Nipun Gupta [Thu, 29 Aug 2019 10:27:11 +0000 (15:57 +0530)]
bus/dpaa: decouple FQ portal alloc and init
The decoupling of FQ portal allocation is required as a
pre-requisite to support Rx interrupts as we need to have
event FD's at portal allocation i.e. before the
initialization of the Frame Queues.
This change will help us get the event fd once the portals
have been allocated for static FQ's.
Hemant Agrawal [Thu, 29 Aug 2019 10:27:08 +0000 (15:57 +0530)]
bus/dpaa: fix dpaa_sec blacklist
The black list of dpaa_sec devices fails.
EAL: failed to parse device "dpaa:dpaa_sec-1"
This patch address following issues:
- bus usages dpaa-sec while the driver usage dpaa_sec
- bus usages numbers from 0 to MAX_SEC - while driver
probe usages sec number form max-fman_device +1
Somnath Kotur [Wed, 2 Oct 2019 17:17:44 +0000 (10:17 -0700)]
net/bnxt: support for QinQ insertion and stripping
Driver will accelerate only outer/S-VLAN insertion by turning on
the appropriate bits in the Tx Buffer Descriptor when the packet
arrives for transmission.
The TPID to be used for this S-VLAN is conveyed by the vlan_tpid_set
dev_op which will terminate in the driver.
In the Rx path, driver will continue providing the stripped vlan tag
in the mbuf's vlan tci field. This would be the outermost vlan tag
in a double-tagged packet or the vlan tag for a single vlan tagged pkt.
The secondary process should not modify device state when
init is called. The pci device information pointed to by
dev_private pointer is not correct in secondary process.
Fixes: b7778e8a1c00 ("net/bnxt: refactor to properly allocate resources for PF/VF") Cc: stable@dpdk.org Signed-off-by: Stephen Hemminger <stephen@networkplumber.org> Acked-by: Ajit Khaparde <ajit.khaparde@broadcom.com>
Kalesh AP [Wed, 2 Oct 2019 17:17:37 +0000 (10:17 -0700)]
net/bnxt: fix deferred start of Tx queues
Driver should not change "deferred_start" state of the tx queues.
It should get the state in queue_setup_op() and use that value.
Since the deferred start state was being used in the packet transmit
functions to determine whether the queue has been stopped already,
introduced a per-txq flag to track queue stopped/started state.
Kalesh AP [Wed, 2 Oct 2019 17:17:36 +0000 (10:17 -0700)]
net/bnxt: fix L2 context calculation for Thor
On Wh+, number of L2 context supported by FW is the sum of number of
EM flow count and number of L2 context count returned in HWRM_FUNC_QCFG.
This is not true for Thor. Restrict this only for Whitney for now.
Fixes: ff9c0ca47e81 ("net/bnxt: save the number of EM flow count") Cc: stable@dpdk.org Signed-off-by: Kalesh AP <kalesh-anakkur.purayil@broadcom.com> Signed-off-by: Ajit Khaparde <ajit.khaparde@broadcom.com> Reviewed-by: Somnath Kotur <somnath.kotur@broadcom.com> Reviewed-by: Lance Richardson <lance.richardson@broadcom.com>
Ajit Khaparde [Wed, 2 Oct 2019 17:17:34 +0000 (10:17 -0700)]
net/bnxt: fix async link handling and update
When updating the link because of an async link notification
there is no need to set wait_for_completion. At this point
the link related information should be available without need to poll.
Use rte_eth_linkstatus_set instead of memcpy to ensure atomicity
while updating the link status.
We force the physical link down as a part of device stop.
But we are not waiting there enough and handling the async notification
before exiting. It just sits in the default CQ till we do a device
start.
Fix it by calling the CQ handler in device stop.
Fixes: 7bc8e9a227cc ("net/bnxt: support async link notification") Cc: stable@dpdk.org Signed-off-by: Ajit Khaparde <ajit.khaparde@broadcom.com> Reviewed-by: Kalesh AP <kalesh-anakkur.purayil@broadcom.com> Reviewed-by: Lance Richardson <lance.richardson@broadcom.com>
Kalesh AP [Wed, 2 Oct 2019 17:17:32 +0000 (10:17 -0700)]
net/bnxt: fix Rx queue start/stop
Driver should not change "deferred_start" state of the rx queues.
It should get the state in queue_setup_op() and use that value.
Since the deferred start state was being used in the packet receive
functions to determine whether a stopped rx ring should be polled,
introduced a per-rxq flag to track queue stopped/started state.
Maximum frame length supported by hw is 9600 bytes.
When user launches testpmd with --mbuf-size=9729,
we are posting buffer descriptors of size 9601 to the ring.
This was causing packets getting discarded.
Fixes: daef48efe5e5 ("net/bnxt: support set MTU") Cc: stable@dpdk.org Signed-off-by: Santoshkumar Karanappa Rastapur <santosh.rastapur@broadcom.com> Signed-off-by: Kalesh AP <kalesh-anakkur.purayil@broadcom.com> Signed-off-by: Ajit Khaparde <ajit.khaparde@broadcom.com> Reviewed-by: Somnath Kotur <somnath.kotur@broadcom.com> Reviewed-by: Lance Richardson <lance.richardson@broadcom.com>
When user creates a flow similar to an existing flow with just the
destination queue change, we delete the old filter and allocate a new
one with this destination queue change. We were also allocating a new L2
filter matching the same destination mac resulting in 2 L2 filters for the
same destination mac.
This was causing any flow matching the destination mac to be steered to
this queue instead of the default queue.
Gavin Hu [Mon, 16 Sep 2019 11:27:18 +0000 (19:27 +0800)]
net/bnxt: enforce IO barrier for doorbell command
The doorbell ringing operation requires a rte_io_mb immediately to make
the command complete and visible to the device before reading the
response, otherwise it may read stale or invalid responses.
Fixes: ca241d9a0952 ("net/bnxt: use I/O device memory read/write API") Cc: stable@dpdk.org Signed-off-by: Gavin Hu <gavin.hu@arm.com> Acked-by: Ajit Khaparde <ajit.khaparde@broadcom.com>
Gavin Hu [Mon, 16 Sep 2019 11:27:16 +0000 (19:27 +0800)]
net/bnxt: remove duplicate barrier
As there is an inclusive rte_io_wmb within the following rte_write32()
API who rings the doorbell, this makes the above rte_wmb unnecessary and
remove it.
Fixes: 1cd45aeb3270 ("net/bnxt: support Stratus VF device") Cc: stable@dpdk.org Signed-off-by: Gavin Hu <gavin.hu@arm.com> Reviewed-by: Steve Capper <steve.capper@arm.com> Reviewed-by: Ruifeng Wang <ruifeng.wang@arm.com> Reviewed-by: Phil Yang <phil.yang@arm.com> Acked-by: Ajit Khaparde <ajit.khaparde@broadcom.com>
Gavin Hu [Mon, 16 Sep 2019 11:27:15 +0000 (19:27 +0800)]
net/ice: use relaxed and remove duplicate barrier
To guarantee the orderings of successive stores to CIO and MMIO memory,
a lighter weight rte_io_wmb [1] can be used instead of rte_wmb, and since
the ICE_PCI_REG_WRITE API already has an inclusive rte_io_wmb, this
explicit call can even be saved.
Gavin Hu [Mon, 16 Sep 2019 11:27:14 +0000 (19:27 +0800)]
net/i40e: use relaxed and remove duplicate barrier
To guarantee the orderings of successive stores to CIO and MMIO memory,
a lighter weight rte_io_wmb [1] can be used instead of rte_wmb, and since
the I40E_PCI_REG_WRITE API already has an inclusive rte_io_wmb, this
explicit call can be even saved.
Kalesh AP [Wed, 2 Oct 2019 01:23:35 +0000 (18:23 -0700)]
net/bnxt: support PTP for Thor
On Thor, direct access to PTP registers (via GRC) is not supported.
Driver must use HWRM to access the timestamp information.
Vectorized Rx/Tx cannot be enabled if RTE_LIBRTE_IEEE1588=y.
Remove the PTP flags handling code from the vector Rx path.
Add support to read tx timestamp value and the time from the
timesync clock.
On Thor, Rx timestamps are provided directly in the Rx completion
records to the driver. Only 32 bits of the timestamp is present in
the completion. Driver needs to read the current 48 bit free running
timer using the HWRM_PORT_TS_QUERY command and combine the upper
16 bits from the HWRM response with the lower 32 bits in the
Rx completion to produce the 48 bit timestamp for the Rx packet.
Kalesh AP [Wed, 2 Oct 2019 01:23:33 +0000 (18:23 -0700)]
net/bnxt: avoid null pointer dereference
Commit "bd0a14c99f65" enables the creation of a dedicated completion
ring for asynchronous event handling instead of handling these
events on a receive completion ring on non Stingray Platforms.
This causes a segfault due to NULL pointer dereference in
bnxt_alloc_async_cp_ring() on stingray. Fix this by checking the
pointer validity before accessing it.
Fixes: bd0a14c99f65 ("net/bnxt: use dedicated CPR for async events") Cc: stable@dpdk.org Signed-off-by: Kalesh AP <kalesh-anakkur.purayil@broadcom.com> Signed-off-by: Ajit Khaparde <ajit.khaparde@broadcom.com> Reviewed-by: Rahul Gupta <rahul.gupta@broadcom.com> Reviewed-by: Lance Richardson <lance.richardson@broadcom.com>
Kalesh AP [Wed, 2 Oct 2019 01:23:32 +0000 (18:23 -0700)]
net/bnxt: reduce verbosity of logs
When IOMMU is available, EAL picks IOVA as VA as the default IOVA mode.
This causes the bnxt driver to log warning messages saying
"Memzone physical address same as virtual." and
"Using rte_mem_virt2iova()" during load.
Reduce the verbosity of logs to DEBUG. Reduced couple of other
logs level to DEBUG as well.
Kalesh AP [Wed, 2 Oct 2019 01:23:31 +0000 (18:23 -0700)]
net/bnxt: support hot firmware upgrade for Stingray
FW sets HWRM_FUNC_QCAPS_OUTPUT_FLAGS_ERR_RECOVER_RELOAD
in HWRM_FUNC_QCAPS command, if device requires to invoke
fastboot FW during FW reset.
Driver has to poll for shutdown bit in fw_status register:
1. in case of hot fw upgrade, this bit will be set after all
function drivers unregistered with fw.
2. in case of fw initiated error recovery, this bit will be
set after fw has collected the core dump
Kalesh AP [Wed, 2 Oct 2019 01:23:30 +0000 (18:23 -0700)]
net/bnxt: support FW reset
Added code to perform FW_RESET. When the driver detects error in FW,
it has to initiate the recovery by resetting the cores. FW advertise
the method to do a core reset, reset register offsets and values
to perform reset in response of HWRM_ERROR_RECOVERY_QCFG command.
There are 2 ways to recover from the error.
1. Master function issues core resets to recover from error.
2. Master function detects chimp dead condition and notify the Kong
processor about the chimp dead case through FW_RESET HWRM command.
Kong Processor send an RESET_NOTIFY async event with
REASON_CODE_FW_EXCEPTION_FATAL to all the PF’s/VF’s that
chimp is dead and it is going to reset the chimp.
Kalesh AP [Wed, 2 Oct 2019 01:23:29 +0000 (18:23 -0700)]
net/bnxt: support periodic FW health monitoring
Periodically poll the FW heartbeat register and FW recovery counter
registers to check the FW health. Polling frequency will be
advertised by the FW in HWRM_ERROR_RECOVERY_QCFG response.
Schedule the task upon receiving the async event from FW.
Kalesh AP [Wed, 2 Oct 2019 01:23:28 +0000 (18:23 -0700)]
net/bnxt: handle error recovery FW event
1. Advertise HWRM_FUNC_DRV_RGTR_INPUT_FLAGS_ERROR_RECOVERY_SUPPORT flag
in the FUNC_DRV_RGTR command.
2. request for the async event ASYNC_EVENT_CMPL_EVENT_ID_ERROR_RECOVERY
in the FUNC_DRV_RGTR command.
3. handle the async event EVENT_ID_ERROR_RECOVERY from FW.
Error recovery support will be used by firmware only if all the driver
instances support error recovery process.
In Driver initiated error recovery process, driver has to know about
the registers offset and values to initiate FW reset. The HWRM command
HWRM_ERROR_RECOVERY_QCFG is used to obtain all the registers and values
required to initiate FW reset. This command response includes
FW heart_beat register, health status register, Error counter register,
register offsets and values to do chip reset if firmware crashes and
becomes unresponsive.
Kalesh AP [Wed, 2 Oct 2019 01:23:25 +0000 (18:23 -0700)]
net/bnxt: handle fatal event from FW
When firmware hit some unrecoverable error conditions, firmware initiate
the recovery by sending an async event EVENT_CMPL_EVENT_ID_RESET_NOTIFY
with data1 set to RESET_NOTIFY_EVENT_DATA1_REASON_CODE_FW_EXCEPTION_FATAL
to all host drivers and will reset the chip.
The recovery procedure is same sequence as the one for hot FW upgrade.
Kalesh AP [Wed, 2 Oct 2019 01:23:24 +0000 (18:23 -0700)]
net/bnxt: inform firmware about IF state changes
Use latest firmware API to inform firmware about IF state changes.
Firmware has the option to clean up resources during IF down and
to require the driver to reserve resources again during IF up.
During port start, HWRM_FUNC_DRV_IF_CHANGE command response
flags indicates that firmware has reset. Add logic to re-probe
the firmware and re-setup resources.
Kalesh AP [Wed, 2 Oct 2019 01:23:23 +0000 (18:23 -0700)]
net/bnxt: handle reset notify async event from FW
When the FW upgrade is initiated the current instance
of FW issues a HWRM_ASYNC_EVENT_CMPL_EVENT_ID_RESET_NOTIFY
async notification to the driver. On receiving this notification,
the PMD shall quiesce itself and poll on the HWRM_VER_GET FW
command at regular intervals.
Once the VER_GET command succeeds, the driver should go through
the rediscovery process and re-initialize the device.
Also register with FW for the reset notify async event.
Kalesh AP [Wed, 2 Oct 2019 01:23:22 +0000 (18:23 -0700)]
net/bnxt: prevent device access when device is in reset
Refactor init and uninit functions so that the driver can fail
the eth_dev_ops callbacks and accessing Tx and Rx queues
when device is in reset or in error state.
Transmit and receive queues are freed during reset cleanup and
reallocated during recovery. So we block all data path handling
in this state. The eth_dev dev_started field is updated depending
on the status of the device.
The devices backed by mlx5 PMD might share the same multiport
Infiniband device context. It regards representors and slaves
of bonding device. These ports are spawned with devargs.
These patch check whether configuration deduced from these
devargs is compatible with configurations if devices
sharing the same context. It prevents the incorrect
whitelists, like:
net/mlx5: update switch port id in bonding configuration
With bonding configuration multiple PFs may represent the
single switching device with multiple ports as representors.
To distinguish representors belonging to different PFs we
should generated unique port ID. It is proposed to use
the PF index in bonding configuration to generate this
unique port IDs.
With bonding configurations the switch domain may be shared
between multiple PCI devices, we should search the switch
sibling devices within the entire set of present ethernet
devices backed by the mlx5 PMD.
net/mlx5: update source and destination vport translations
There new kernel/rdma_core [1] supports matching on metadata
register instead of vport field to provide operations over
VF LAG bonding configurations. This patch provides correct
translations for flow matchers and destination port actions
if united E-Switch (for VF LAG) is configured and/or new vport
matching mode is engaged.
net/mlx5: elaborate E-Switch port parameters query
The routine mlx5_port_to_eswitch_info() is elaborated
to two ones (get E-Switch port parameters by port and
by device pointer) and simplified to returning structure
containing all parameters instead of copying.
net/mlx5: query vport index match mode and parameters
There new kernel/rdma_core [1] supports matching on metadata
register instead of vport field to provide operations over
VF LAG bonding configurations. The patch retrieves parameters
and information about the way is engaged to match vport on E-Switch.
If bonding Infiniband device is found the unified E-Switch
is supposed and the extra rdma-core/kernel support is needed
to retrieve vport indices. The patch introduces this feature
defines, bonding support check is added to probe routine.
net/mlx5: add VF LAG mode bonding device recognition
The Mellanox NICs starting from ConnectX-5 support LAG over
NIC ports internally, implemented by the NIC firmware and hardware.
The multiport NIC presents multiple physical PCI functions (PF),
with SR-IOV multiple virtual PCI functions (VFs) might be presented.
With switchdev mode the VF representors are engaged and PFs and their
VFs are connected by internal E-Switch feature. Each PF and related VFs
have dedicated E-Switch and belong to dedicated switch domain.
If NIC ports are combined to support NIC the kernel drivers introduce
the single unified Infiniband multiport devices, and all only one
unified E-Switch with single switch domain combines master PF
all all VFs. No extra DPDK bonding device is needed.
At device probing the device list to spawn was allocated
as dynamic size local variable. It was no possible to have
one unified exit point from routine due to compiler warnings.
This patch allocates the spawn device list directly with
rte_zmalloc() and it is possible to goto to unified exit
label from anywhere of the routine.
The routine mlx5_ibv_device_to_pci_addr() takes Infiniband
device list object, takes the device sysfs path from there
and retrieves PCI address. The routine may be implemented
in more generic way by taking sysfs path directly as parameter
and can be used for getting PCI address of netdevs.
The generic routine is renamed to mlx5_dev_to_pci_addr()
net/mlx5: move backing PCI device to private context
Now all devices created over the same multiport IB device
have shared context containing the backing PCI device field.
For the VF LAG configurations it becomes possible the
representors might be connected to VF created over different
PFs. In this case representors have the different backing
PCI devices and mentioned field should be moved to device
private area.
John Daley [Thu, 5 Sep 2019 21:33:17 +0000 (14:33 -0700)]
net/enic: add PCI id for new virtual function
Probe for the PCI ID of a new mode of VF which will be added to VIC
adapter firmware. When fully implemented, the new mode will operate
independent of the Cisco Virtual Machine Fabric Extender (VM-FEX) and
will not need to be provisioned through libvirt. The new mode is dubbed
"Standalone vNic" mode or "SN" mode.
Also, minor formatting changes.
Signed-off-by: John Daley <johndale@cisco.com> Reviewed-by: Hyong Youb Kim <hyonkim@cisco.com>
Packet processing in secondary process cannot work because rx_pkt_burst
and tx_pkt_burst in eth_dev are not initialized while probing device.
This patch is to the initialization.
Fixes: ee27edbe0c10 ("drivers/net: share vdev data to secondary process") Cc: stable@dpdk.org Signed-off-by: Yasufumi Ogawa <yasufum.o@gmail.com> Reviewed-by: Ferruh Yigit <ferruh.yigit@intel.com>
Simei Su [Tue, 1 Oct 2019 09:22:12 +0000 (17:22 +0800)]
ethdev: add symmetric Toeplitz hash
Currently, there are DEFAULT,TOEPLITZ and SIMPLE_XOR hash function.
To support symmetric hash by rte_flow RSS action, this patch adds
new hash function "Symmetric Toeplitz" which is supported by some hardware.
Signed-off-by: Simei Su <simei.su@intel.com> Reviewed-by: Qi Zhang <qi.z.zhang@intel.com> Acked-by: Ori Kam <orika@mellanox.com> Acked-by: Andrew Rybchenko <arybchenko@solarflare.com>
Ying A Wang [Wed, 28 Aug 2019 06:00:38 +0000 (14:00 +0800)]
ethdev: add PPPoE to flow API
- RTE_FLOW_ITEM_TYPE_PPPOES: matches a PPPoE session header.
- RTE_FLOW_ITEM_TYPE_PPPOED: matches a PPPoE discovery header.
- RTE_FLOW_ITEM_TYPE_PPPOE_PROTO_ID: matches a PPPoE session
protocol identifier.
Signed-off-by: Ying A Wang <ying.a.wang@intel.com> Acked-by: Ori Kam <orika@mellanox.com> Reviewed-by: Qi Zhang <qi.z.zhang@intel.com> Reviewed-by: Ferruh Yigit <ferruh.yigit@intel.com>
Ying A Wang [Wed, 28 Aug 2019 06:00:37 +0000 (14:00 +0800)]
ethdev: add GTP extension header to flow API
- RTE_FLOW_ITEM_TYPE_GTP_PSC: matches a GTP
- RTE_FLOW_ITEM_TYPE_GTP_PSC: matches a GTP
PDU extension header (PDU session container).
Signed-off-by: Ying A Wang <ying.a.wang@intel.com> Acked-by: Ori Kam <orika@mellanox.com> Reviewed-by: Qi Zhang <qi.z.zhang@intel.com> Reviewed-by: Ferruh Yigit <ferruh.yigit@intel.com>
Hui Zhao [Thu, 21 Mar 2019 20:28:14 +0000 (21:28 +0100)]
net/bonding: fix OOB access in other aggregator modes
slave aggregator_port_id is in [0, RTE_MAX_ETHPORTS-1] range.
If RTE_MAX_ETHPORTS is > 8, we can hit out of bound accesses on
agg_bandwidth[] and agg_count[] arrays.
Fixes: 6d72657ce379 ("net/bonding: add other aggregator modes") Cc: stable@dpdk.org Signed-off-by: Hui Zhao <zhaohui8@huawei.com> Signed-off-by: David Marchand <david.marchand@redhat.com> Reviewed-by: Maxime Coquelin <maxime.coquelin@redhat.com> Acked-by: Chas Williams <chas3@att.com>
During PF/VF link update, a default speed value of 100M will be set
if get_link_info has failed or speed is unknown.
Consequently if PF is put in no-carrier state, VFs will switch to
"in carrier" state due to a link up + a link speed set to 100M
(default value if no speed detected).
To be consistent with linux drivers on which PF and VFs are in
same carrier state, sets default speed to undefined (instead of 100M)
and updates a link status of VF only if link is up and speed is
different from undefined.
The speed capability of X553 1GbE should be ETH_LINK_SPEED_1G |
ETH_LINK_SPEED_100M | ETH_LINK_SPEED_10M rather than ETH_LINK_SPEED_1G |
ETH_LINK_SPEED_10G. Correct it to fix the issue.
Add support for setting VLAN PCP field via rte_flow API. Hardware
overwrites the entire 16-bit VLAN TCI field. So, both VLAN VID and
PCP actions must be specified.
Add support for matching VLAN fields via rte_flow API.
When matching VLAN pattern, the ethertype field in hardware
filter specification must contain VLAN header's ethertype, and
not Ethernet header's ethertype. The hardware automatically
searches for ethertype 0x8100 in Ethernet header, when
parsing incoming packet against VLAN pattern.
Remove compile time option to control Tx coalescing Latency vs
Throughput behavior. Add tx_mode_latency devarg instead, to
dynamically control Tx coalescing behavior.
net/cxgbe: avoid polling link status before device start
Link updates come in firmware event queue, which is only created
when device starts. So, don't poll for link status if firmware
event queue is not yet created.
This fixes NULL dereference when accessing non existent firmware
event queue.
Fixes: 265af08e75ba ("net/cxgbe: add link up and down ops") Cc: stable@dpdk.org Signed-off-by: Rahul Lakkireddy <rahul.lakkireddy@chelsio.com>
Set VLAN action mode to VLAN_REWRITE only if VLAN_INSERT has not been
set yet. Otherwise, the resulting VLAN packets will have their VLAN
header rewritten, instead of pushing a new outer VLAN header.
Also fix the VLAN ID extraction logic and endianness issues.
In the process of recovery from error CQE, when using vectorized Rx
burst, the initialization of data length in mbufs was not done.
As a result the wrong length was left written in mbuf, causing
memory overwrite or wrong error report.
This patch fixes the initialization of mbuf data length during
recovery from error CQE, when using vectorized Rx burst,
net/mlx5: fix UAR remap initialization for 32-bit systems
The txq_uar_init() routine uses the uninitialized uar_mmap_offset
field in 32-bit configurations due to this field is initialized
after txq_uar_init() call.
Noa Ezra [Wed, 19 Jun 2019 06:13:44 +0000 (06:13 +0000)]
net/vhost: support TSO disabling
TSO (TCP Segmentation Offload) is enabled by default on vhost.
Add the ability to disable TSO on vhost.
The user should also disable the feature on the virtual machine's xml.
Michal Krawczyk [Thu, 26 Sep 2019 10:45:27 +0000 (12:45 +0200)]
maintainers: update for ena
Add Igor Chauskin from the Amazon as another maintainer of the driver.
Igor is another person from the Amazon team that is responsible for the
ENA DPDK driver.
Marvin Liu [Mon, 23 Sep 2019 14:05:11 +0000 (22:05 +0800)]
net/virtio: fix mbuf data and packet length mismatch
If reserve virtio header room by function rte_pktmbuf_prepend, both
segment data length and packet length of mbuf will be increased.
Data length will be equal to descriptor length, while packet length
should be decreased as virtio-net header won't be taken into packet.
Thus will cause mismatch in mbuf structure. Fix this issue by access
mbuf data directly and increase descriptor length if it is needed.
Fixes: 58169a9c8153 ("net/virtio: support Tx checksum offload") Fixes: 892dc798fa9c ("net/virtio: implement Tx path for packed queues") Fixes: 4905ed3a523f ("net/virtio: optimize Tx enqueue for packed ring") Fixes: e5f456a98d3c ("net/virtio: support in-order Rx and Tx") Cc: stable@dpdk.org Reported-by: Stephen Hemminger <stephen@networkplumber.org> Signed-off-by: Marvin Liu <yong.liu@intel.com> Reviewed-by: Maxime Coquelin <maxime.coquelin@redhat.com>
net/virtio: get all pending Rx packets in vectorized paths
The loop to read packets does not take all packets as the number of
available packets (nb_used) is decremented in the loop.
Take all available packets provides a performance improvement of 3%.
Adrian Moreno [Fri, 6 Sep 2019 12:50:20 +0000 (14:50 +0200)]
vhost: fix vring memory partially mapped
Only the mapping of the vring addresses is being ensured. This causes
errors when the vring size is larger than the IOTLB page size. E.g:
queue sizes > 256 for 4K IOTLB pages
Ensure the entire vring memory range gets mapped. Refactor duplicated
code for for IOTLB UPDATE and IOTLB INVALIDATE and add packed virtqueue
support.
Fixes: 09927b524969 ("vhost: translate ring addresses when IOMMU enabled") Cc: stable@dpdk.org Signed-off-by: Adrian Moreno <amorenoz@redhat.com> Reviewed-by: Maxime Coquelin <maxime.coquelin@redhat.com>
Marvin Liu [Wed, 18 Sep 2019 17:06:55 +0000 (01:06 +0800)]
net/virtio: cleanup on demand when in-order Tx
Check whether space are enough before burst enqueue operation. If more
space is needed, will try to clean up used descriptors for space on
demand. It can give more chances to free used descriptors, thus will
help RFC2544 performance. Also deduct failed xmit packets from total
xmit number.
Fixes: e5f456a98d3c ("net/virtio: support in-order Rx and Tx") Cc: stable@dpdk.org Signed-off-by: Marvin Liu <yong.liu@intel.com> Reviewed-by: Maxime Coquelin <maxime.coquelin@redhat.com>
Marvin Liu [Wed, 18 Sep 2019 17:06:54 +0000 (01:06 +0800)]
net/virtio: update stats when in-order Tx done
When doing xmit in-order enqueue, packets are buffered and then flushed
into avail ring. Buffered packets can be dropped due to insufficient
space. Moving stats update action just after successful avail ring
updates can guarantee correctness.
Fixes: e5f456a98d3c ("net/virtio: support in-order Rx and Tx") Cc: stable@dpdk.org Signed-off-by: Marvin Liu <yong.liu@intel.com> Reviewed-by: Maxime Coquelin <maxime.coquelin@redhat.com>
Noa Ezra [Thu, 20 Jun 2019 06:33:03 +0000 (06:33 +0000)]
net/vhost: fix redundant queue state event
In some situations, when a virtual machine is starting,
vring_state_changed can be called while there was no change in the
queue state. This fix makes sure that there was really a change in the
queue state before calling the callback for EVENT_QUEUE_STATE.
Fixes: ee584e9710b9 ("vhost: add driver on top of the library") Cc: stable@dpdk.org Signed-off-by: Noa Ezra <noae@mellanox.com> Reviewed-by: Matan Azrad <matan@mellanox.com> Reviewed-by: Maxime Coquelin <maxime.coquelin@redhat.com>
Tiwei Bie [Mon, 19 Aug 2019 11:34:57 +0000 (19:34 +0800)]
vhost: protect vring access done by application
Besides the enqueue/dequeue API, other APIs of the builtin net
backend should also be protected.
Fixes: a3688046995f ("vhost: protect active rings from async ring changes") Cc: stable@dpdk.org Reported-by: Peng He <xnhp0320@icloud.com> Signed-off-by: Tiwei Bie <tiwei.bie@intel.com> Reviewed-by: Maxime Coquelin <maxime.coquelin@redhat.com>
Tiwei Bie [Mon, 19 Aug 2019 11:34:56 +0000 (19:34 +0800)]
vhost: fix vring address handling during live migration
When live migration starts, QEMU will set ring addrs again for
each virtqueue. In this case, we should try to translate ring
addrs after we invalidating the ring, otherwise virtqueues can
be enabled with the addrs untranslated. Besides, also leverage
the access_ok flag in non-IOMMU case to prevent the data path
accessing invalidated virtqueues.
Fixes: 5a4933e56be4 ("vhost: postpone ring address translations at kick time only") Cc: stable@dpdk.org Reported-by: Yilong Lv <lvyilong.lyl@alibaba-inc.com> Signed-off-by: Tiwei Bie <tiwei.bie@intel.com> Reviewed-by: Maxime Coquelin <maxime.coquelin@redhat.com>
Tiwei Bie [Mon, 19 Aug 2019 11:34:55 +0000 (19:34 +0800)]
vhost: forbid reallocation when running
When the device has been started, don't do the reallocation anymore.
Otherwise the pointers used in application threads can be invalidated
without proper protection. Instead of introducing a global lock to
protect the change of device pointers which will hurt the performance,
let's just do the reallocation during setup.
Fixes: af295ad4698c ("vhost: realloc device and queues to same numa node as vring desc") Cc: stable@dpdk.org Reported-by: Yinan Wang <yinan.wang@intel.com> Signed-off-by: Tiwei Bie <tiwei.bie@intel.com> Reviewed-by: Maxime Coquelin <maxime.coquelin@redhat.com>
Jim Harris [Tue, 24 Sep 2019 09:23:34 +0000 (02:23 -0700)]
vhost: add missing experimental flag
This function is listed under EXPERIMENTAL in the
rte_vhost_version.map, so it needs to be marked
with __rte_experimental in the header file as well.
Found by check-experimental-syms.sh when trying to compile
DPDK with -finstrument-functions. This script didn't
catch this in the normal case, since the function is
declared __rte_always_inline.
This also requires updating the vhost_scsi example to allow
use of this newly marked experimental API.
Signed-off-by: Jim Harris <james.r.harris@intel.com> Reviewed-by: Maxime Coquelin <maxime.coquelin@redhat.com>
git.droids-corp.org / dpdk.git / log