Transmit: Add test_tx_bandwidth_dump_enable support to
transmit packets from the driver which helps measure
line rate for 25G and 10G.
pkt1500, pkt8192, pkt64 packet sizes can be configured.
Receive- Drop Packets at the driver level.
Bug 4742868
Change-Id: I52fe1b73b169dfedc06afea5fb62d7bc21fd9ba1
Signed-off-by: Nagaraj Annaiah <nannaiah@nvidia.com>
Reviewed-on: https://git-master.nvidia.com/r/c/linux-nv-oot/+/3197198
Reviewed-by: Ashutosh Jha <ajha@nvidia.com>
GVS: buildbot_gerritrpt <buildbot_gerritrpt@nvidia.com>
- Adding WAR to ignore lane bringup restart task
as it is causing lane bring up failures in SLT EQOS/MGBE
- Populate RSS hash table with enabled num_of_dma channels
Bug 4709627
Change-Id: I195db0371f69dfcedc1c67023c1279af426dd7e6
Signed-off-by: Mahesh Patil <maheshp@nvidia.com>
Reviewed-on: https://git-master.nvidia.com/r/c/linux-nv-oot/+/3162313
Reviewed-by: Nagaraj Annaiah <nannaiah@nvidia.com>
Reviewed-by: Ashutosh Jha <ajha@nvidia.com>
Existing implementation uses a single page_pool
for all Rx DMA channels. As by default all irqs
are on CPU 0, this does not cause any issue.
Routing an irq to some other CPU causes race
conditon for page allocation from page_pool which
leads to random memory corruption.
[ 158.416637] Call trace:
[ 158.416644] arm_lpae_map_pages+0xb4/0x1e0
[ 158.416649] arm_smmu_map_pages+0x8c/0x160
[ 158.416661] __iommu_map+0xf8/0x2b0
[ 158.416677] iommu_map_atomic+0x58/0xb0
[ 158.416683] __iommu_dma_map+0xac/0x150
[ 158.416687] iommu_dma_map_page+0xf4/0x220
[ 158.416690] dma_map_page_attrs+0x1e8/0x260
[ 158.416727] page_pool_dma_map+0x48/0xd0
[ 158.416750] __page_pool_alloc_pages_slow+0xc4/0x390
[ 158.416757] page_pool_alloc_pages+0x64/0x90
[ 158.416762] ether_padctrl_mii_rx_pins+0x164/0x9d0 [nvethernet]
[ 158.416807] ether_padctrl_mii_rx_pins+0x478/0x9d0 [nvethernet]
[ 158.416822] osi_process_rx_completions+0x284/0x4d0 [nvethernet]
[ 158.416832] 0xffffa26218b8f71c
[ 158.416855] __napi_poll+0x48/0x230
[ 158.416871] net_rx_action+0xf4/0x290
[ 158.416875] __do_softirq+0x130/0x3e8
[ 158.416889] __irq_exit_rcu+0xe8/0x110
This change creates a page_pool per Rx DMA channel.
This ensures there is no race conditon for page
alloc/dealloc from each DMA napi context.
Bug 4541158
Change-Id: I12668ee7d824fd30d54a874bbbdf190d02943478
Signed-off-by: Aniruddha Paul <anpaul@nvidia.com>
Reviewed-on: https://git-master.nvidia.com/r/c/linux-nv-oot/+/3091494
Reviewed-by: Jon Hunter <jonathanh@nvidia.com>
GVS: Gerrit_Virtual_Submit <buildbot_gerritrpt@nvidia.com>
Reviewed-by: Ashutosh Jha <ajha@nvidia.com>
Issue: If OSI_RX_SWCX_REUSE flag is set in OSI then
OSD will not reallocate the buffer for that descriptor
index. It will re-use the same buffer virt/physical
addresses. If that buffer is a reserved buffer then OSI
will not process that so it will result in packet drops.
Fix: Reuse the buffer is OSI_RX_SWCX_REUSE flag is set
and the buffer is not a reserved buffer.
Bug 3727058
Change-Id: I3331604a90b2cbe902d91f2082e1fe36f830decc
Signed-off-by: Bhadram Varka <vbhadram@nvidia.com>
Reviewed-on: https://git-master.nvidia.com/r/c/linux-nvidia/+/2745194
Tested-by: mobile promotions <svcmobile_promotions@nvidia.com>
Reviewed-by: mobile promotions <svcmobile_promotions@nvidia.com>
Issue:
1) dev_queue_xmit acquired the per queue tx lock and called driver
transmit routine.
2) During the transmit - common interrupt asserted which tries to
acquire the same tx lock which resulted in lock recursion.
Fix: Move the lock acquire to tasklet context under common isr.
Bug 3773016
Change-Id: I7cfd49beb1238286d3bccd9e4b9ccc054c4f6d30
Signed-off-by: Bhadram Varka <vbhadram@nvidia.com>
Reviewed-on: https://git-master.nvidia.com/r/c/linux-nvidia/+/2770227
Reviewed-by: Narayan Reddy <narayanr@nvidia.com>
Reviewed-by: Bitan Biswas <bbiswas@nvidia.com>
Reviewed-by: svcacv <svcacv@nvidia.com>
GVS: Gerrit_Virtual_Submit <buildbot_gerritrpt@nvidia.com>
Issue: In the current scenrio, when the remote
interface is brought down and up. There is no way
of getting the link up again with the remote device
which in turn fails the data transfers even though
interface from other end is brought up.
Fix: When interface is gone down on the other side,
MAC receives local faults and in this scenario
no data should be sent to MAC, so disable network
queues and initiate the lane bring up process for
monitoring the link status. Once the link is up,
re enable the network queues for data transfers.
Also added support for enabling driver logs using
msglvl of ethtool
Bug 3744088
Bug 3654543
Bug 3665378
Change-Id: I16cdee74e4e3ff6cd176924575f5005f385d4c5d
Signed-off-by: Narayan Reddy <narayanr@nvidia.com>
Reviewed-on: https://git-master.nvidia.com/r/c/linux-nvidia/+/2730876
Reviewed-by: svc-mobile-coverity <svc-mobile-coverity@nvidia.com>
Reviewed-by: svc-mobile-cert <svc-mobile-cert@nvidia.com>
Reviewed-by: svc_kernel_abi <svc_kernel_abi@nvidia.com>
Reviewed-by: Rakesh Goyal <rgoyal@nvidia.com>
Reviewed-by: Bhadram Varka <vbhadram@nvidia.com>
Reviewed-by: Srinivas Ramachandran <srinivasra@nvidia.com>
GVS: Gerrit_Virtual_Submit
Issue: Observed Tx timestamp timeout from application since
workqueue is not getting scheduled at the driver layer to consume
the Tx timestamp from the OSI layer which resulted in tx_timeout
from APP layer.
Fix: Invoke the timestamp consume function in tx done instead
of scheduling through workqueue. Schedule the work queue only
if the timestamp is not found.
Bug 3658701
Change-Id: If97aebc1980f685e3cc34f135ace91298ee78657
Signed-off-by: Bhadram Varka <vbhadram@nvidia.com>
Reviewed-on: https://git-master.nvidia.com/r/c/linux-nvidia/+/2732444
Reviewed-by: Troy Kong <troyk@nvidia.com>
Reviewed-by: Narayan Reddy <narayanr@nvidia.com>
Reviewed-by: svc-mobile-coverity <svc-mobile-coverity@nvidia.com>
Reviewed-by: svc_kernel_abi <svc_kernel_abi@nvidia.com>
Reviewed-by: Rakesh Goyal <rgoyal@nvidia.com>
Reviewed-by: Bitan Biswas <bbiswas@nvidia.com>
Tested-by: Troy Kong <troyk@nvidia.com>
GVS: Gerrit_Virtual_Submit
Dma phy address, virtual address and packet length can be
obtained from tx swcx structure. Hence passing
pointer to tx swcx is sufficient. In future, if more information
from osi is needed, it can be embedded into tx swcx itself rather than
adding more arguments to osd tx complete call back.
Bug 3576506
Change-Id: I3b3aca993c52359668c41f337a79ea4d0d1854fb
Signed-off-by: Hareesh Kesireddy <hkesireddy@nvidia.com>
Reviewed-on: https://git-master.nvidia.com/r/c/linux-nvidia/+/2700392
Tested-by: mobile promotions <svcmobile_promotions@nvidia.com>
Reviewed-by: mobile promotions <svcmobile_promotions@nvidia.com>
Issue:
- Workqueue for Tx timestamp is running without
delay, use many CPU cycle at hard isr thread
priority.
- Incorrect return value from ether_handle_tso()
Fix:
- Update code to use delayed workqueue
- Return correct value
Bug 200780891
Bug 3400623
Change-Id: I2095a0634e079bf870ef87cd2de1d35fe24bafd4
Signed-off-by: Rakesh Goyal <rgoyal@nvidia.com>
Reviewed-on: https://git-master.nvidia.com/r/c/linux-nvidia/+/2610986
Tested-by: mobile promotions <svcmobile_promotions@nvidia.com>
Reviewed-by: mobile promotions <svcmobile_promotions@nvidia.com>
Algorithm:
- There is one master queue and one final queue to maintain the
packets.
- Final queue will start when a packet received with TTL=1
- TTL = 2 and 3 packets will be queued into final if those are in
sequence else those will be queued to master queue.
- When packet received with TTL = 2 then only merge process will start
- For every valid final queue merge starts by looking at the IPID
in master queue in increment order. This loop continues until
TTL = 2 found in master queue.
- Once TTL = 2 found then entire final queue will be given GRO
layer in sequential order.
Bug 3106903
Change-Id: I6cc08f3faae67a04813d32d59f7cce02f06dccbf
Signed-off-by: Bhadram Varka <vbhadram@nvidia.com>
Reviewed-on: https://git-master.nvidia.com/r/c/linux-nvidia/+/2554423
Tested-by: mobile promotions <svcmobile_promotions@nvidia.com>
Reviewed-by: svcacv <svcacv@nvidia.com>
Reviewed-by: Nagarjuna Kristam <nkristam@nvidia.com>
Reviewed-by: Narayan Reddy <narayanr@nvidia.com>
Reviewed-by: Sachin Nikam <snikam@nvidia.com>
Reviewed-by: mobile promotions <svcmobile_promotions@nvidia.com>
- On transmit complete callback
-- If OSI_TXDONE_CX_TS_DELAY flag is set in Tx done
flag, store buffer pointer along with pkt Id
in its internal array/list
- On Common interrupt:
-- Call osi to handle common interrupt.
- On trasmit complete interrupt:
-- Parse through internal array/list for pending TS
to read, by issuing OSI_CMD_GET_TS ioctl for pkt_id.
Loop until list is cleared or error is received.
-- On interface up
Remove old timestamp of channels initialized.
Signed-off-by: Rakesh Goyal <rgoyal@nvidia.com>
Change-Id: I5104277de66dd240023f921fa5f7b4fcc035f74d
Reviewed-on: https://git-master.nvidia.com/r/c/linux-nvidia/+/2535962
Reviewed-by: svc_kernel_abi <svc_kernel_abi@nvidia.com>
Reviewed-by: Nagarjuna Kristam <nkristam@nvidia.com>
Reviewed-by: Bhadram Varka <vbhadram@nvidia.com>
Reviewed-by: Sachin Nikam <snikam@nvidia.com>
Reviewed-by: mobile promotions <svcmobile_promotions@nvidia.com>
Tested-by: mobile promotions <svcmobile_promotions@nvidia.com>
GVS: Gerrit_Virtual_Submit
Issue: DMA will be in suspend state if there is no rx buffer allocation
happen to HW. If DMA is in suspend state, no interrupt will
occur for channel which leads to not calling of
osd_receive_packet().
Fix: Whenever there is rx buffer kernel memory allocation failure,
desc will be updated with reserve buffer. This reserved RX buffer
is already allocated at the time of eqos_open. Aim of same to
make sure rx DMA always have buffer and not go in suspend state.
Bug 200650229
Change-Id: Ic18d9f8f03c69c142d91288a1a7288d26e1c8b42
Signed-off-by: rakesh goyal <rgoyal@nvidia.com>
Reviewed-on: https://git-master.nvidia.com/r/c/linux-nvidia/+/2399011
Reviewed-by: Bhadram Varka <vbhadram@nvidia.com>
Reviewed-by: Bitan Biswas <bbiswas@nvidia.com>
Reviewed-by: mobile promotions <svcmobile_promotions@nvidia.com>
Tested-by: mobile promotions <svcmobile_promotions@nvidia.com>
GVS: Gerrit_Virtual_Submit
Issue: osi_process_tx_completions is run in the BH for
processing Tx done irq. This function always loops
till the Tx status for all packets that are Tx'd so
far is processed.
1) The NAPI budget for this Tx completions handler is
not honored.
2) The condition to invoke napi_complete() is incorrect.
Currently, processed count should be 0, which can happen
only if BH is scheduled to run and it finds no packets that
have completed Tx. THis is a wasted schedule for the BH.
3) netdev stats for Tx packets is incremented for every Tx
descriptor instead of packet.
Fix: 1) Add budget parameter to osi_process_tx_completions() API.
2) Change condition to invoke napi_complete() to avoid wasted
scheduling of the napi handler.
3) Increment the netdev stats only when skb is consumed ie.
entire packet is transmitted).
Bug 2756439
Change-Id: I273bdb2667112b8d252b119dee5cc92055cb2913
Signed-off-by: Srinivas Ramachandran <srinivasra@nvidia.com>
Reviewed-on: https://git-master.nvidia.com/r/2237821
Reviewed-by: mobile promotions <svcmobile_promotions@nvidia.com>
Tested-by: mobile promotions <svcmobile_promotions@nvidia.com>
