There is opencoded send completion logic all over all
the drivers.

We need to convert to this routine to enforce ordering
issues for completions.  This routine fixes an ordering
issue where the read of the SWQE fields necessary for creating
the completion can race with a post send if the post send catches
a send queue at the edge of being full.  Is is possible in that situation
to read SWQE fields that are being written.

This new routine insures that SWQE fields are read prior to advancing
the index that post send uses to determine queue fullness.
Reviewed-by: NMichael J. Ruhl <michael.j.ruhl@intel.com>
Signed-off-by: NMike Marciniszyn <mike.marciniszyn@intel.com>
Signed-off-by: NDennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: NDoug Ledford <dledford@redhat.com>

The reference count adjustments on reference count completion
are open coded throughout.

Add a routine to do all reference count adjustments and use.
Reviewed-by: NMichael J. Ruhl <michael.j.ruhl@intel.com>
Signed-off-by: NMike Marciniszyn <mike.marciniszyn@intel.com>
Signed-off-by: NDennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: NJason Gunthorpe <jgg@mellanox.com>

The currently include file ordering for rdmavt headers has an
ab/ba include issue the precludes using inlines from rdma_vt.h
in rdmavt_qp.h.

At the heart of the issue is that rdma_vt.h includes rdmavt_qp.h.

Fix the ordering issue by adjusting rdma_vt.h to not require rdmavt_qp.h
and move qp related inlines to rdmavt_qp.h.

Additionally, promote rvt_mmap_info to rdma_vt.h since it is shared
by rdmavt_cq.h and rdmavt_qp.h.
Reviewed-by: NMichael J. Ruhl <michael.j.ruhl@intel.com>
Signed-off-by: NMike Marciniszyn <mike.marciniszyn@intel.com>
Signed-off-by: NDennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: NJason Gunthorpe <jgg@mellanox.com>

This patch fixes miscellaneous comment errors.
Reviewed-by: NMike Marciniszyn <mike.marciniszyn@intel.com>
Signed-off-by: NKaike Wan <kaike.wan@intel.com>
Signed-off-by: NDennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: NJason Gunthorpe <jgg@mellanox.com>

This patch integrates TID RDMA WRITE protocol into normal RDMA verbs
framework. The TID RDMA WRITE protocol is an end-to-end protocol
between the hfi1 drivers on two OPA nodes that converts a qualified
RDMA WRITE request into a TID RDMA WRITE request to avoid data copying
on the responder side.
Reviewed-by: NMike Marciniszyn <mike.marciniszyn@intel.com>
Signed-off-by: NMitko Haralanov <mitko.haralanov@intel.com>
Signed-off-by: NKaike Wan <kaike.wan@intel.com>
Signed-off-by: NDennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: NDoug Ledford <dledford@redhat.com>

The s_ack_queue is managed by two pointers into the ring:
r_head_ack_queue and s_tail_ack_queue. r_head_ack_queue is the index of
where the next received request is going to be placed and s_tail_ack_queue
is the entry of the request currently being processed. This works
perfectly fine for normal Verbs as the requests are processed one at a
time and the s_tail_ack_queue is not moved until the request that it
points to is fully completed.

In this fashion, s_tail_ack_queue constantly chases r_head_ack_queue and
the two pointers can easily be used to determine "queue full" and "queue
empty" conditions.

The detection of these two conditions are imported in determining when an
old entry can safely be overwritten with a new received request and the
resources associated with the old request be safely released.

When pipelined TID RDMA WRITE is introduced into this mix, things look
very different. r_head_ack_queue is still the point at which a newly
received request will be inserted, s_tail_ack_queue is still the
currently processed request. However, with pipelined TID RDMA WRITE
requests, s_tail_ack_queue moves to the next request once all TID RDMA
WRITE responses for that request have been sent. The rest of the protocol
for a particular request is managed by other pointers specific to TID RDMA
- r_tid_tail and r_tid_ack - which point to the entries for which the next
TID RDMA DATA packets are going to arrive and the request for which
the next TID RDMA ACK packets are to be generated, respectively.

What this means is that entries in the ring, which are "behind"
s_tail_ack_queue (entries which s_tail_ack_queue has gone past) are no
longer considered complete. This is where the problem is - a newly
received request could potentially overwrite a still active TID RDMA WRITE
request.

The reason why the TID RDMA pointers trail s_tail_ack_queue is that the
normal Verbs send engine uses s_tail_ack_queue as the pointer for the next
response. Since TID RDMA WRITE responses are processed by the normal Verbs
send engine, s_tail_ack_queue had to be moved to the next entry once all
TID RDMA WRITE response packets were sent to get the desired pipelining
between requests. Doing otherwise would mean that the normal Verbs send
engine would not be able to send the TID RDMA WRITE responses for the next
TID RDMA request until the current one is fully completed.

This patch introduces the s_acked_ack_queue index to point to the next
request to complete on the responder side. For requests other than TID
RDMA WRITE, s_acked_ack_queue should always be kept in sync with
s_tail_ack_queue. For TID RDMA WRITE request, it may fall behind
s_tail_ack_queue.
Reviewed-by: NMike Marciniszyn <mike.marciniszyn@intel.com>
Signed-off-by: NMitko Haralanov <mitko.haralanov@intel.com>
Signed-off-by: NKaike Wan <kaike.wan@intel.com>
Signed-off-by: NDennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: NDoug Ledford <dledford@redhat.com>

The RC retry timeout value is based on the estimated time for the
response packet to come back. However, for TID RDMA READ request, due
to the use of header suppression, the driver is normally not notified
for each incoming response packet until the last TID RDMA READ response
packet. Consequently, the retry timeout value should be extended to
cover the transaction time for the entire length of a segment (default
256K) instead of that for a single packet. This patch addresses the
issue by introducing new retry timer functions to account for multiple
packets and wrapper functions for backward compatibility.
Reviewed-by: NMike Marciniszyn <mike.marciniszyn@intel.com>
Signed-off-by: NKaike Wan <kaike.wan@intel.com>
Signed-off-by: NDennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: NDoug Ledford <dledford@redhat.com>

TID entries are used by hfi1 hardware to receive data payload from
incoming packets directly into a user buffer and thus avoid data copying
by software. This patch implements the functions for TID allocation,
freeing, and programming TID RcvArray entries in hardware for kernel
clients. TID entries are managed via lists of TID groups similar to PSM.
Furthermore, to track TID resource allocation for each request, software
flows are also allocated and freed as needed. Since software flows
consume large amount of memory for tracking TID allocation and freeing,
it is generally desirable to allocate them dynamically in the send queue
and only for TID RDMA requests, but pre-allocate them for receive queue
because the send queue could have thousands of entries while the receive
queue has only a limited number of entries.
Signed-off-by: NMitko Haralanov <mitko.haralanov@intel.com>
Signed-off-by: NAshutosh Dixit <ashutosh.dixit@intel.com>
Signed-off-by: NMike Marciniszyn <mike.marciniszyn@intel.com>
Signed-off-by: NKaike Wan <kaike.wan@intel.com>
Signed-off-by: NDennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: NDoug Ledford <dledford@redhat.com>

This patch moves some RC helper functions into a header file so that
they can be called from both RC and  TID RDMA functions. In addition,
a common function for rewinding a request is created in rdmavt so that
it can be shared between qib and hfi1 driver.
Reviewed-by: NMike Marciniszyn <mike.marciniszyn@intel.com>
Signed-off-by: NMitko Haralanov <mitko.haralanov@intel.com>
Signed-off-by: NKaike Wan <kaike.wan@intel.com>
Signed-off-by: NDennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: NDoug Ledford <dledford@redhat.com>

This patch moves ruc_loopback() from hfi1 into rdmavt for code sharing
with the qib driver.
Reviewed-by: NBrian Welty <brian.welty@intel.com>
Reviewed-by: NMike Marciniszyn <mike.marciniszyn@intel.com>
Reviewed-by: NDennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: NVenkata Sandeep Dhanalakota <venkata.s.dhanalakota@intel.com>
Signed-off-by: NHarish Chegondi <harish.chegondi@intel.com>
Signed-off-by: NDennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: NJason Gunthorpe <jgg@mellanox.com>

Moving send completion code into rdmavt in order to have shared logic
between qib and hfi1 drivers.
Reviewed-by: NMike Marciniszyn <mike.marciniszyn@intel.com>
Reviewed-by: NBrian Welty <brian.welty@intel.com>
Signed-off-by: NVenkata Sandeep Dhanalakota <venkata.s.dhanalakota@intel.com>
Signed-off-by: NHarish Chegondi <harish.chegondi@intel.com>
Signed-off-by: NDennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: NJason Gunthorpe <jgg@mellanox.com>

This patch moves hfi1_copy_sge() into rdmavt for sharing with qib.
This patch also moves all the wss_*() functions into rdmavt as
several wss_*() functions are called from hfi1_copy_sge()

When SGE copy mode is adaptive, cacheless copy may be done in some cases
for performance reasons. In those cases, X86 cacheless copy function
is called since the drivers that use rdmavt and may set SGE copy mode
to adaptive are X86 only. For this reason, this patch adds
"depends on X86_64" to rdmavt/Kconfig.
Reviewed-by: NAshutosh Dixit <ashutosh.dixit@intel.com>
Reviewed-by: NMichael J. Ruhl <michael.j.ruhl@intel.com>
Reviewed-by: NMike Marciniszyn <mike.marciniszyn@intel.com>
Reviewed-by: NDennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: NBrian Welty <brian.welty@intel.com>
Signed-off-by: NHarish Chegondi <harish.chegondi@intel.com>
Signed-off-by: NDennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: NJason Gunthorpe <jgg@mellanox.com>

Move some s_flags defines out of rdmavt and into hfi1 because they are
hfi1 specific and therefore should remain in the driver instead of
bubbling up to rdmavt.

Document device specific ranges in rdmavt and remap
those in hfi1.
Reviewed-by: NDennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: NMike Marciniszyn <mike.marciniszyn@intel.com>
Signed-off-by: NKaike Wan <kaike.wan@intel.com>
Signed-off-by: NDennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: NJason Gunthorpe <jgg@mellanox.com>

Moving receive-side WQE allocation logic into rdmavt will allow
further code reuse between qib and hfi1 drivers.
Reviewed-by: NMike Marciniszyn <mike.marciniszyn@intel.com>
Reviewed-by: NDennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: NBrian Welty <brian.welty@intel.com>
Signed-off-by: NHarish Chegondi <harish.chegondi@intel.com>
Signed-off-by: NDennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: NDoug Ledford <dledford@redhat.com>

The s_ahgpsn was incorrectly placed in the read-mostly section of the QP
and the s_curr_size and s_hdrwords are oversized. The misplaced
s_ahgpsn will cause the read-mostly cachelines to thrash.

Place s_ahgpsn in the send side cache lines and correctly size and
s_hdrwords and s_cur_size to keep the send side cachelines at the same
size.
Reviewed-by: NMike Marciniszyn <mike.marciniszyn@intel.com>
Signed-off-by: NSebastian Sanchez <sebastian.sanchez@intel.com>
Signed-off-by: NDennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: NDoug Ledford <dledford@redhat.com>

A destroy of an MR prior to destroying the QP can cause the following
diagnostic if the QP is referencing the MR being de-registered:

hfi1 0000:05:00.0: hfi1_0: rvt_dereg_mr timeout mr ffff8808562108
              00 pd ffff880859b20b00

The solution is to when the a non-zero refcount is encountered when
the MR is destroyed the QPs needs to be iterated looking for QPs in
the same PD as the MR.  If rvt_qp_mr_clean() detects any such QP
references the rkey/lkey, the QP needs to be put into an error state
via a call to rvt_qp_error() which will trigger the clean up of any
stuck references.

This solution is as specified in IBTA 1.3 Volume 1 11.2.10.5.

[This is reproduced with the 0.4.9 version of qperf and the rc_bw test]
Reviewed-by: NDennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: NMike Marciniszyn <mike.marciniszyn@intel.com>
Signed-off-by: NDennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: NDoug Ledford <dledford@redhat.com>

There are currently 3 spots in the qib and hfi1 driver that have
knowledge of the internal QP hash list that should only be in
scope to rdmavt QP code.

Add an iterator API for processing all QPs to hide the
nature of the RCU hashlist.

The API consists of:
- rvt_qp_iter_init()
  * For iterating QPs one at a time for seq_file semantics
- rvt_qp_iter_next()
  * For iterating QPs one at a time for seq_file semantics
- rvt_qp_iter()
  * For iterating all QPs

The first two are used for things like seq_file prints.

The last is for code that just needs to iterate all QPs
in the system.
Reviewed-by: NDennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: NMike Marciniszyn <mike.marciniszyn@intel.com>
Signed-off-by: NDennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: NDoug Ledford <dledford@redhat.com>

The pmtu field doens't have be stored in the QP structure
as it can easily be calculated when needed.
Reviewed-by: NMike Marciniszyn <mike.marciniszyn@intel.com>
Signed-off-by: NSebastian Sanchez <sebastian.sanchez@intel.com>
Signed-off-by: NDennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: NDoug Ledford <dledford@redhat.com>

Current computation of qp->timeout_jiffies in rvt_modify_qp() will cause
overflow due to the fact that the input to the function usecs_to_jiffies
is only 32-bit ( unsigned int). Overflow will occur when attr->timeout is
equal to or greater than 30. The consequence is unnecessarily excessive
retry and thus degradation of the system performance.

This patch fixes the problem by limiting the input to 5-bit and calling
usecs_to_jiffies() before multiplying the scaling factor.
Reviewed-by: NMike Marciniszyn <mike.marciniszyn@intel.com>
Signed-off-by: NKaike Wan <kaike.wan@intel.com>
Signed-off-by: NDennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: NDoug Ledford <dledford@redhat.com>

Improve code readablity by adding inline functions
to read specific BTH/IB fields without knowledge of
byte offsets.
Reviewed-by: NBrian Welty <brian.welty@intel.com>
Reviewed-by: NDasaratharaman Chandramouli <dasaratharaman.chandramouli@intel.com>
Reviewed-by: NDennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: NDon Hiatt <don.hiatt@intel.com>
Signed-off-by: NDennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: NDoug Ledford <dledford@redhat.com>

This field is causing excessive cache line bouncing.

There are spare bytes in the r_lock cache line so the best approach
is to make an rvt QP field and remove from the hfi1 priv field.
Signed-off-by: NSebastian Sanchez <sebastian.sanchez@intel.com>
Signed-off-by: NMike Marciniszyn <mike.marciniszyn@intel.com>
Signed-off-by: NDennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: NDoug Ledford <dledford@redhat.com>

This patch simply renames struct ib_ah_attr to
rdma_ah_attr as these fields specify attributes that are
not necessarily specific to IB.
Reviewed-by: NIra Weiny <ira.weiny@intel.com>
Reviewed-by: NDon Hiatt <don.hiatt@intel.com>
Reviewed-by: NNiranjana Vishwanathapura <niranjana.vishwanathapura@intel.com>
Reviewed-by: NSean Hefty <sean.hefty@intel.com>
Signed-off-by: NDasaratharaman Chandramouli <dasaratharaman.chandramouli@intel.com>
Signed-off-by: NDoug Ledford <dledford@redhat.com>

The Infiniband spec defines "A multicast address is defined by a
MGID and a MLID" (section 10.5).

The current code only uses the MGID for identifying multicast groups.
Update the driver to be compliant with this definition.
Reviewed-by: NIra Weiny <ira.weiny@intel.com>
Reviewed-by: NDasaratharaman Chandramouli <dasaratharaman.chandramouli@intel.com>
Signed-off-by: NMichael J. Ruhl <michael.j.ruhl@intel.com>
Signed-off-by: NDennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: NDoug Ledford <dledford@redhat.com>

The wqe should be read only and in fact the superfluous reset of the
RVT_SEND_RESERVE_USED flag causes an issue where reserved operations
elicit a bad completion to the ULP.

The maintenance of the flag is now entirely within rvt_post_one_wr()
where a reserved operation will set the flag and a non-reserved operation
will insure the operation that is about to be posted has the flag reset.

Fixes: Commit 856cc4c2 ("IB/hfi1: Add the capability for reserved operations")
Reviewed-by: NDon Hiatt <don.hiatt@intel.com>
Signed-off-by: NMike Marciniszyn <mike.marciniszyn@intel.com>
Signed-off-by: NDennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: NDoug Ledford <dledford@redhat.com>

The work to create a completion helper moved the translation of send
wqe operations to completion opcodes to rdmvat.

This precludes having driver dependent operations.  Make the translation
driver dependent by doing the translation in the driver prior to the
rvt_qp_swqe_complete() call using restored translation tables.

Fixes: Commit f2dc9cdc ("IB/rdmavt: Add a send completion helper")
Fixes: Commit 0771da5a ("IB/hfi1,IB/qib: Use new send completion helper")
Reviewed-by: NDennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: NMike Marciniszyn <mike.marciniszyn@intel.com>
Signed-off-by: NDennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: NDoug Ledford <dledford@redhat.com>

Rename RVT AETH defines and export in rdma/ib_hdrs.h
Reviewed-by: NMike Marciniszyn <mike.marciniszyn@intel.com>
Signed-off-by: NDon Hiatt <don.hiatt@intel.com>
Signed-off-by: NDennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: NDoug Ledford <dledford@redhat.com>

Return usec from an index into ib_rvt_rnr_table.
Reviewed-by: NMike Marciniszyn <mike.marciniszyn@intel.com>
Signed-off-by: NDon Hiatt <don.hiatt@intel.com>
Signed-off-by: NDennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: NDoug Ledford <dledford@redhat.com>

The send complete for RC QPs mismanages the ack count when the
responder side is only in RTR.

A QP in that state cannot send requests, but it can be the target
for operations that elicit responses.

Adjust the RC completion logic to correct the count maintenance
by reflecting RECV_OK in a new state test.
Reviewed-by: NKaike Wan <kaike.wan@intel.com>
Signed-off-by: NMike Marciniszyn <mike.marciniszyn@intel.com>
Signed-off-by: NDennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: NDoug Ledford <dledford@redhat.com>

Convert copy_sge and related SGE state functions to use boolean.
For determining if QP is in user mode, add helper function in rdmavt_qp.h.
This is used to determine if QP needs the last byte ordering.
While here, change rvt_pd.user to a boolean.
Reviewed-by: NMike Marciniszyn <mike.marciniszyn@intel.com>
Reviewed-by: NDean Luick <dean.luick@intel.com>
Signed-off-by: NBrian Welty <brian.welty@intel.com>
Signed-off-by: NDennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: NDoug Ledford <dledford@redhat.com>

To move common code across target to rdmavt for code reuse.
Reviewed-by: NMike Marciniszyn <mike.marciniszyn@intel.com>
Reviewed-by: NBrian Welty <brian.welty@intel.com>
Signed-off-by: NVenkata Sandeep Dhanalakota <venkata.s.dhanalakota@intel.com>
Signed-off-by: NDennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: NDoug Ledford <dledford@redhat.com>

Add rvt_compute_aeth() and rvt_get_credit() as shared functions in
rdmavt, moved from hfi1/qib logic.
Reviewed-by: NDennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: NBrian Welty <brian.welty@intel.com>
Signed-off-by: NDennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: NDoug Ledford <dledford@redhat.com>

Add rvt_rc_error() and rvt_comm_est() as shared functions in
rdmavt, moved from hfi1/qib logic.
Reviewed-by: NDennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: NBrian Welty <brian.welty@intel.com>
Signed-off-by: NDennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: NDoug Ledford <dledford@redhat.com>

Add rvt_div_round_up_mtu() and rvt_div_mtu() routines to
do the computation based on the pmtu and the log_pmtu.

Change divides in qib, hfi1 to use the new inlines.
Reviewed-by: NKaike Wan <kaike.wan@intel.com>
Signed-off-by: NMike Marciniszyn <mike.marciniszyn@intel.com>
Signed-off-by: NDennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: NDoug Ledford <dledford@redhat.com>

Add a helper to release mr references held by
an swqe.
Reviewed-by: NBrian Welty <brian.welty@intel.com>
Signed-off-by: NMike Marciniszyn <mike.marciniszyn@intel.com>
Signed-off-by: NDennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: NDoug Ledford <dledford@redhat.com>

This is for use by client drivers to drive
send completions into a CQ.

A new exported table allows for the mapping
of ib_wr_opcode into a ib_wc_opcode.
Reviewed-by: NAshutosh Dixit <ashutosh.dixit@intel.com>
Signed-off-by: NMike Marciniszyn <mike.marciniszyn@intel.com>
Signed-off-by: NDennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: NDoug Ledford <dledford@redhat.com>

This centralizes the function and improves code readability.
Reviewed-by: NDennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: NMike Marciniszyn <mike.marciniszyn@intel.com>
Signed-off-by: NDennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: NDoug Ledford <dledford@redhat.com>

The use of the specific opcode test is redundant since
all ack entry users correctly manipulate the mr pointer
to selectively trigger the reference clearing.

The overly specific test hinders the use of implementation
specific operations.

The change needs to get rid of the union to insure that
an atomic value is not seen as an MR pointer.
Reviewed-by: NAshutosh Dixit <ashutosh.dixit@intel.com>
Signed-off-by: NMike Marciniszyn <mike.marciniszyn@intel.com>
Signed-off-by: NIra Weiny <ira.weiny@intel.com>
Signed-off-by: NDoug Ledford <dledford@redhat.com>

Hanging has been observed while writing a file over NFSoRDMA. Dmesg on
the server contains messages like these:

[  931.992501] svcrdma: Error -22 posting RDMA_READ
[  952.076879] svcrdma: Error -22 posting RDMA_READ
[  982.154127] svcrdma: Error -22 posting RDMA_READ
[ 1012.235884] svcrdma: Error -22 posting RDMA_READ
[ 1042.319194] svcrdma: Error -22 posting RDMA_READ

Here is why:

With the base memory management extension enabled, FRMR is used instead
of FMR. The xprtrdma server issues each RDMA read request as the following
bundle:

(1)IB_WR_REG_MR, signaled;
(2)IB_WR_RDMA_READ, signaled;
(3)IB_WR_LOCAL_INV, signaled & fencing.

These requests are signaled. In order to generate completion, the fast
register work request is processed by the hfi1 send engine after being
posted to the work queue, and the corresponding lkey is not valid until
the request is processed. However, the rdmavt driver validates lkey when
the RDMA read request is posted and thus it fails immediately with error
-EINVAL (-22).

This patch changes the work flow of local operations (fast register and
local invalidate) so that fast register work requests are always
processed immediately to ensure that the corresponding lkey is valid
when subsequent work requests are posted. Local invalidate requests are
processed immediately if fencing is not required and no previous local
invalidate request is pending.

To allow completion generation for signaled local operations that have
been processed before posting to the work queue, an internal send flag
RVT_SEND_COMPLETION_ONLY is added. The hfi1 send engine checks this flag
and only generates completion for such requests.
Reviewed-by: NMike Marciniszyn <mike.marciniszyn@intel.com>
Signed-off-by: NJianxin Xiong <jianxin.xiong@intel.com>
Signed-off-by: NDennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: NDoug Ledford <dledford@redhat.com>

This fix allows for support of in-kernel reserved operations
without impacting the ULP user.

The low level driver can register a non-zero value which
will be transparently added to the send queue size and hidden
from the ULP in every respect.

ULP post sends will never see a full queue due to a reserved
post send and reserved operations will never exceed that
registered value.

The s_avail will continue to track the ULP swqe availability
and the difference between the reserved value and the reserved
in use will track reserved availabity.
Reviewed-by: NAshutosh Dixit <ashutosh.dixit@intel.com>
Signed-off-by: NMike Marciniszyn <mike.marciniszyn@intel.com>
Signed-off-by: NDennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: NDoug Ledford <dledford@redhat.com>

Some work requests are local operations, such as IB_WR_REG_MR and
IB_WR_LOCAL_INV. They differ from non-local operations in that:

(1) Local operations can be processed immediately without being posted
to the send queue if neither fencing nor completion generation is needed.
However, to ensure correct ordering, once a local operation is posted to
the work queue due to fencing or completion requiement, all subsequent
local operations must also be posted to the work queue until all the
local operations on the work queue have completed.

(2) Local operations don't send packets over the wire and thus don't
need (and shouldn't update) the packet sequence numbers.

Define a new a flag bit for the post send table to identify local
operations.

Add a new field to the QP structure to track the number of local
operations on the send queue to determine if direct processing of new
local operations should be enabled/disabled.
Reviewed-by: NMike Marciniszyn <mike.marciniszyn@intel.com>
Reviewed-by: NDennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: NJianxin Xiong <jianxin.xiong@intel.com>
Signed-off-by: NDennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: NDoug Ledford <dledford@redhat.com>