Remove duplicate settings for vendor_err and qp_num.

Link: https://lore.kernel.org/r/20210930094813.226888-5-yangx.jy@fujitsu.comSigned-off-by: NXiao Yang <yangx.jy@fujitsu.com>
Signed-off-by: NJason Gunthorpe <jgg@nvidia.com>

To reflect the intention, since it is not just a single bit.

Link: https://lore.kernel.org/r/20210914080253.1145353-3-yangx.jy@fujitsu.comSigned-off-by: NXiao Yang <yangx.jy@fujitsu.com>
Signed-off-by: NJason Gunthorpe <jgg@nvidia.com>

1) Replace (RXE_READ_MASK | RXE_WRITE_MASK) with RXE_READ_OR_WRITE_MASK.
2) Change (RXE_READ_MASK | RXE_WRITE_OR_SEND) to RXE_READ_OR_WRITE_MASK
   because we don't need to check RETH for RXE_SEND_MASK.

Link: https://lore.kernel.org/r/20210914080253.1145353-2-yangx.jy@fujitsu.comSigned-off-by: NXiao Yang <yangx.jy@fujitsu.com>
Signed-off-by: NJason 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.comSigned-off-by: NBob Pearson <rpearsonhpe@gmail.com>
Signed-off-by: NJason Gunthorpe <jgg@nvidia.com>

The memcpy() that copies a WQE from a SRQ the QP uses an incorrect size.
The size should have been the size of the rxe_send_wqe struct not the size
of a pointer to it. The result is that IO operations using a SRQ on the
responder side will fail.

Fixes: ec0fa244 ("RDMA/rxe: Fix over copying in get_srq_wqe")
Link: https://lore.kernel.org/r/20210729220039.18549-2-rpearsonhpe@gmail.comSigned-off-by: NBob Pearson <rpearsonhpe@gmail.com>
Signed-off-by: NJason Gunthorpe <jgg@nvidia.com>

Isolate ICRC generation into a single subroutine named rxe_generate_icrc()
in rxe_icrc.c. Remove scattered crc generation code from elsewhere.

Link: https://lore.kernel.org/r/20210707040040.15434-5-rpearsonhpe@gmail.comSigned-off-by: NBob Pearson <rpearsonhpe@gmail.com>
Signed-off-by: NJason Gunthorpe <jgg@nvidia.com>

This error path needs to unlock before returning.

Fixes: ec0fa244 ("RDMA/rxe: Fix over copying in get_srq_wqe")
Link: https://lore.kernel.org/r/YNXUCmnPsSkPyhkm@mwandaSigned-off-by: NDan Carpenter <dan.carpenter@oracle.com>
Reviewed-by: NMajd Dibbiny <majd@nvidia.com>
Reviewed-by: NBob Pearson <rpearsonhpe@gmail.com>
Signed-off-by: NJason Gunthorpe <jgg@nvidia.com>

Currently prepare_ack_packet writes almost all the fields of the BTH in
the ack packet twice. Replace code with the subroutine init_bth().

Fixes: 8700e3e7 ("Soft RoCE driver")
Link: https://lore.kernel.org/r/20210618045742.204195-6-rpearsonhpe@gmail.comSigned-off-by: NBob Pearson <rpearsonhpe@gmail.com>
Signed-off-by: NJason Gunthorpe <jgg@nvidia.com>

Currently get_srq_wqe() in rxe_resp.c copies the maximum possible number
of bytes from the wqe into the QPs copy of the SRQ wqe. This is usually
extra work and risks reading past the end of the SRQ circular buffer if
the SRQ is configured with less than the maximum possible number of SGEs.

Check the number of SGEs is not too large.
Compute the actual number of bytes in the WR and copy only those.

Fixes: 8700e3e7 ("Soft RoCE driver")
Link: https://lore.kernel.org/r/20210618045742.204195-5-rpearsonhpe@gmail.comSigned-off-by: NBob Pearson <rpearsonhpe@gmail.com>
Signed-off-by: NJason Gunthorpe <jgg@nvidia.com>

build_rdma_network_hdr() in rxe_resp.c does more copying than is
needed. Remove this subroutine and eliminate the extra copies for IPV6 and
reduce the extra copying for IPV4.

Fixes: e404f945 ("IB/rxe: improved debug prints & code cleanup")
Link: https://lore.kernel.org/r/20210618045742.204195-4-rpearsonhpe@gmail.comSigned-off-by: NBob Pearson <rpearsonhpe@gmail.com>
Signed-off-by: NJason Gunthorpe <jgg@nvidia.com>

In send_atomic_ack() in rxe_resp.c there is code copying ack_pkt into the
skb->cb[]. This doesn't do anything useful because the cb[] is not used in
the transmit path by the rxe driver.

Remove this code.

Fixes: 4c93496f ("IB/rxe: do not copy extra stack memory to skb")
Link: https://lore.kernel.org/r/20210618045742.204195-2-rpearsonhpe@gmail.comSigned-off-by: NBob Pearson <rpearson@hpe.com>
Signed-off-by: NJason Gunthorpe <jgg@nvidia.com>

Add code to implement memory access through memory windows.

Link: https://lore.kernel.org/r/20210608042552.33275-10-rpearsonhpe@gmail.comSigned-off-by: NBob Pearson <rpearsonhpe@gmail.com>
Signed-off-by: NJason Gunthorpe <jgg@nvidia.com>

Implement invalidate MW and cleaned up invalidate MR operations.

Added code to perform remote invalidate for send with invalidate.  Added
code to perform local invalidation. Deleted some blank lines in rxe_loc.h.

Link: https://lore.kernel.org/r/20210608042552.33275-9-rpearsonhpe@gmail.comSigned-off-by: NBob Pearson <rpearsonhpe@gmail.com>
Signed-off-by: NJason Gunthorpe <jgg@nvidia.com>

Currently the rdma_rxe driver attempts to protect atomic responder
resources by taking a reference to the qp which is only freed when the
resource is recycled for a new read or atomic operation. This means that
in normal circumstances there is almost always an extra qp reference once
an atomic operation has been executed which prevents cleaning up the qp
and associated pd and cqs when the qp is destroyed.

This patch removes the call to rxe_add_ref() in send_atomic_ack() and the
call to rxe_drop_ref() in free_rd_atomic_resource(). If the qp is
destroyed while a peer is retrying an atomic op it will cause the
operation to fail which is acceptable.

Link: https://lore.kernel.org/r/20210604230558.4812-1-rpearsonhpe@gmail.comReported-by: NZhu Yanjun <zyjzyj2000@gmail.com>
Fixes: 86af6176 ("IB/rxe: remove unnecessary skb_clone")
Signed-off-by: NBob Pearson <rpearsonhpe@gmail.com>
Signed-off-by: NJason 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: NBob Pearson <rpearsonhpe@gmail.com>
Signed-off-by: NJason Gunthorpe <jgg@nvidia.com>

All responder errors from request packets that do not consume a receive
WQE fail to generate acks for RC QPs.  This patch corrects this behavior
by making the flow follow the same path as request packets that do consume
a WQE after the completion.

Link: https://lore.kernel.org/r/20210402001016.3210-1-rpearson@hpe.com
Link: https://lore.kernel.org/linux-rdma/1a7286ac-bcea-40fb-2267-480134dd301b@gmail.com/Signed-off-by: NBob Pearson <rpearson@hpe.com>
Signed-off-by: NJason 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.comSigned-off-by: NBob Pearson <rpearson@hpe.com>
Acked-by: NZhu Yanjun <zyjzyj2000@gmail.com>
Signed-off-by: NJason Gunthorpe <jgg@nvidia.com>

The pkt->offset field is never used except to assign it to 0.  But it adds
lots of unneeded code. This patch removes the field and related code. This
causes a measurable improvement in performance.

Link: https://lore.kernel.org/r/20210211210455.3274-1-rpearson@hpe.comSigned-off-by: NBob Pearson <rpearson@hpe.com>
Signed-off-by: NJason 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: 4c173f59 ("RDMA/rxe: Use ib_device_get_by_netdev() instead of open coding")
Link: https://lore.kernel.org/r/20210128233318.2591-1-rpearson@hpe.comSigned-off-by: NBob Pearson <rpearson@hpe.com>
Signed-off-by: NJason Gunthorpe <jgg@nvidia.com>

This reverts commit b2d24404.

It's true that creating rxe on top of 802.1q interfaces doesn't work.
Thus, commit fd49ddaf ("RDMA/rxe: prevent rxe creation on top of vlan
interface") was absolutely correct.

But b2d24404 was incorrect assuming that with this change, RDMA and
VLAN don't work togehter at all. It just has to be set up
differently. Rather than creating rxe on top of the VLAN interface, rxe
must be created on top of the physical interface.  RDMA then works just
fine through VLAN interfaces on top of that physical interface, via the
"upper device" logic.

This is hard to see in the rxe logic because it never talks about vlan,
but instead rxe carefully selects upper vlan netdevices when working with
packets which in turn imply certain vlan tagging. This is all done
correctly and interacts with the gid table with VLAN support the same as
real HW does.

b2d24404 broke this setup deliberately and should thus be
reverted. Also, b2d24404 removed rxe_dma_device(), so adapt the revert
to discard that hunk.

Fixes: b2d24404 ("RDMA/rxe: Remove VLAN code leftovers from RXE")
Link: https://lore.kernel.org/r/20210120161913.7347-1-mwilck@suse.comSigned-off-by: NMartin Wilck <mwilck@suse.com>
Signed-off-by: NJason Gunthorpe <jgg@nvidia.com>

Since the commit fd49ddaf ("RDMA/rxe: prevent rxe creation on top of
vlan interface") does not permit rxe on top of vlan device, all the stuff
related with vlan should be removed.

Fixes: fd49ddaf ("RDMA/rxe: prevent rxe creation on top of vlan interface")
Link: https://lore.kernel.org/r/1604326422-18625-1-git-send-email-yanjunz@nvidia.comSigned-off-by: NZhu Yanjun <yanjunz@nvidia.com>
Reviewed-by: NLeon Romanovsky <leonro@nvidia.com>
Signed-off-by: NJason Gunthorpe <jgg@nvidia.com>

Add SPDX headers to all rxe .c and .h files.

Link: https://lore.kernel.org/r/20200827145439.2273-1-rpearson@hpe.comSigned-off-by: NBob Pearson <rpearson@hpe.com>
Signed-off-by: NJason Gunthorpe <jgg@nvidia.com>

If RoCE PDUs being sent or received contain pad bytes, then the iCRC
is miscalculated, resulting in PDUs being emitted by RXE with an incorrect
iCRC, as well as ingress PDUs being dropped due to erroneously detecting
a bad iCRC in the PDU.  The fix is to include the pad bytes, if any,
in iCRC computations.

Note: This bug has caused broken on-the-wire compatibility with actual
hardware RoCE devices since the soft-RoCE driver was first put into the
mainstream kernel.  Fixing it will create an incompatibility with the
original soft-RoCE devices, but is necessary to be compatible with real
hardware devices.

Fixes: 8700e3e7 ("Soft RoCE driver")
Signed-off-by: NSteve Wise <larrystevenwise@gmail.com>
Link: https://lore.kernel.org/r/20191203020319.15036-2-larrystevenwise@gmail.comSigned-off-by: NDoug Ledford <dledford@redhat.com>

Calculate the correct byte_len on the receiving side when a work
completion is generated with IB_WC_RECV_RDMA_WITH_IMM opcode.

According to the IBA byte_len must indicate the number of written bytes,
whereas it was always equal to zero for the IB_WC_RECV_RDMA_WITH_IMM
opcode, even though data was transferred.

Fixes: 8700e3e7 ("Soft RoCE driver")
Signed-off-by: NKonstantin Taranov <konstantin.taranov@inf.ethz.ch>
Signed-off-by: NJason Gunthorpe <jgg@mellanox.com>

The variable rxe in the function is not used. So it is removed.
Signed-off-by: NZhu Yanjun <yanjun.zhu@oracle.com>
Signed-off-by: NJason Gunthorpe <jgg@mellanox.com>

The variable rxe is only used in the function rxe_xmit_packet, and the
caller functions do not use it. So move this variable into the function
rxe_xmit_packet.
Signed-off-by: NZhu Yanjun <yanjun.zhu@oracle.com>
Signed-off-by: NDoug Ledford <dledford@redhat.com>
Signed-off-by: NJason 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: NAndrew Boyer <andrew.boyer@dell.com>
Signed-off-by: NDoug Ledford <dledford@redhat.com>
Signed-off-by: NJason Gunthorpe <jgg@mellanox.com>

Error completions must still contain a valid wr_id and
qp_num such that the consumer can rely on. Correctly
fill these fields in receive error completions.
Reported-by: NWalker Benjamin <benjamin.walker@intel.com>
Cc: stable@vger.kernel.org
Signed-off-by: NSagi Grimberg <sagi@grimberg.me>
Reviewed-by: NZhu Yanjun <yanjun.zhu@oracle.com>
Tested-by: NZhu Yanjun <yanjun.zhu@oracle.com>
Signed-off-by: NDoug Ledford <dledford@redhat.com>

When resp is in error state, the queued SKBs will not be handled.
The function get_req cleans up the skb queue directly.

CC: Srinivas Eeda <srinivas.eeda@oracle.com>
CC: Junxiao Bi <junxiao.bi@oracle.com>
Signed-off-by: NZhu Yanjun <yanjun.zhu@oracle.com>
Signed-off-by: NDoug Ledford <dledford@redhat.com>

Don't reset the resp opcode for a replayed read response.
The resp opcode could be in the middle of a write or send
sequence, when the duplicate read request was received.
An example sequence is as follows:
- Receive read request for 12KB PSN 20. Transmit read response
  first, middle and last with PSNs 20,21,22.
- Receive write first PSN 23.
  At this point the resp psn is 24 and resp opcode is write first.
- The sender notices that PSN 20 is dropped and retransmits.
  Receive read request for 12KB PSN 20. Transmit read response
  first, middle and last with PSNs 20,21,22. The resp opcode is
  set to -1, the resp psn remains 24.
- Receive write first PSN 23. This is processed by duplicate_request().
  The resp opcode remains -1 and resp psn remains 24.
- Receive write middle PSN 24. check_op_seq() reports a missing
  first error since the resp opcode is -1.

When sending an ack for a duplicate send or write request,
use the psn of the previous ack sent. Do not use the psn
of a read response for the ack.
An example sequence is as follows:
- Receive write PSN 30. Transmit ACK for PSN 30.
- Receive read request 4KB PSN 31. Transmit read response with
  PSN 31. The resp psn is now 32.
- The sender notices that PSN 30 is dropped and retransmits.
  Receive write PSN 30. duplicate_request() sends an ACK with
  PSN 31. That is incorrect since PSN 31 was a read request.
Signed-off-by: NVijay Immanuel <vijayi@attalasystems.com>
Signed-off-by: NDoug Ledford <dledford@redhat.com>

rxe_prepare() is called on an skb which has ndev already initialized by
rxe_init_packet().
Therefore avoid querying the GID attribute again and use the available
netdevice from the skb->dev.
Signed-off-by: NParav Pandit <parav@mellanox.com>
Reviewed-by: NDaniel Jurgens <danielj@mellanox.com>
Reviewed-by: NYuval Shaia <yuval.shaia@oracle.com>
Tested-by: NYuval Shaia <yuval.shaia@oracle.com>
Signed-off-by: NZhu Yanjun <yanjun.zhu@oracle.com>
Signed-off-by: NLeon Romanovsky <leonro@mellanox.com>
Signed-off-by: NDoug Ledford <dledford@redhat.com>

Set the vlan flag and vlan_id field in the wc for rdma_listen()
to work over VLAN. This is required by ib_init_ah_attr_from_wc()
which is called by the CM REQ handler.
Signed-off-by: NVijay Immanuel <vijayi@attalasystems.com>
Reviewed-by: NYonatan Cohen <yonatanc@mellanox.com>
Signed-off-by: NJason Gunthorpe <jgg@mellanox.com>

When skb is sent, it will pass the following functions in soft roce.

rxe_send [rdma_rxe]
    ip_local_out
        __ip_local_out
        ip_output
            ip_finish_output
                ip_finish_output2
                    dev_queue_xmit
                        __dev_queue_xmit
                            dev_hard_start_xmit

In the above functions, if error occurs in the above functions or
iptables rules drop skb after ip_local_out, kfree_skb will be called.
So it is not necessary to call kfree_skb in soft roce module again.
Or else crash will occur.

The steps to reproduce:

     server                       client
    ---------                    ---------
    |1.1.1.1|<----rxe-channel--->|1.1.1.2|
    ---------                    ---------

On server: rping -s -a 1.1.1.1 -v -C 10000 -S 512
On client: rping -c -a 1.1.1.1 -v -C 10000 -S 512

The kernel configs CONFIG_DEBUG_KMEMLEAK and
CONFIG_DEBUG_OBJECTS are enabled on both server and client.

When rping runs, run the following command in server:

iptables -I OUTPUT -p udp  --dport 4791 -j DROP

Without this patch, crash will occur.

CC: Srinivas Eeda <srinivas.eeda@oracle.com>
CC: Junxiao Bi <junxiao.bi@oracle.com>
Signed-off-by: NZhu Yanjun <yanjun.zhu@oracle.com>
Reviewed-by: NYuval Shaia <yuval.shaia@oracle.com>
Signed-off-by: NDoug Ledford <dledford@redhat.com>

In the functions rxe_mem_init_dma, rxe_mem_init_user, rxe_mem_init_fast
and copy_data, the function variable rxe is not used. So this function
variable rxe is removed.

CC: Srinivas Eeda <srinivas.eeda@oracle.com>
CC: Junxiao Bi <junxiao.bi@oracle.com>
Signed-off-by: NZhu Yanjun <yanjun.zhu@oracle.com>
Reviewed-by: NLeon Romanovsky <leonro@mellanox.com>
Signed-off-by: NDoug Ledford <dledford@redhat.com>

Follow the advice from Bart, the function refcount_inc is replaced
with skb_get in commit 99dae690 ("IB/rxe: optimize mcast recv process")
and commit 86af6176 ("IB/rxe: remove unnecessary skb_clone").

CC: Srinivas Eeda <srinivas.eeda@oracle.com>
CC: Junxiao Bi <junxiao.bi@oracle.com>
Suggested-by: NBart Van Assche <Bart.VanAssche@wdc.com>
Signed-off-by: NZhu Yanjun <yanjun.zhu@oracle.com>
Signed-off-by: NDoug Ledford <dledford@redhat.com>

In the function duplicate_request, the reference of skb can be increased
to replace the function skb_clone.

This will make rxe performace better and save memory.

CC: Srinivas Eeda <srinivas.eeda@oracle.com>
CC: Junxiao Bi <junxiao.bi@oracle.com>
Signed-off-by: NZhu Yanjun <yanjun.zhu@oracle.com>
Reviewed-by: NBart Van Assche <bart.vanassche@wdc.com>
Signed-off-by: NDoug Ledford <dledford@redhat.com>

In send_atomic_ack function, it is not necessary to make a
skb_clone. To gain better performance (high throughput and
low latency), this skb_clone is removed.

The following tests are made.

 server                       client
---------                    ---------
|1.1.1.1|<----rxe-channel--->|1.1.1.2|
---------                    ---------

On server: rping -s -a 1.1.1.1 -v -C 1000 -S 512
On client: rping -c -a 1.1.1.1 -v -C 1000 -S 512

The kernel config CONFIG_DEBUG_KMEMLEAK is enabled on both server
and client.

This test runs for several hours. There is no memory leak and the whole
system can work well.

Based on the above network, the following tests are made.

Server: ibv_rc_pingpong -d rxe0 -g 1
Client: ibv_rc_pingpong -d rxe0 -g 1 1.1.1.1

The test results on Server(10 tests are made).
Before:
Throughput is 137.07 Mbit/sec
Latency is 517.76 usec/iter

After:
Throughput is 148.85 Mbit/sec
Latency is 476.64 usec/iter

The throughput is enhanced and the latency is reduced.

CC: Srinivas Eeda <srinivas.eeda@oracle.com>
CC: Junxiao Bi <junxiao.bi@oracle.com>
Signed-off-by: NZhu Yanjun <yanjun.zhu@oracle.com>
Signed-off-by: NDoug Ledford <dledford@redhat.com>

The rxe driver works as follows:
* The send queue, receive queue and completion queues are implemented as
  circular buffers.
* ib_post_send() and ib_post_recv() calls are serialized through a spinlock.
* Removing elements from various queues happens from tasklet
  context. Tasklets are guaranteed to run on at most one CPU. This serializes
  access to these queues. See also rxe_completer(), rxe_requester() and
  rxe_responder().
* rxe_completer() processes the skbs queued onto qp->resp_pkts.
* rxe_requester() handles the send queue (qp->sq.queue).
* rxe_responder() processes the skbs queued onto qp->req_pkts.

Since rxe_drain_req_pkts() processes qp->req_pkts, calling
rxe_drain_req_pkts() from rxe_requester() is racy. Hence this patch.
Reported-by: NMoni Shoua <monis@mellanox.com>
Signed-off-by: NBart Van Assche <bart.vanassche@wdc.com>
Cc: stable@vger.kernel.org
Signed-off-by: NDoug Ledford <dledford@redhat.com>

This matches what the userspace copy of this header has been doing
for a while. imm_data is an opaque 4 byte array carried over the network,
and invalidate_rkey is in CPU byte order.
Signed-off-by: NJason Gunthorpe <jgg@mellanox.com>

The resource array is sized by max_dest_rd_atomic, not max_rd_atomic.
Iterating over max_rd_atomic entries of qp->resp.resources[] will cause
incorrect behavior when the two attributes are different (or even
crash if max_rd_atomic is larger).

Fixes: 8700e3e7 ("Soft RoCE driver")
Signed-off-by: NAndrew Boyer <andrew.boyer@dell.com>
Reviewed-by: NYuval Shaia <yuval.shaia@oracle.com>
Signed-off-by: NDoug Ledford <dledford@redhat.com>