Replace calls to rxe_run_task() with rxe_sched_task(). This prevents the
tasks from all running on the same cpu.
This change slightly reduces performance for single qp send and write
benchmarks in loopback mode but greatly improves the performance with
multiple qps because if run task is used all the work tends to be
performed on one cpu. For actual on the wire benchmarks there is no
noticeable performance change.
Link: https://lore.kernel.org/r/20240329145513.35381-11-rpearsonhpe@gmail.com
Signed-off-by: Bob Pearson <rpearsonhpe@gmail.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
Currently the rxe driver has three work queue tasks per qp. These are the
req.task, comp.task and resp.task which call rxe_requester(),
rxe_completer() and rxe_responder() respectively directly or on work
queues. Each of these subroutines checks to see if there is work to be
performed on the send queue or on the response packet queue or the request
packet queue and will run until there is no work remaining or yield the
cpu and reschedule itself until there is no work remaining.
This commit combines the req.task and comp.task into a single send.task
and renames the resp.task to the recv.task. The combined send.task calls
rxe_requester() and rxe_completer() serially and continues until all work
on both the send queue and the response packet queue are done.
In various benchmarks the performance is either improved or left the
same. At high scale there is a significant reduction in the load on the
cpu.
This is the first step in combining these two tasks. Once they are
serialized cross rescheduling of req.task and comp.task can be more
efficiently handled by just letting the send.task continue to run. This
will be done in the next several patches.
Link: https://lore.kernel.org/r/20240329145513.35381-7-rpearsonhpe@gmail.com
Signed-off-by: Bob Pearson <rpearsonhpe@gmail.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
In rxe_comp_queue_pkt() an incoming response packet skb is enqueued to the
resp_pkts queue and then a decision is made whether to run the completer
task inline or schedule it. Finally the skb is dereferenced to bump a 'hw'
performance counter. This is wrong because if the completer task is
already running in a separate thread it may have already processed the skb
and freed it which can cause a seg fault. This has been observed
infrequently in testing at high scale.
This patch fixes this by changing the order of enqueuing the packet until
after the counter is accessed.
Link: https://lore.kernel.org/r/20240329145513.35381-4-rpearsonhpe@gmail.com
Signed-off-by: Bob Pearson <rpearsonhpe@gmail.com>
Fixes: 0b1e5b99a4 ("IB/rxe: Add port protocol stats")
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
Currently the rxe driver makes little effort to make the changes to qp
state (which includes qp->attr.qp_state, qp->attr.sq_draining and
qp->valid) atomic between different client threads and IO threads. In
particular a common template is for an RDMA application to call
ib_modify_qp() to move a qp to ERR state and then wait until all the
packet and work queues have drained before calling ib_destroy_qp(). None
of these state changes are protected by locks to assure that the changes
are executed atomically and that memory barriers are included. This has
been observed to lead to incorrect behavior around qp cleanup.
This patch continues the work of the previous patches in this series and
adds locking code around qp state changes and lookups.
Link: https://lore.kernel.org/r/20230405042611.6467-5-rpearsonhpe@gmail.com
Signed-off-by: Bob Pearson <rpearsonhpe@gmail.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
The rxe driver has four different QP state variables,
qp->attr.qp_state,
qp->req.state,
qp->comp.state, and
qp->resp.state.
All of these basically carry the same information.
This patch replaces uses of qp->req.state by qp->attr.qp_state and enum
rxe_qp_state. This is the third of three patches which will remove all
but the qp->attr.qp_state variable. This will bring the driver closer to
the IBA description.
Link: https://lore.kernel.org/r/20230405042611.6467-3-rpearsonhpe@gmail.com
Signed-off-by: Bob Pearson <rpearsonhpe@gmail.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
The rxe driver has four different QP state variables,
qp->attr.qp_state,
qp->req.state,
qp->comp.state, and
qp->resp.state.
All of these basically carry the same information.
This patch replaces uses of qp->comp.state by qp->attr.qp_state. This is
the second of three patches which will remove all but the
qp->attr.qp_state variable. This will bring the driver closer to the IBA
description.
Link: https://lore.kernel.org/r/20230405042611.6467-2-rpearsonhpe@gmail.com
Signed-off-by: Bob Pearson <rpearsonhpe@gmail.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
Currently each of the three tasklets requester, completer and responder in
the rxe driver take and release a reference to the qp argument at the
beginning and end of the subroutines. The caller passing in the qp
argument should be responsible for holding a reference to qp so these are
not required. Further doing so breaks the qp cleanup code in
rxe_qp_do_cleanup which calls these routines after all the references have
been dropped so they cannot drain the packet and work request queues as
intended.
In fact if these routines are deferred by calling tasklet_schedule there
is no guarantee that the calling code does have a qp reference. That is a
bug in rxe_task.c which will be fixed later in this series.
Link: https://lore.kernel.org/r/20230304174533.11296-6-rpearsonhpe@gmail.com
Signed-off-by: Bob Pearson <rpearsonhpe@gmail.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
Originally is was thought that the tasklet machinery in rxe_task.c would
be used in other applications but that has not happened for years. This
patch replaces the 'void *arg' by struct 'rxe_qp *qp' in the parameters to
the tasklet calls. This change will have no affect on performance but may
make the code a little clearer.
Link: https://lore.kernel.org/r/20230304174533.11296-2-rpearsonhpe@gmail.com
Signed-off-by: Bob Pearson <rpearsonhpe@gmail.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
This patch adds error and debug messages so that every interaction
with rdma-core through a verbs API call or a completion error return
will generate at least one error message backed up by debug messages
with more detail.
With dynamic debugging one can follow up after seeing an error message
by turning on the appropriate debug messages.
Link: https://lore.kernel.org/r/20230303221623.8053-5-rpearsonhpe@gmail.com
Signed-off-by: Bob Pearson <rpearsonhpe@gmail.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
Currently the requester can continue to process send wqes after an local
qp operation error is detected because the setting of the qp state to the
error state is deferred until later. This patch splits the qp state for
the completer and requester into two separate states and sets
qp->req.state = QP_STATE_ERROR as soon as the error is detected before
another wqe can be executed.
Link: https://lore.kernel.org/r/1658307368-1851-4-git-send-email-lizhijian@fujitsu.com
Signed-off-by: Bob Pearson <rpearsonhpe@gmail.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
In the tasklets (completer, responder, and requester) check the return
value from rxe_get() to detect failures to get a reference. This only
occurs if the qp has had its reference count drop to zero which indicates
that it no longer should be used.
The ref is never 0 today because the tasklets are flushed before the ref
is dropped. The next patch changes this so that the ref is dropped then
the tasklets are flushed.
Link: https://lore.kernel.org/r/20220421014042.26985-4-rpearsonhpe@gmail.com
Signed-off-by: Bob Pearson <rpearsonhpe@gmail.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
Most of the locks in the rxe driver are _irqsave/restore locks but in fact
there are no interrupt threads that run rxe code or share data with
rxe. There are softirq threads and data sharing so the appropriate lock
type is _bh. This patch replaces all irqsave type locks with bh type
locks.
Link: https://lore.kernel.org/r/20211103050241.61293-2-rpearsonhpe@gmail.com
Signed-off-by: Bob Pearson <rpearsonhpe@gmail.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
Earlier patches added memory barriers to protect user space to kernel
space communications. The user space queues were previously shown to have
occasional memory synchonization errors which were removed by adding
smp_load_acquire, smp_store_release barriers. This patch extends that to
the case where queues are used between kernel space threads.
This patch also extends the queue types to include kernel ULP queues which
access the other end of the queues in kernel verbs calls like poll_cq and
post_send/recv.
Link: https://lore.kernel.org/r/20210914164206.19768-2-rpearsonhpe@gmail.com
Signed-off-by: Bob Pearson <rpearsonhpe@gmail.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
Linux 5.13-rc7
Needed for dependencies in following patches. Merge conflict in rxe_cmop.c
resolved by compining both patches.
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
Add support for bind MW work requests from user space. Since rdma/core
does not support bind mw in ib_send_wr there is no way to support bind mw
in kernel space.
Added bind_mw local operation in rxe_req.c. Added bind_mw WR operation in
rxe_opcode.c. Added bind_mw WC in rxe_comp.c. Added additional fields to
rxe_mw in rxe_verbs.h. Added rxe_do_dealloc_mw() subroutine to cleanup an
mw when rxe_dealloc_mw is called. Added code to implement bind_mw
operation in rxe_mw.c
Link: https://lore.kernel.org/r/20210608042552.33275-8-rpearsonhpe@gmail.com
Signed-off-by: Bob Pearson <rpearsonhpe@gmail.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
In order to prevent user space from modifying the index that belongs to
the kernel for shared queues let the kernel use a local copy of the index
and copy any new values of that index to the shared rxe_queue_bus struct.
This adds more switch statements which decreases the performance of the
queue API. Move the type into the parameter list for these functions so
that the compiler can optimize out the switch statements when the explicit
type is known. Modify all the calls in the driver on performance paths to
pass in the explicit queue type.
Link: https://lore.kernel.org/r/20210527194748.662636-4-rpearsonhpe@gmail.com
Link: https://lore.kernel.org/linux-rdma/20210526165239.GP1002214@@nvidia.com/
Signed-off-by: Bob Pearson <rpearsonhpe@gmail.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
In the original rxe implementation it was intended to use a common object
to represent MRs and MWs but they are different enough to separate these
into two objects.
This allows replacing the mem name with mr for MRs which is more
consistent with the style for the other objects and less likely to be
confusing. This is a long patch that mostly changes mem to mr where it
makes sense and adds a new rxe_mw struct.
Link: https://lore.kernel.org/r/20210325212425.2792-1-rpearson@hpe.com
Signed-off-by: Bob Pearson <rpearson@hpe.com>
Acked-by: Zhu Yanjun <zyjzyj2000@gmail.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
In rxe_comp.c in rxe_completer() the function free_pkt() did not clear skb
which triggered a warning at 'done:' and could possibly at 'exit:'. The
WARN_ONCE() calls are not actually needed. The call to free_pkt() is
moved to the end to clearly show that all skbs are freed.
Fixes: 899aba891c ("RDMA/rxe: Fix FIXME in rxe_udp_encap_recv()")
Link: https://lore.kernel.org/r/20210304192048.2958-1-rpearson@hpe.com
Signed-off-by: Bob Pearson <rpearsonhpe@gmail.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
rxe_udp_encap_recv() drops the reference to rxe->ib_dev taken by
rxe_get_dev_from_net() which should be held until each received skb is
freed. This patch moves the calls to ib_device_put() to each place a
received skb is freed. It also takes references to the ib_device for each
cloned skb created to process received multicast packets.
Fixes: 4c173f596b ("RDMA/rxe: Use ib_device_get_by_netdev() instead of open coding")
Link: https://lore.kernel.org/r/20210128233318.2591-1-rpearson@hpe.com
Signed-off-by: Bob Pearson <rpearson@hpe.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
When run stress tests with RXE, the following Call Traces often occur
watchdog: BUG: soft lockup - CPU#2 stuck for 22s! [swapper/2:0]
...
Call Trace:
<IRQ>
create_object+0x3f/0x3b0
kmem_cache_alloc_node_trace+0x129/0x2d0
__kmalloc_reserve.isra.52+0x2e/0x80
__alloc_skb+0x83/0x270
rxe_init_packet+0x99/0x150 [rdma_rxe]
rxe_requester+0x34e/0x11a0 [rdma_rxe]
rxe_do_task+0x85/0xf0 [rdma_rxe]
tasklet_action_common.isra.21+0xeb/0x100
__do_softirq+0xd0/0x298
irq_exit+0xc5/0xd0
smp_apic_timer_interrupt+0x68/0x120
apic_timer_interrupt+0xf/0x20
</IRQ>
...
The root cause is that tasklet is actually a softirq. In a tasklet
handler, another softirq handler is triggered. Usually these softirq
handlers run on the same cpu core. So this will cause "soft lockup Bug".
Fixes: 8700e3e7c4 ("Soft RoCE driver")
Link: https://lore.kernel.org/r/20200212072635.682689-8-leon@kernel.org
Signed-off-by: Zhu Yanjun <yanjunz@mellanox.com>
Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
In some cases (not in this particular one) variable self-initialization
can lead to undefined behavior. In this case, it is just obscure code.
Signed-off-by: Maksym Planeta <mplaneta@os.inf.tu-dresden.de>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
The variable rxe in the function is not used. So it is removed.
Signed-off-by: Zhu Yanjun <yanjun.zhu@oracle.com>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
link_down is self-explanatory.
rdma_sends and rdma_recvs count the number of RDMA Send and RDMA Receive
operations completed successfully. This is different from the existing
sent_pkts and rcvd_pkts counters because the existing counters measure
packets, not RDMA operations.
ack_deffered is renamed to ack_deferred to fix the spelling.
out_of_sequence is renamed to out_of_seq_request to make clear that it is
counting only requests and not other packets which can be out of sequence.
Signed-off-by: Andrew Boyer <andrew.boyer@dell.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
When a read request is retried for the remaining partial
data, the response may restart from read response first
or read response only. So support those cases.
Do not advance the comp psn beyond the current wqe's last_psn
as that could skip over an entire read wqe and will cause the
req_retry() logic to set an incorrect req psn.
An example sequence is as follows:
Write PSN 40 -- this is the current WQE.
Read request PSN 41
Write PSN 42
Receive ACK PSN 42 -- this will complete the current WQE
for PSN 40, and set the comp psn to 42 which is a problem
because the read request at PSN 41 has been skipped over.
So when req_retry() tries to retransmit the read request,
it sets the req psn to 42 which is incorrect.
When retrying a read request, calculate the number of psns
completed based on the dma resid instead of the wqe first_psn.
The wqe first_psn could have moved if the read request was
retried multiple times.
Set the reth length to the dma resid to handle read retries for
the remaining partial data.
Signed-off-by: Vijay Immanuel <vijayi@attalasystems.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>
