Start switching iomap_copy routines over to use #define and arch provided
inline/macro functions instead of weak symbols.
Inline functions allow more compiler optimization and this is often a
driver hot path.
x86 has the only weak implementation for __iowrite32_copy(), so replace it
with a static inline containing the same single instruction inline
assembly. The compiler will generate the "mov edx,ecx" in a more optimal
way.
Remove iomap_copy_64.S
Link: https://lore.kernel.org/r/1-v3-1893cd8b9369+1925-mlx5_arm_wc_jgg@nvidia.com
Acked-by: Arnd Bergmann <arnd@arndb.de>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
In some situations a sent packet may get queued in the NIC longer than
than timeout of a ULP. Currently if this happens the ULP may try to reset
the link by destroying the qp and setting up an alternate connection but
will fail because the rxe driver is waiting for the packet to finish
getting sent and be returned to the skb destructor function where the qp
reference holding things up will be dropped. This patch modifies the way
that the qp is passed to the destructor to pass the qp index and not a qp
pointer. Then the destructor will attempt to lookup the qp from its index
and if it fails exit early. This requires taking a reference on the struct
sock rather than the qp allowing the qp to be destroyed while the sk is
still around waiting for the packet to finish.
Link: https://lore.kernel.org/r/20240329145513.35381-15-rpearsonhpe@gmail.com
Signed-off-by: Bob Pearson <rpearsonhpe@gmail.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
Currently the rxe_driver detects packet drops by ip_local_out() which
occur before the packet is sent on the wire and attempts to resend
them. This is redundant with the usual retry mechanism which covers
packets that get dropped in transit to or from the remote node.
The way this is implemented is not robust since it sets need_req_skb and
waits for the number of local skbs outstanding for this qp to drop below a
low water mark. This is racy since the skb may be sent to the destructor
before the requester can set the need_req_skb flag. This will cause a
deadlock in the send path for that qp.
This patch removes this mechanism since the normal retry path will correct
the error and resend the packet and it makes no difference if the packet
is dropped locally or later.
Link: https://lore.kernel.org/r/20240329145513.35381-14-rpearsonhpe@gmail.com
Signed-off-by: Bob Pearson <rpearsonhpe@gmail.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
The rxe send path currently counts the number of skbs outstanding between
the rxe driver and the ethernet driver to prevent too many packets to
accumulate waiting to send. This patch makes the local loopback path
behave the same way. The loopback path forwards the packets to the receive
path which will eventually call kfree_skb on all packets and drop the qp
references. This makes the loopback path more useful for software testing.
Link: https://lore.kernel.org/r/20240329145513.35381-13-rpearsonhpe@gmail.com
Signed-off-by: Bob Pearson <rpearsonhpe@gmail.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
In rxe_send() a ref is taken on the qp to keep it alive until the
kfree_skb() has a chance to call the skb destructor rxe_skb_tx_dtor()
which drops the reference. If the packet has an incorrect protocol the
error path just calls kfree_skb() which will call the destructor which
will drop the ref. Currently the driver also calls rxe_put() which is
incorrect. Additionally since the packets sent to rxe_send() are under the
control of the driver and it only ever produces IPV4 or IPV6 packets the
simplest fix is to remove all the code in this block.
Link: https://lore.kernel.org/r/20240329145513.35381-12-rpearsonhpe@gmail.com
Signed-off-by: Bob Pearson <rpearsonhpe@gmail.com>
Fixes: 9eb7f8e44d ("IB/rxe: Move refcounting earlier in rxe_send()")
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
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>
Now that req.task and comp.task are merged it is no longer necessary to
call save_state() before calling rxe_xmit_pkt() and rollback_state() if
rxe_xmit_pkt() fails. This was done originally to prevent races between
rxe_completer() and rxe_requester() which now cannot happen.
Link: https://lore.kernel.org/r/20240329145513.35381-8-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_post_send_kernel() if the qp is in the error state after posting
the work requests the rxe_completer() task is scheduled.
But, the only way to move the qp into the error state is to call
rxe_qp_error() which also schedules the rxe_completer() task to drain the
queues. Calling it a second time has no effect. This commit removes the
redundant call.
Link: https://lore.kernel.org/r/20240329145513.35381-6-rpearsonhpe@gmail.com
Signed-off-by: Bob Pearson <rpearsonhpe@gmail.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
A previous commit incorrectly added an 'if(!err)' before scheduling the
requester task in rxe_post_send_kernel(). But if there were send wrs
successfully added to the send queue before a bad wr they might never get
executed.
This commit fixes this by scheduling the requester task if any wqes were
successfully posted in rxe_post_send_kernel() in rxe_verbs.c.
Link: https://lore.kernel.org/r/20240329145513.35381-5-rpearsonhpe@gmail.com
Signed-off-by: Bob Pearson <rpearsonhpe@gmail.com>
Fixes: 5bf944f241 ("RDMA/rxe: Add error messages")
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>
The existing behavior of ib_umad, which maintains received MAD
packets in an unbounded list, poses a risk of uncontrolled growth.
As user-space applications extract packets from this list, the rate
of extraction may not match the rate of incoming packets, leading
to potential list overflow.
To address this, we introduce a limit to the size of the list. After
considering typical scenarios, such as OpenSM processing, which can
handle approximately 100k packets per second, and the 1-second retry
timeout for most packets, we set the list size limit to 200k. Packets
received beyond this limit are dropped, assuming they are likely timed
out by the time they are handled by user-space.
Notably, packets queued on the receive list due to reasons like
timed-out sends are preserved even when the list is full.
Signed-off-by: Michael Guralnik <michaelgur@nvidia.com>
Reviewed-by: Mark Zhang <markzhang@nvidia.com>
Link: https://lore.kernel.org/r/7197cb58a7d9e78399008f25036205ceab07fbd5.1713268818.git.leon@kernel.org
Signed-off-by: Leon Romanovsky <leon@kernel.org>
