Prior to this patch only UIO/VFIO interrupt handlers types were supported.
This patch adds support for the external interrupt handler type, allowing
external drivers to set their own fds with specific interrupt handlers.
Signed-off-by: Shahaf Shuler <shahafs@mellanox.com>
Acked-by: Yongseok Koh <yskoh@mellanox.com>
This commit adds a signal-based trigger to the Rx burst function in order
to avoid unnecessary system calls while Rx queues are empty.
Triggered Rx bursts put less pressure on the kernel, free up CPU resources
for applications and result in a noticeable performance improvement when
sharing CPU threads with other PMDs.
Measuring the traffic forwarding rate between two physical devices in
testpmd (IO mode, single thread, 64B packets) before and after adding two
tap PMD instances (4 ports total) that do not process any traffic and
comparing results yields:
Without Rx trigger:
-15% (--burst=32)
-62% (--burst=1)
With Rx trigger:
-0.3% (--burst=32)
-6% (--burst=1)
Signed-off-by: Adrien Mazarguil <adrien.mazarguil@6wind.com>
Acked-by: Pascal Mazon <pascal.mazon@6wind.com>
Acked-by: Keith Wiles <keith.wiles@intel.com>
Polling the Tx queue file descriptor before writing to it is not mandatory
since it is configured as non-blocking.
Signed-off-by: Adrien Mazarguil <adrien.mazarguil@6wind.com>
Acked-by: Pascal Mazon <pascal.mazon@6wind.com>
Acked-by: Keith Wiles <keith.wiles@intel.com>
nicvf HW expects the DMA address of the packet data to be
aligned with cache line size.
Packet data offset is a function of struct mbuf size,
mbuf private size and headroom. mbuf private size can
be changed from the application in pool creation, this
check detects HW alignment requirement constraint in pmd
start function.
Signed-off-by: Jerin Jacob <jerin.jacob@caviumnetworks.com>
Acked-by: Hemant Agrawal <hemant.agrawal@nxp.com>
The hardware offload capabilities are not being advertised for the EM PMD.
Because of this, applications that only enable these features if the device
advertises them will never do so.
Normally this is not an issue since normal packet processing should work
even if hardware offload is not available. But, in older versions of
Virtual Box the e1000 device emulation (Intel PRO/1000 MT Desktop 82540EM)
assumes that it should enable VLAN stripping even if the driver does not
request it. This means that any ingress packets that have a VLAN tag will
be stripped. Since the application did not request to enable VLAN
stripping it is not expecting these packets so they are not processed as
VLAN packets.
Regardless of the Virtual Box issue, the driver should be advertising
supported capabilities as is done in other drivers.
Signed-off-by: Allain Legacy <allain.legacy@windriver.com>
Acked-by: Wenzhuo Lu <wenzhuo.lu@intel.com>
According to 'libefx' API requirements, one is allowed to
apply multicast address list to the port in started state
only, otherwise the new array should be copied to a local
storage in order to be applied during the next port start
Coverity issue: 141296
Fixes: 0fa0070e43 ("net/sfc: support multicast addresses list controls")
Fixes: e9ddf37a50 ("net/sfc: fix setting empty multicast list")
Cc: stable@dpdk.org
Signed-off-by: Ivan Malov <ivan.malov@oktetlabs.ru>
Signed-off-by: Andrew Rybchenko <arybchenko@solarflare.com>
Reviewed-by: Andrew Lee <alee@solarflare.com>
Reviewed-by: Andy Moreton <amoreton@solarflare.com>
Support individual/group destination address match (unknown unicast
and all-multicast correspondingly in terms of firmware).
Signed-off-by: Roman Zhukov <roman.zhukov@oktetlabs.ru>
Signed-off-by: Andrew Rybchenko <arybchenko@solarflare.com>
Reviewed-by: Andrew Lee <alee@solarflare.com>
Exact match of source and destination ports is supported by parser.
IP protocol match is enforced to UDP.
Signed-off-by: Roman Zhukov <roman.zhukov@oktetlabs.ru>
Signed-off-by: Andrew Rybchenko <arybchenko@solarflare.com>
Reviewed-by: Andrew Lee <alee@solarflare.com>
Exact match of source and destination ports is supported by parser.
IP protocol match is enforced to TCP.
Signed-off-by: Roman Zhukov <roman.zhukov@oktetlabs.ru>
Signed-off-by: Andrew Rybchenko <arybchenko@solarflare.com>
Reviewed-by: Andrew Lee <alee@solarflare.com>
Exact match of IP protocol, source and destination
addresses is supported by parser.
EtherType match is enforced to IPv6 EtherType.
Signed-off-by: Roman Zhukov <roman.zhukov@oktetlabs.ru>
Signed-off-by: Andrew Rybchenko <arybchenko@solarflare.com>
Reviewed-by: Andrew Lee <alee@solarflare.com>
Exact match of IP protocol, source and destination
addresses is supported by parser.
EtherType match is enforced to IPv4 EtherType.
Signed-off-by: Roman Zhukov <roman.zhukov@oktetlabs.ru>
Signed-off-by: Andrew Rybchenko <arybchenko@solarflare.com>
Reviewed-by: Andrew Lee <alee@solarflare.com>
Exact match of VLAN ID bits is supported only and required in VLAN item.
Mask to match VLAN ID bits only is required, default mask to match entire
TCI is not supported.
Signed-off-by: Roman Zhukov <roman.zhukov@oktetlabs.ru>
Signed-off-by: Andrew Rybchenko <arybchenko@solarflare.com>
Reviewed-by: Andrew Lee <alee@solarflare.com>
Only pattern items VOID, ETH and actions VOID, QUEUE is now
supported.
Signed-off-by: Roman Zhukov <roman.zhukov@oktetlabs.ru>
Signed-off-by: Andrew Rybchenko <arybchenko@solarflare.com>
The information is obtained from firmware on attach. It may
change after MC reboot (firmware version or variant change).
Cache should be refreshed after MC reboot when it is handled
properly (not yet).
Signed-off-by: Andrew Rybchenko <arybchenko@solarflare.com>
Reviewed-by: Andrew Lee <alee@solarflare.com>
Reviewed-by: Andy Moreton <amoreton@solarflare.com>
Reviewed-by: Robert Stonehouse <rstonehouse@solarflare.com>
Just log error for all filter types and return no support indication.
Signed-off-by: Andrew Rybchenko <arybchenko@solarflare.com>
Reviewed-by: Andrew Lee <alee@solarflare.com>
Reviewed-by: Andy Moreton <amoreton@solarflare.com>
Reviewed-by: Robert Stonehouse <rstonehouse@solarflare.com>
The previous API had various problems, including the length of the
caller provided buffer not being specified, no means being available
to discover how big the buffer needs to be, and a lack of clarity of
what the resulting list contains.
To improve it:
- add the buffer length as a parameter
- if the provided buffer is too short, fail with ENOSPC and return
the required length
- ensure that the list contents are valid and add comments describing it
It is safe to change this API as, unsurprisingly, it has no users.
Signed-off-by: Mark Spender <mspender@solarflare.com>
Signed-off-by: Andrew Rybchenko <arybchenko@solarflare.com>
The flag EFX_FILTER_MATCH_LOC_MAC_IG to represent filtering on the
individual/group bit of the MAC address (with the two cases being
distinguished by the MAC address in the filter specification) was
introduced to mirror the Linux driver filtering code, but the
implementations are different enough anyway that it isn't of much
value.
Having separate flags for unknown unicast and multicast simplifies
the code and allows the set of flags to match those used by MCDI.
It will also makes it easier to report whether these filters are
supported.
In the MCDI definitions, the unknown multicast and unicast flags
have the values 0x40000000 and 0x80000000 respectively, and so using
the same values for simplicity requires 32 bits in the filter
specification to store the flags. This means the structure is now
a little bigger than 64 bytes, but filters are not often used on
critical paths so this shouldn't have much impact - on Linux they
are also bigger than they used to be.
Signed-off-by: Mark Spender <mspender@solarflare.com>
Signed-off-by: Andrew Rybchenko <arybchenko@solarflare.com>
Without setting the id, calling xstats_get twice with same array causes
memory corruption.
Also, if IXGBEVF_NB_XSTATS will be different than 1 in the future,
this will cause issues.
Fixes: 156712ba40 ("ixgbevf: add extending stats")
Signed-off-by: Ido Barnea <ibarnea@cisco.com>
Since PKT_TX_TCP_SEG implies PKT_TX_TCP_CKSUM, the PMD must force this
flag.
The fix applied for both tunneled and non-tunneled packets.
Fixes: 3f13f8c23a ("net/mlx5: support hardware TSO")
Fixes: b247f34601 ("net/mlx5: support hardware TSO for VXLAN and GRE")
Signed-off-by: Shahaf Shuler <shahafs@mellanox.com>
Acked-by: Nelio Laranjeiro <nelio.laranjeiro@6wind.com>
Multiple IES API resets can cause a race condition where the mailbox
interrupt request bits can be cleared before being handled. This can
leave certain mailbox messages from the PF to be untreated and the PF
will enter in some inactive state. If this situation occurs, the IES API
will initiate a mailbox version reset which, then, trigger a mailbox
state change. Once this mailbox transition occurs (from OPEN to CONNECT
state), a request for reset will be returned.
Signed-off-by: Qi Zhang <qi.z.zhang@intel.com>
Other shared code bases are planning on using
IS_MULTICAST_ETHER_ADDR and friends without leaving the driver
name in the macro.
Remove reference to FM10K here so that we can re-use the specific
compat flags from Linux.
Signed-off-by: Qi Zhang <qi.z.zhang@intel.com>
Avoid potential FUM fault errors on a VF when updating MAC address
and VLAN information. Only use the register flow when the mailbox is
disconnected, by checking if the enqueue_tx returns
FM10K_MBX_ERR_NO_MBX. If the mailbox message can be sent, there is no
reason to bother with the register writes which are only intended to
be used during VF driver initialization.
Signed-off-by: Qi Zhang <qi.z.zhang@intel.com>
Don't report FM10K_ERR_REQUESTS_PENDING when we fail to disable queues
within the timeout. This can occur due to a hardware Tx hang, or when
the switch ethernet fabric is resetting while we are transmitting
traffic. It can sometimes take up to 500ms before the Tx DMA engine
gives up. Instead, just skip the DMA engine check and perform
a data-path reset anyways. Add a statistic counter to keep track of the
number of resets occurring while we have pending DMA on the rings.
In order to prevent having to assign err = FM10K_SUCCESS, re-order the
last few items of the reset_hw_pf function so that we don't perform
"return err" at the end.
Signed-off-by: Qi Zhang <qi.z.zhang@intel.com>
If the fm10k interface is brought up, but the switch manager software is
not running, the driver will continuously request the lport map every
few seconds in the base driver watchdog routine. Eventually after
several minutes the switch mailbox Tx fifo will fill up and the mailbox
will timeout, resulting in a reset. This reset will appear as if for no
reason, and occurs regularly every few minutes until the switch manager
software is loaded.
Signed-off-by: Qi Zhang <qi.z.zhang@intel.com>
The VF uses a multi-bit update request to clear unused VLANs whenever it
resets. However, an accident in a previous refactor broke multi-bit
updates for VFs, due to misreading a comment in fm10k_vf.c and
attempting to reduce code duplication. The problem occurs because
a multi-bit request has a non-zero length, and the PF would simply drop
any request with the upper 16 bits set. In addition, a multi-bit vlan
update does not have a concept for "VLAN 0" as the single bit update
does.
A previous revision of this patch resolved the issue by simply removing
the upper 16 bit check and the iov_select_vid checks. However, this would
remove the checks for default VID and for ensuring no other VLANs can be
enabled except pf_vid when it has been set. To resolve that issue, this
revision uses the iov_select_vid when we have a single-bit update, and
denies any multi-bit update when the VLAN was administratively set by
the PF. This should be ok since the PF properly updates VLAN_TABLE when
it assigns the PF vid. This ensures that requests to add or "remove" the
PF vid work as expected, but a rogue VF could not use the multi-bit
update as a loophole to attempt receiving traffic on other VLANs.
Signed-off-by: Qi Zhang <qi.z.zhang@intel.com>
The original comment may be read incorrectly as referring to checking
the *entire* length is zero. However, it merely checks only the reserved
bits of both length and reserved in a small amount of code. Update the
comment to indicate this is a clever trick and clearly spell out that it
only checks the reserve bits.
Signed-off-by: Qi Zhang <qi.z.zhang@intel.com>
Use a new #define FM10K_VLAN_OVERRIDE even though we're using
the exact same bit. The reason for this is clarity in the code,
otherwise you can read FM10K_VLAN_CLEAR and think it should be
removed. Also add a comment explaining why the FM10K_VLAN_OVERRIDE
bit is set.
Signed-off-by: Qi Zhang <qi.z.zhang@intel.com>
The diagram represents bit layout of the multi-bit VLAN update
message format. Re-draw the numbers using base 8, and mark the
bit values every 8 bits at the top. This should make it more
easy to grasp the table quickly.
Signed-off-by: Qi Zhang <qi.z.zhang@intel.com>
Add FM10K_PF_ATTR_ID_ERR, since it is possible for the switch manager
to send out an error message indicating status of the LPORT_MAP due to
zero allocated bandwidth.
Signed-off-by: Qi Zhang <qi.z.zhang@intel.com>
Clean up the logic in fm10k_tlv_attr_parse, we
should not reply on that FM10K_NOT_IMPLEMENTED is
greater than zero, as this can easily cause confusion.
The patch also correct a minor document error.
Signed-off-by: Qi Zhang <qi.z.zhang@intel.com>
Deleting lport when multicast mod is configured to
FM10K_XCAST_MODE_ALLMULTI or FM10K_XCAST_MODE_PROMISC will
result in generating orphaned multicast-group entries in the
switch manager.
Before deleting the lport, reset multicast mode to
FM10K_XCAST_MODE_NONE to flush out these multicast-group
entries.
Signed-off-by: Qi Zhang <qi.z.zhang@intel.com>
Store the device name in dev->data->name, to have symmetrical behavior
between rte_pmd_tap_probe(name) and rte_pmd_tap_remove(name).
The netdevice name (linux interface name) is stored in the name field of
struct pmd_internals.
snprintf(data->name) has been moved closer to the rte_ethdev_allocate()
as it should use the same name.
Fixes: 02f96a0a82 ("net/tap: add TUN/TAP device PMD")
Signed-off-by: Pascal Mazon <pascal.mazon@6wind.com>
Acked-by: Keith Wiles <keith.wiles@intel.com>
EF10 supported by the PMD has no limitations on address boundary
crossing by Tx DMA descriptors.
Fixes: 428c7ddd2f ("net/sfc: send bursts of packets")
Fixes: fec33d5bb3 ("net/sfc: support firmware-assisted TSO")
Signed-off-by: Andrew Rybchenko <arybchenko@solarflare.com>
Siena has limitation on maximum byte count and 4k boundary crosssing
(which is stricter than maximum byte count).
EF10 has limitation on maximum byte count only.
Fixes: f7dc06bf35 ("net/sfc/base: import 5xxx/6xxx family support")
Fixes: e7cd430c86 ("net/sfc/base: import SFN7xxx family support")
Fixes: 94190e3543 ("net/sfc/base: import SFN8xxx family support")
Signed-off-by: Andrew Rybchenko <arybchenko@solarflare.com>
If a libefx-based driver needs some way to clear port statistics,
then an MCDI agnostic method is required.
Signed-off-by: Ivan Malov <ivan.malov@oktetlabs.ru>
Signed-off-by: Andrew Rybchenko <arybchenko@solarflare.com>
All kinds of filter need to hardware mac type check
to make sure the hardware support that type of filter.
If not, it may cause serious issue.
Fixes: 11777435c7 ("net/ixgbe: parse flow director filter")
Fixes: 672be56d76 ("net/ixgbe: parse n-tuple filter")
Fixes: eb3539fc85 ("net/ixgbe: parse ethertype filter")
Fixes: 429f6ebb42 ("net/ixgbe: parse TCP SYN filter")
Cc: stable@dpdk.org
Signed-off-by: Wei Zhao <wei.zhao1@intel.com>
Signed-off-by: Wenzhuo Lu <wenzhuo.lu@intel.com>
Move ixgbe 2 mac type check macros to ixgbe_ethdev.h in
order to be used by filter parser functions in file
ixgbe_flow.c.
Signed-off-by: Wei Zhao <wei.zhao1@intel.com>
Signed-off-by: Wenzhuo Lu <wenzhuo.lu@intel.com>
Delete useless function declarations in file ixgbe_flow.c and
adjust function definition position to avoid compile error.
Signed-off-by: Wei Zhao <wei.zhao1@intel.com>
With all vmxnet3 version 3 changes incorporated in the vmxnet3 driver,
the driver can configure emulation to run at vmxnet3 version 3, provided
the emulation advertises support for version 3.
This patch also updates release notes.
Signed-off-by: Shrikrishna Khare <skhare@vmware.com>
Acked-by: Yong Wang <yongwang@vmware.com>
Acked-by: Jin Heo <heoj@vmware.com>
In vmxnet3 version 3, the emulation added support for the vmxnet3 driver
to communicate information about the memory regions the driver will use
for rx/tx buffers. The driver can also indicate which rx/tx queue the
memory region is applicable for. If this information is communicated
to the emulation, the emulation will always keep these memory regions
mapped, thereby avoiding the mapping/unmapping overhead for every packet.
Signed-off-by: Shrikrishna Khare <skhare@vmware.com>
Signed-off-by: Guolin Yang <gyang@vmware.com>
Acked-by: Yong Wang <yongwang@vmware.com>
Acked-by: Jin Heo <heoj@vmware.com>
