mlx5: Move IB event processing onto a workqueue")
Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
Reviewed-by: Yuval Shaia <yuval.shaia@oracle.com>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
Now that all callers who care about RoCE addresses have been
converted to use rdma_read_gids() simplify rdma_addr_get_sgid()
to only support real GID addresses.
Callers should only use it for OPA and IB transports.
The now deleted implementation for RoCE has several bugs related to IPv6
support and incorrect/inconsistent 'GID' addresses compared to the CM
paths.
Signed-off-by: Parav Pandit <parav@mellanox.com>
Reviewed-by: Daniel Jurgens <danielj@mellanox.com>
Signed-off-by: Leon Romanovsky <leon@kernel.org>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
Use the newly introduced rdma_read_gids() to read the SGID and DGID for
the connection which returns GID correctly for RoCE transport as well.
rdma_addr_get_dgid() for RoCE for client side connections returns MAC
address, instead of DGID.
rdma_addr_get_sgid() for RoCE doesn't return correct SGID for IPv6 and
when more than one IP address is assigned to the netdevice.
Therefore use transport agnostic rdma_read_gids() API provided by rdma_cm
module.
Signed-off-by: Parav Pandit <parav@mellanox.com>
Reviewed-by: Daniel Jurgens <danielj@mellanox.com>
Acked-by: Santosh Shilimkar <santosh.shilimkar@oracle.com>
Signed-off-by: Leon Romanovsky <leon@kernel.org>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
Make use of rdma_read_gids() API to read SGID and DGID which returns
correct GIDs for RoCE and other transports.
rdma_addr_get_dgid() for RoCE for client side connections returns MAC
address, instead of DGID.
rdma_addr_get_sgid() for RoCE doesn't return correct SGID for IPv6 and
when more than one IP address is assigned to the netdevice.
Therefore use transport agnostic rdma_read_gids() API provided by rdma_cm
module.
Signed-off-by: Parav Pandit <parav@mellanox.com>
Reviewed-by: Daniel Jurgens <danielj@mellanox.com>
Signed-off-by: Leon Romanovsky <leon@kernel.org>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
This patch introduces an API that allows legacy applications to query
GIDs for a rdma_cm_id which is used during connection establishment.
GIDs are stored and created differently for iWarp, IB and RoCE transports.
Therefore rdma_read_gids() returns GID for all the transports hiding
such internal details to caller.
It is usable for client side and server side connections.
In general continued use of GID based addressing outside of IB is
discouraged, so rdma_read_gids() should not be used by any new ULPs.
Signed-off-by: Parav Pandit <parav@mellanox.com>
Reviewed-by: Daniel Jurgens <danielj@mellanox.com>
Signed-off-by: Leon Romanovsky <leon@kernel.org>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
Lorenzo Bianconi says:
====================
l2tp: set l2specific_len based on l2specific_type
Do not rely on l2specific_len value provided by userspace but set sublayer
length according to l2specific_type.
Mark L2TP_ATTR_L2SPEC_LEN attribute as not used
Changes since v2:
- drop the patch related to a fix in the switch default case in
l2tp_nl_cmd_session_create()
- use L2SPECTYPE_NONE as default case in l2tp_get_l2specific_len()
Changes since v1:
- remove l2specific_len parameter
- add sanity check on l2specific_type provided by userspace
====================
Signed-off-by: David S. Miller <davem@davemloft.net>
Remove l2specific_len configuration parameter since now L2-Specific
Sublayer length is computed according to l2specific_type provided by
userspace.
Reviewed-by: Guillaume Nault <g.nault@alphalink.fr>
Tested-by: Guillaume Nault <g.nault@alphalink.fr>
Signed-off-by: Lorenzo Bianconi <lorenzo.bianconi@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Remove l2specific_len dependency while building l2tpv3 header or
parsing the received frame since default L2-Specific Sublayer is
always four bytes long and we don't need to rely on a user supplied
value.
Moreover in l2tp netlink code there are no sanity checks to
enforce the relation between l2specific_len and l2specific_type,
so sending a malformed netlink message is possible to set
l2specific_type to L2TP_L2SPECTYPE_DEFAULT (or even
L2TP_L2SPECTYPE_NONE) and set l2specific_len to a value greater than
4 leaking memory on the wire and sending corrupted frames.
Reviewed-by: Guillaume Nault <g.nault@alphalink.fr>
Tested-by: Guillaume Nault <g.nault@alphalink.fr>
Signed-off-by: Lorenzo Bianconi <lorenzo.bianconi@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
