The recent changes in Broadcom's ethernet driver(L2 driver) broke
RoCE functionality in terms of MSIx vector allocation and
de-allocation.
There is a possibility that L2 driver would initiate MSIx vector
reallocation depending upon the requests coming from administrator.
In such cases L2 driver needs to free up all the MSIx vectors
allocated previously and reallocate/initialize those.
If RoCE driver is loaded and reshuffling is attempted, there will be
kernel crashes because RoCE driver would still be holding the MSIx
vectors but L2 driver would attempt to free in-use vectors. Thus
leading to a kernel crash.
Making changes in roce driver to fix crashes described above.
As part of solution L2 driver tells RoCE driver to release
the MSIx vector whenever there is a need. When RoCE driver
get message it sync up with all the running tasklets and IRQ
handlers and releases the vectors. L2 driver send one more
message to RoCE driver to resume the MSIx vectors. L2 driver
guarantees that RoCE vector do not change during reshuffling.
Fixes: ec86f14ea5 ("bnxt_en: Add ULP calls to stop and restart IRQs.")
Fixes: 08654eb213 ("bnxt_en: Change IRQ assignment for RDMA driver.")
Signed-off-by: Devesh Sharma <devesh.sharma@broadcom.com>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
Hitting the following hardlockup due to a race condition in
error CQE processing.
[26146.879798] bnxt_en 0000:04:00.0: QPLIB: FP: CQ Processed Req
[26146.886346] bnxt_en 0000:04:00.0: QPLIB: wr_id[1251] = 0x0 with status 0xa
[26156.350935] NMI watchdog: Watchdog detected hard LOCKUP on cpu 4
[26156.357470] Modules linked in: nfsd auth_rpcgss nfs_acl lockd grace
[26156.447957] CPU: 4 PID: 3413 Comm: kworker/4:1H Kdump: loaded
[26156.457994] Hardware name: Dell Inc. PowerEdge R430/0CN7X8,
[26156.466390] Workqueue: ib-comp-wq ib_cq_poll_work [ib_core]
[26156.472639] Call Trace:
[26156.475379] <NMI> [<ffffffff98d0d722>] dump_stack+0x19/0x1b
[26156.481833] [<ffffffff9873f775>] watchdog_overflow_callback+0x135/0x140
[26156.489341] [<ffffffff9877f237>] __perf_event_overflow+0x57/0x100
[26156.496256] [<ffffffff98787c24>] perf_event_overflow+0x14/0x20
[26156.502887] [<ffffffff9860a580>] intel_pmu_handle_irq+0x220/0x510
[26156.509813] [<ffffffff98d16031>] perf_event_nmi_handler+0x31/0x50
[26156.516738] [<ffffffff98d1790c>] nmi_handle.isra.0+0x8c/0x150
[26156.523273] [<ffffffff98d17be8>] do_nmi+0x218/0x460
[26156.528834] [<ffffffff98d16d79>] end_repeat_nmi+0x1e/0x7e
[26156.534980] [<ffffffff987089c0>] ? native_queued_spin_lock_slowpath+0x1d0/0x200
[26156.543268] [<ffffffff987089c0>] ? native_queued_spin_lock_slowpath+0x1d0/0x200
[26156.551556] [<ffffffff987089c0>] ? native_queued_spin_lock_slowpath+0x1d0/0x200
[26156.559842] <EOE> [<ffffffff98d083e4>] queued_spin_lock_slowpath+0xb/0xf
[26156.567555] [<ffffffff98d15690>] _raw_spin_lock+0x20/0x30
[26156.573696] [<ffffffffc08381a1>] bnxt_qplib_lock_buddy_cq+0x31/0x40 [bnxt_re]
[26156.581789] [<ffffffffc083bbaa>] bnxt_qplib_poll_cq+0x43a/0xf10 [bnxt_re]
[26156.589493] [<ffffffffc083239b>] bnxt_re_poll_cq+0x9b/0x760 [bnxt_re]
The issue happens if RQ poll_cq or SQ poll_cq or Async error event tries to
put the error QP in flush list. Since SQ and RQ of each error qp are added
to two different flush list, we need to protect it using locks of
corresponding CQs. Difference in order of acquiring the lock in
SQ poll_cq and RQ poll_cq can cause a hard lockup.
Revisits the locking strategy and removes the usage of qplib_cq.hwq.lock.
Instead of this lock, introduces qplib_cq.flush_lock to handle
addition/deletion of QPs in flush list. Also, always invoke the flush_lock
in order (SQ CQ lock first and then RQ CQ lock) to avoid any potential
deadlock.
Other than the poll_cq context, the movement of QP to/from flush list can
be done in modify_qp context or from an async error event from HW.
Synchronize these operations using the bnxt_re verbs layer CQ locks.
To achieve this, adds a call back to the HW abstraction layer(qplib) to
bnxt_re ib_verbs layer in case of async error event. Also, removes the
buddy cq functions as it is no longer required.
Signed-off-by: Sriharsha Basavapatna <sriharsha.basavapatna@broadcom.com>
Signed-off-by: Somnath Kotur <somnath.kotur@broadcom.com>
Signed-off-by: Devesh Sharma <devesh.sharma@broadcom.com>
Signed-off-by: Selvin Xavier <selvin.xavier@broadcom.com>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
Avoid system crash when destroy_qp is invoked while
the driver is processing the poll_cq. Synchronize these
functions using the cq_lock.
Signed-off-by: Selvin Xavier <selvin.xavier@broadcom.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>
Shared receive queue (SRQ) is defined as a pool of
receive buffers shared among multiple QPs which belong
to same protection domain in a given process context.
Use of SRQ reduces the memory foot print of IB applications.
Broadcom adapters support SRQ, adding code-changes to enable
shared receive queue.
Signed-off-by: Devesh Sharma <devesh.sharma@broadcom.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>
Destroy_qp shall wait for any outstanding CQ notification to be
flushed out before proceeding with QP destroy. Flushing the WQ
before destroying the QP.
Signed-off-by: Selvin Xavier <selvin.xavier@broadcom.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>
Enables multiple Interrupt vectors. Driver is requesting the max
MSIX vectors based on the number of online cpus and creates upto
9 MSIx vectors (1 for control path and 8 for data path).
A tasklet is created for each of these vectors. NQs are assigned
to CQs in round robin fashion.
This patch also adds IRQ affinity hint for the MSIX vector of each NQ.
Signed-off-by: Ray Jui <ray.jui@broadcom.com>
Signed-off-by: Selvin Xavier <selvin.xavier@broadcom.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>
This patch allows driver to post send and receive
requests on QPs which are in error state.
Instead of flushing the QP in the context of polling
error CQEs, the QPs will be added to a flush list
maintained per CQ. QP state is moved to error.
QP is added to flush list if the user moves it
to error state using modify_qp also. After polling the HW
CQ in poll_cq routine, this flush list is traversed
and driver completes work requests on each QP in the flush
list, till the budget expires. The QP is moved out of
flush list during QP destroy or during modify_QP to RESET.
When ULPs post Work Requests while QP is in error state,
driver will store the ULP data and then increment the
QP producer s/w index, without ringing doorbell. It then
schedules a worker to invoke the CQ handler since the
interrupts wont be generated from the HW for this request.
Signed-off-by: Sriharsha Basavapatna <sriharsha.basavapatna@broadcom.com>
Signed-off-by: Selvin Xavier <selvin.xavier@broadcom.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>
While invoking the req_notify_cq hook, ULPs can request
whether the CQs have any CQEs pending. If CQEs are pending,
drivers can indicate it by returning 1 for req_notify_cq.
The stack will poll CQ again till CQ is empty.
This patch peeks the CQ for any valid entries and return accordingly.
Signed-off-by: Selvin Xavier <selvin.xavier@broadcom.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>
HW stalls out after 0x800000 WQEs are posted for UD QPs.
To workaround this problem, driver will send a modify_qp cmd
to the HW at around the halfway mark(0x400000) so that FW
can accordingly modify the QP context in the HW to prevent this
stall.
This workaround needs to be done for UD, QP1 and Raw Ethertype
packets. Added a counter to keep track of WQEs posted during post_send.
Signed-off-by: Somnath Kotur <somnath.kotur@broadcom.com>
Signed-off-by: Selvin Xavier <selvin.xavier@broadcom.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>
This patch implements the following HW workarounds
1. The SQ depth needs to be augmented by 128 + 1 to avoid running
into an Out of order CQE issue
2. Workaround to handle the problem where the HW fast path engine continues
to access DMA memory in retranmission mode even after the WQE has
already been completed. If the HW reports this condition, driver detects
it and posts a Fence WQE. The driver stops reporting the completions
to stack until it receives completion for Fence WQE.
Signed-off-by: Eddie Wai <eddie.wai@broadcom.com>
Signed-off-by: Sriharsha Basavapatna <sriharsha.basavapatna@broadcom.com>
Signed-off-by: Selvin Xavier <selvin.xavier@broadcom.com>
Reviewed-by: Leon Romanovsky <leonro@mellanox.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>
This patch introduces the RoCE driver for the Broadcom
NetXtreme-E 10/25/40/50G RoCE HCAs.
The RoCE driver is a two part driver that relies on the parent
bnxt_en NIC driver to operate. The changes needed in the bnxt_en
driver have already been incorporated via Dave Miller's net tree
into the mainline kernel.
The vendor official git repository for this driver is available
on github as:
https://github.com/Broadcom/linux-rdma-nxt/
Signed-off-by: Eddie Wai <eddie.wai@broadcom.com>
Signed-off-by: Devesh Sharma <devesh.sharma@broadcom.com>
Signed-off-by: Somnath Kotur <somnath.kotur@broadcom.com>
Signed-off-by: Sriharsha Basavapatna <sriharsha.basavapatna@broadcom.com>
Signed-off-by: Selvin Xavier <selvin.xavier@broadcom.com>
Reviewed-by: Leon Romanovsky <leonro@mellanox.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>
