staging/rdma/hfi1: Fix bug that could block the process on context exit

A race was discovred in the user SDMA code, which could result
in an process being stuck in the kernel call indefinitely in
certain error conditions.

If, during the processing of a user SDMA request, there was an
error *and* all outstanding SDMA descriptor had been completed
by the time the that error case was handled in the calling function,
the state of the packet queue would not get correctly updated
resulting in the process subsequently getting stuck, thinking that
there are more descriptors to be completed.

To handle this scenario, the driver now checks the submitted
packet count vs. the completed. If all submitted packets have also
been completed, the driver can safely free the request and signal
user level. Otherwise, this will be handled by the completion
callback.

Reviewed-by: Dennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: Mitko Haralanov <mitko.haralanov@intel.com>
Signed-off-by: Jubin John <jubin.john@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>

staging/rdma/hfi1: Remove unused variable nsbr

Remove unused nsbr count from PCIe Gen3 code

Reviewed-by: Stuart Summers <john.s.summers@intel.com>
Signed-off-by: Dean Luick <dean.luick@intel.com>
Signed-off-by: Jubin John <jubin.john@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>

staging/rdma/hfi1: Make EPROM check per device

Add a variable eprom_available to each device, replacing the
global of the same name.  This is to allow multiple HFI devices
with different EPROM availability to operate correctly on the
the same system.

Reviewed-by: Dennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: Dean Luick <dean.luick@intel.com>
Signed-off-by: Jubin John <jubin.john@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>

staging/rdma/hfi1: Add credits for VL0 to VL7 in snoop mode

Add a new option to the snoop ioctl which allows credits to be allocated
across all VLs. Previously only VL0 and VL15 had credits allocated.
The new option used in the ioctl HFI1_SNOOP_IOCSET_OPTS allows credits
to be allocated so that VL15 will have at least 8.5KB credits and the
other VLs will have the rest of the credits divided equally across
themselves.

The total number of credits are stored in the upper 16 bits of the
integer passed and the cumulative value should ensure that VL0 has at
least 8.5KB and each VL a minimum of 2KB + 128 bytes

Reviewed-by: Dennis Dalessandro <dennis.dalessandro@intel.com>
Reviewed-by: Dean Luick <dean.luick@intel.com>
Signed-off-by: Sadanand Warrier <sadanand.warrier@intel.com>
Signed-off-by: Jubin John <jubin.john@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>

staging/rdma/hfi1: Improve performance of user SDMA

To facilitate locked page counting, the user SDMA
routines would maintain a list of io vectors, which
were freed in the completion callback and then unpin
the associated pages during the next call into the
kernel.

Since the size of this list was unbounded, doing this
was bad for performance because the driver ended up
spending too much time freeing the io vectors.

This commit changes how the io vector freeing is done
by moving the actual page unpinning in the callback and
maintaining a count of unpinned pages. This count can
then be used during the next call into the kernel to
update the mm->pinned_vm variable (since that requires
process context and the ability to sleep.)

Reviewed-by: Mike Marciniszyn <mike.marciniszyn@intel.com>
Signed-off-by: Mitko Haralanov <mitko.haralanov@intel.com>
Signed-off-by: Jubin John <jubin.john@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>

staging/rdma/hfi1, IB/core: Fix LinkDownReason define for consistency

LinkDownReason LocalMediaNotInstalled lacked an underscore
and was inconsistent with other defines in the same family.
This patch fixes this.

Reviewed-by: Ira Weiny <ira.weiny@intel.com>
Signed-off-by: Easwar Hariharan <easwar.hariharan@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>

staging/rdma/hfi1: Remove modify_port and port_immutable functions

Delete code from query_port which has been moved into rvt_query_port
Create a call back function to shut down a port which may be called from
rvt_modify_port

Signed-off-by: Harish Chegondi <harish.chegondi@intel.com>
Signed-off-by: Dennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>

staging/rdma/hfi1: Support query gid in rdmavt

Query gid is in rdmavt, but still relies on the driver to maintain the
guid table. Add the necessary driver call back and remove the existing
verb handler.

Reviewed-by: Harish Chegondi <harish.chegondi@intel.com>
Signed-off-by: Dennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>

staging/rdma/hfi1: Clean up init_cntrs()

Clean up init_cntrs() by removing unnecessary memsets and debug
statements

Suggested-by: Dan Carpenter <dan.carpenter@oracle.com>
Reviewed-by: Dennis Dalessandro <dennis.dalessandro@intel.com>
Reviewed-by: Mike Marciniszyn <mike.marciniszyn@intel.com>
Signed-off-by: Jubin John <jubin.john@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>

staging/rdma/hfi1: Fix snoop packet length calculation

The LRH has a 12 bit packet length field, not 11 bit. This caused a
snoop packet length miscalculation leading to a crash when sending a
large ping over IPoIB while running opapacketcapture.

Reviewed-by: Mike Marciniszyn <mike.marciniszyn@intel.com>
Signed-off-by: Dean Luick <dean.luick@intel.com>
Signed-off-by: Jubin John <jubin.john@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>

staging/rdma/hfi1: Correct TWSI reset

Change the TWSI reset function so it will stop the reset
once the lines are in an expected state.

Reviewed-by: Easwar Hariharan <easwar.hariharan@intel.com>
Reviewed-by: Dean Luick <dean.luick@intel.com>
Signed-off-by: Pablo Cacho <pablo.cacho@intel.com>
Signed-off-by: Jubin John <jubin.john@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>

staging/rdma/hfi1: Remove PCIe AER diagnostic message

There are several reasons why PCIE AER cannot be enabled.  Do not
report the failure to enable as an error.

Reviewed-by: Dennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: Dean Luick <dean.luick@intel.com>
Signed-off-by: Jubin John <jubin.john@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>

staging/rdma/hfi1: Implement LED beaconing for maintenance

This patch implements LED beaconing for maintenance. A MAD packet with
the LEDInfo attribute set to 1 will enable LED beaconing with a duty
cycle of 2s on and 1.5s off. A MAD packet with the LEDInfo attribute
set to 0 will disable beaconing and return the LED to normal operation.

Reviewed-by: Dennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: Easwar Hariharan <easwar.hariharan@intel.com>
Signed-off-by: Mitko Haralanov <mitko.haralanov@intel.com>
Signed-off-by: Jubin John <jubin.john@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>

staging/rdma/hfi1: Split last 8 bytes of copy to user buffer

Copy the last 8 bytes of user mode RC WRITE_ONLY and WRITE_LAST
opcodes separately from the rest of the data.

It is a de-facto standard for some MPI implementations to use a
poll on the last few bytes of a verbs message to indicate that
the message has been received rather than follow the required
function method.  The driver uses the kernel memcpy routine, which
becomes "rep movsb" on modern machines.  This copy, while very
fast, does not guarantee in-order copy completion and the result
is an occasional perceived corrupted packet.  Avoid the issue by
splitting the last 8 bytes to copy from the verbs opcodes where it
matters and performing an in-order byte copy.

Reviewed-by: Mike Marciniszyn <mike.marciniszyn@intel.com>
Signed-off-by: Dean Luick <dean.luick@intel.com>
Signed-off-by: Jubin John <jubin.john@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>

staging/rdma/hfi1: Fix fabric serdes reset by re-downloading firmware

A host fabric serdes reset is required to go back to polling.
However, access to the fabric serdes may have been invalidated
by the sibling HFI when it downloads its fabric serdes firmware.
Work around this by re-downloading and re-validating the serdes
firmware at reset time on Bx hardware.

Reviewed-by: Dennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: Dean Luick <dean.luick@intel.com>
Signed-off-by: Jubin John <jubin.john@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>

staging/rdma/hfi1: Report physical state changes per device instead of globally

Make physical state change reporting be per-device, not global
to reduce excessive reports of "physical state changed"

Reviewed-by: Dennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: Dean Luick <dean.luick@intel.com>
Signed-off-by: Jubin John <jubin.john@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>

staging/rdma/hfi1: Properly determine error status of SDMA slots

To ensure correct operation between the driver and PSM
with respect to managing the SDMA request ring, it is
important that the status for a particular request slot
is set at the correct time. Otherwise, PSM can get out
of sync with the driver, which could lead to hangs or
errors on new requests.

Properly determining of when to set the error status of
a SDMA slot depends on knowing exactly when the last txreq
for that request has been completed. This in turn requires
that the driver knows exactly how many requests have been
generated and how many of those requests have been successfully
submitted to the SDMA queue.

The previous implementation of the mid-layer SDMA API did not
provide a way for the caller of sdma_send_txlist() to know how
many of the txreqs in the input list have actually been submitted
without traversing the list and counting. Since sdma_send_txlist()
already traverses the list in order to process it, requiring
such traversal in the caller is completely unnecessary. Therefore,
it is much easier to enhance sdma_send_txlist() to return the
number of successfully submitted txreqs.

This, in turn, allows the caller to accurately determine the
progress of the SDMA request and, therefore, correctly set the
error status at the right time.

Reviewed-by: Mike Marciniszyn <mike.marciniszyn@intel.com>
Signed-off-by: Mitko Haralanov <mitko.haralanov@intel.com>
Signed-off-by: Jubin John <jubin.john@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>

staging/rdma/hfi1: correctly check for post-interrupt packets

At the end of the packet processing interrupt and thread handler,
the RcvAvail interrupt is finally cleared down.  There is a window
between the last packet check (via DMA to memory) and interrupt
clear-down.  The code to recheck for a packet once the RcvAVail
interrupt is enabled must ultimately use a CSR read of RcvHdrTail
rather than depend on DMA'ed memory.

This change adds a CSR read of RcvHdrTail if the memory check does
not show a packet preset.  The memory check is retained as a quick
test before doing the more expensive, but always correct, CSR read.

In the ASIC, the CSR read used to force the RcvAvail clear-down write
to complete may bypass queued DMA writes to memory.  The only correct
way to decide if a packet has arrived without an interrupt to push DMA
to memory ahead of itself is to read the tail directly after RcvAvail
has been cleared down.  It is not sufficient to just read the tail and
skip pushing the clear-down.  Both must be done.  The tail read will not
push clear-down write due to it being in a different area of the chip.

At this point, it is OK to have packet data still being DMA'ed to
memory.  This is the end of packet processing for previous packets.
If the driver detects a new packet has arrived before interrputs were
re-enabled, it will force a new interrupt and the interrupt will push
the packet DMAs to memory, where the driver will then react to the
interrupt and do normal packet processing.

Reviewed-by: Mike Marciniszyn <mike.marciniszyn@intel.com>
Signed-off-by: Dean Luick <dean.luick@intel.com>
Signed-off-by: Jubin John <jubin.john@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>

staging/rdma/hfi1: Improve performance of SDMA transfers

Commit a0d406934a46 ("staging/rdma/hfi1: Add page lock limit
check for SDMA requests") added a mechanism to
delay the clean-up of user SDMA requests in order to facilitate
proper locked page counting.

This delayed processing was done using a kernel workqueue, which
meant that a kernel thread would have to spin up and take CPU
cycles to do the clean-up.

This proved detrimental to performance because now there are two
execution threads (the kernel workqueue and the user process)
needing cycles on the same CPU.

Performance-wise, it is much better to do as much of the clean-up
as can be done in interrupt context (during the callback) and do
the remaining work in-line during subsequent calls of the user
process into the driver.

The changes required to implement the above also significantly
simplify the entire SDMA completion processing code and eliminate
a memory corruption causing the following observed crash:

    [ 2881.703362] BUG: unable to handle kernel NULL pointer dereference at        (null)
    [ 2881.703389] IP: [<ffffffffa02897e4>] user_sdma_send_pkts+0xcd4/0x18e0 [hfi1]
    [ 2881.703422] PGD 7d4d25067 PUD 77d96d067 PMD 0
    [ 2881.703427] Oops: 0000 [#1] SMP
    [ 2881.703431] Modules linked in:
    [ 2881.703504] CPU: 28 PID: 6668 Comm: mpi_stress Tainted: G           OENX 3.12.28-4-default #1
    [ 2881.703508] Hardware name: Intel Corporation S2600KP/S2600KP, BIOS SE5C610.86B.11.01.0044.090
    [ 2881.703512] task: ffff88077da8e0c0 ti: ffff880856772000 task.ti: ffff880856772000
    [ 2881.703515] RIP: 0010:[<ffffffffa02897e4>]  [<ffffffffa02897e4>] user_sdma_send_pkts+0xcd4/0x
    [ 2881.703529] RSP: 0018:ffff880856773c48  EFLAGS: 00010287
    [ 2881.703531] RAX: 0000000000000000 RBX: 0000000000001000 RCX: 0000000000002000
    [ 2881.703534] RDX: 0000000000000000 RSI: 0000000000000000 RDI: 0000000000002000
    [ 2881.703537] RBP: 0000000000000000 R08: 0000000000000001 R09: 0000000000000000
    [ 2881.703540] R10: 0000000000000000 R11: 0000000000000000 R12: 0000000000000000
    [ 2881.703543] R13: 0000000000000000 R14: ffff88071e782e68 R15: ffff8810532955c0
    [ 2881.703546] FS:  00007f9c4375e700(0000) GS:ffff88107eec0000(0000) knlGS:0000000000000000
    [ 2881.703549] CS:  0010 DS: 0000 ES: 0000 CR0: 0000000080050033
    [ 2881.703551] CR2: 0000000000000000 CR3: 00000007d4cba000 CR4: 00000000003407e0
    [ 2881.703554] DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000
    [ 2881.703556] DR3: 0000000000000000 DR6: 00000000fffe0ff0 DR7: 0000000000000400
    [ 2881.703558] Stack:
    [ 2881.703559]  ffffffff00002000 ffff881000001800 ffffffff00000000 00000000000080d0
    [ 2881.703570]  0000000000000000 0000200000000000 0000000000000000 ffff88071e782db8
    [ 2881.703580]  ffff8807d4d08d80 ffff881053295600 0000000000000008 ffff88071e782fc8
    [ 2881.703589] Call Trace:
    [ 2881.703691]  [<ffffffffa028b5da>] hfi1_user_sdma_process_request+0x84a/0xab0 [hfi1]
    [ 2881.703777]  [<ffffffffa0255412>] hfi1_aio_write+0xd2/0x110 [hfi1]
    [ 2881.703828]  [<ffffffff8119e3d8>] do_sync_readv_writev+0x48/0x80
    [ 2881.703837]  [<ffffffff8119f78b>] do_readv_writev+0xbb/0x230
    [ 2881.703843]  [<ffffffff8119fab8>] SyS_writev+0x48/0xc0

This commit also addresses issues related to notification of user
processes of SDMA request slot availability. The slot should be
cleaned up first before the user processes is notified of its
availability.

Reviewed-by: Arthur Kepner <arthur.kepner@intel.com>
Reviewed-by: Dennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: Mitko Haralanov <mitko.haralanov@intel.com>
Signed-off-by: Jubin John <jubin.john@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>

staging/rdma/hfi1: Use device file minor to identify EPROM

When writing to the EPROM, the driver will always use the
"first" device.  This is incorrect for multiple cards.

Use the device file minor to determine the device to use.
Reject the generic device file.

Reviewed-by: Mitko Haralanov <mitko.haralanov@intel.com>
Signed-off-by: Dean Luick <dean.luick@intel.com>
Signed-off-by: Jubin John <jubin.john@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>

staging/rdma/hfi1: Reduce syslog message severity and provide speed information

The syslog message causes unnecessary alarm for the single and dual port
x8 cards by reporting at an error level. This patch reduces the severity
to informational only and adds speed information.

Reviewed-by: Dennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: Easwar Hariharan <easwar.hariharan@intel.com>
Signed-off-by: Jubin John <jubin.john@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>

staging/rdma/hfi1: Improve performance of TID cache look up

When TID caching was enabled, the way the driver found
RB nodes when PSM was unprogramming TID entries was by
traversing the RB tree, looking for a match on the
RcvArray entry index.

The performance of this algorithm was not only poor but
also inconsistent depending on how many RB nodes would
have to be traversed before a match was found.

The lower performance was especially evident in cases where
there was a cache miss with the cache full, requiring the
unprogramming of several TID entries.

This commit changes how RB nodes are looked up when being
free'd by PSM to a index-based lookup into a flat array on
the index of the RcvArray entry. This turns the entire
look-up process into an O(1) algorithm.

Special care needs to be taken for situations when TID
caching is disabled. In those cases, there is no need to
insert the RB nodes into an actual RB tree. Since the entire
RcvArray management mechanism is managed by an index-based
algorithm, the RB nodes can be saved into the flat array,
making both "insertion" and "removal" faster.

Reviewed-by: Arthur Kepner <arthur.kepner@intel.com>
Reviewed-by: Dennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: Mitko Haralanov <mitko.haralanov@intel.com>
Signed-off-by: Jubin John <jubin.john@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>

staging/rdma/hfi1: Fix for module parameter rcvhdrcnt when it's 2097152

The driver crashes when loaded with parameter rcvhdrcnt=2097152.
The root cause was that rcvhdrcnt was initially a 32 bit variable
and its value was assigned to a 16 bit variable, truncating the
upper 16 bits. This patch prevents the user from passing a value
for rcvhdrcnt greater than 16352 (Maximum number for rcvhdrcnt).

Reviewed-by: Dean Luick <dean.luick@intel.com>
Reviewed-by: Mitko Haralanov <mitko.haralanov@intel.com>
Signed-off-by: Sebastian Sanchez <sebastian.sanchez@intel.com>
Signed-off-by: Jubin John <jubin.john@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>

staging/rdma/hfi1: Allow a fair scheduling of QPs

This patch fixes the fairness issues in QP scheduling
- the timeout for cond_resched is changed to a ratio of
qp->timeout_jiffies
- workqueue_congested is used to determine if qp needs to
reschedule itself

Reviewed-by: Mike Marciniszyn <mike.marciniszyn@intel.com>
Signed-off-by: Vennila Megavannan <vennila.megavannan@intel.com>
Signed-off-by: Jubin John <jubin.john@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>

staging/rdma/hfi1: Fix for generic I2C interface

The original I2C interface was geared for QSFP accesses. Modify
the interface to behave more like a generic I2C controller such
that reads and writes can accept multi-byte offsets. Removed
reads following writes and moved reset to top level.

Reviewed-by: Easwar Hariharan <easwar.hariharan@intel.com>
Reviewed-by: Dean Luick <dean.luick@intel.com>
Signed-off-by: Pablo Cacho <pablo.cacho@intel.com>
Signed-off-by: Jubin John <jubin.john@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>

staging/rdma/hfi1: Change send_schedule counter to a per cpu counter

A patch to fix fairness issues in QP scheduling requires
n_send_schedule counter to be converted to a per cpu counter to reduce
cache misses.

Reviewed-by: Mike Marciniszyn <mike.marciniszyn@intel.com>
Signed-off-by: Vennila Megavannan <vennila.megavannan@intel.com>
Signed-off-by: Jubin John <jubin.john@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>

staging/rdma/hfi1: Verbs Mem affinity support

Change verbs memory allocations to the device numa node.  This keeps memory
close to the device for optimal performance.

Reviewed-by: Mike Marciniszyn <mike.marciniszyn@intel.com>
Reviewed-by: Ira Weiny <ira.weiny@intel.com>
Signed-off-by: Mitko Haralanov <mitko.haralanov@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>

staging/rdma/hfi1: Allocate send ctxt on device NUMA node

Allocate the user mode send context memory on the numa node which the
device is attached to for better performance.

Reviewed-by: Dean Luick <dean.luick@intel.com>
Reviewed-by: Mike Marciniszyn <mike.marciniszyn@intel.com>
Reviewed-by: Ira Weiny <ira.weiny@intel.com>
Signed-off-by: Mitko Haralanov <mitko.haralanov@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>

staging/rdma/hfi1: Consolidate CPU/IRQ affinity support

This patch unifies the affinity support for CPU and IRQ allocations into
a single code base. The goal is to allow the driver to make intelligent
placement decision based on an overall view of processes and IRQs across
as much of the driver as possible.

Pulling all the scattered affinity code into a single code base lays the
ground work for accomplishing the above goal.  For example, previous
implementations made user process placement decision solely based on
other user processes.  This algorithm is limited as it did not take into
account IRQ placement and could result in overloading certain CPUs.

A single code base also provides a much easier way to maintain and debug
any performance issues related to affinity.

Reviewed-by: Mike Marciniszyn <mike.marciniszyn@intel.com>
Reviewed-by: Dean Luick <dean.luick@intel.com>
Reviewed-by: Ira Weiny <ira.weiny@intel.com>
Signed-off-by: Mitko Haralanov <mitko.haralanov@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>

staging/rdma/hfi1: Remove unnecessary duplicated variable

struct hfi1_devdata contained 2 variables which represented the numa
node the device is attached to.  Remove the duplicated one.

Reviewed-by: Ira Weiny <ira.weiny@intel.com>
Signed-off-by: Mitko Haralanov <mitko.haralanov@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>

staging/rdma/hfi1: Remove unused code

This comment and code was unused.  Just remove it.

Reviewed-by: Ira Weiny <ira.weiny@intel.com>
Signed-off-by: Mitko Haralanov <mitko.haralanov@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>

staging/rdma/hfi1: Fix SL->SC checks

SLs which are mapped to SC15 are invalid and should fail the
operation.

For RC/UC QP types, verify the AH information at modify_qp time and
fail the modify_qp if the SL is invalid.

For other QP types check the SL during post_send via the new rdmavt
callback.

Reviewed-by: Mike Marciniszyn <mike.marciniszyn@intel.com>
Reviewed-by: Dennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: Ira Weiny <ira.weiny@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>

staging/rdma/hfi1: Add support for enabling/disabling PCIe ASPM

hfi1 HW has a high PCIe ASPM L1 exit latency and also advertises an
acceptable latency less than actual ASPM latencies. Additional
mechanisms than those provided by BIOS/OS are therefore required to
enable/disable ASPM for hfi1 to provide acceptable power/performance
trade offs. This patch adds this support.

By means of a module parameter ASPM can be either (a) always enabled
(power save mode) (b) always disabled (performance mode) (c)
enabled/disabled dynamically. The dynamic mode implements two
heuristics to alleviate possible problems with high ASPM L1 exit
latency. ASPM is normally enabled but is disabled if (a) there are any
active user space PSM contexts, or (b) for verbs, ASPM is disabled as
interrupt activity for a context starts to increase.

A few more points about the verbs implementation. In order to reduce
lock/cache contention between multiple verbs contexts, some processing
is done at the context layer before contending for device layer
locks. ASPM is disabled when two interrupts for a context happen
within 1 millisec. A timer is scheduled which will re-enable ASPM
after 1 second should the interrupt activity cease. Normally, every
interrupt, or interrupt-pair should push the timer out
further. However, since this might increase the processing load per
interrupt, pushing the timer out is postponed for half a second. If
after half a second we get two interrupts within 1 millisec the timer
is pushed out by another second.

Finally, the kernel ASPM API is not used in this patch. This is
because this patch does several non-standard things as SW workarounds
for HW issues. As mentioned above, it enables ASPM even when advertised
actual latencies are greater than acceptable latencies. Also, whereas
the kernel API only allows drivers to disable ASPM from driver probe,
this patch enables/disables ASPM directly from interrupt context. Due
to these reasons the kernel ASPM API was not used.

Reviewed-by: Mike Marciniszyn <mike.marciniszyn@intel.com>
Reviewed-by: Dean Luick <dean.luick@intel.com>
Reviewed-by: Ira Weiny <ira.weiny@intel.com>
Signed-off-by: Ashutosh Dixit <ashutosh.dixit@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>

staging/rdma/hfi1: Method to toggle "fast ECN" detection

Add a per port sysfs paramter to toggle cc_prescan/Fast ECN Detection and
remove the Kconfig option which was previously used to control this.

While am updating the sysfs documentation, fix the name of CCMgtA.

Reviewed-by: Arthur Kepner <arthur.kepner@intel.com>
Reviewed-by: Mike Marciniszyn <mike.marciniszyn@intel.com>
Signed-off-by: Vennila Megavannan <vennila.megavannan@intel.com>
Signed-off-by: Ira Weiny <ira.weiny@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>

staging/rdma/hfi1: Correctly set RcvCtxtCtrl register

The RcvCtxtCtrl register was being incorrectly set upon context
initialization and clean up resulting, in many cases, of contexts using
settings from previous contexts' initialization. This resulted in bad
and unexpected behavior. This was especially important for the TailUpd
bit, which requires special handling and if set incorrectly could lead
to severely degraded performance.

This patch fixes the handling of the RcvCtxtCtrl register, ensuring that
each context gets initialized with settings applicable only for that
context. It also ensures the proper setting for the TailUpd bit by
setting it to either 0 or 1 (as needed by the context's configuration)
explicitly.

Reviewed-by: Ira Weiny <ira.weiny@intel.com>
Signed-off-by: Mitko Haralanov <mitko.haralanov@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>

staging/rdma/hfi1: Fix for 32-bit counter overflow in driver and hfi1stats

When 32-bit hardware counters overflow, hfi1stats misinterprets
the counters as being 64 bits causing the deltas for the
counters to be a huge number. This patch makes hfi1stats
aware that a counter is 32 bits by making the driver write
<counter name>,32 to debugfs.

Reviewed-by: Dean Luick <dean.luick@intel.com>
Signed-off-by: Sebastian Sanchez <sebastian.sanchez@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>

staging/rdma/hfi1: Skip lcb init for simulation

The simulator does not correctly handle LCB cclk loopback.
Skip that step for simulation - it is not needed.

Reviewed-by: Dennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: Dean Luick <dean.luick@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>

staging/rdma/hfi1: No firmware retry for simulation

Simulation has no firmware, so it will never move firmware
acquire to the FINAL state.  Avoid that by skiping the TRY
state and moving directly to FINAL.

Reviewed-by: Dennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: Dean Luick <dean.luick@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>

staging/rdma/hfi1: Don't attempt to qualify or tune loopback plugs

Loopback plugs used for testing hardware don't need to be qualified to
bring the link up unlike production cables. This patch adds an exception
for loopback plugs to the QSFP and SerDes tuning algortihm.

Reviewed-by: Dean Luick <dean.luick@intel.com>
Signed-off-by: Easwar Hariharan <easwar.hariharan@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>

staging/rdma/hfi1: Make firmware failure messages warnings

Make firmware validation failure and missing firmware messages
a warning since alternates can be tried.  Add an error message
when all attempts fail.

Reviewed-by: Dennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: Dean Luick <dean.luick@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>

staging/rdma/hfi1: Only warn when board description is not found

Change-Id: Icc4ad27c4c67e51df8c8a203c4f16973793678ec
Reviewed-by: Dennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: Dean Luick <dean.luick@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>

staging/rdma/hfi1: Fix per-VL transmit discard counts

Implement per-VL transmit counters.  Not all errors can be
attributed to a particular VL, so make a best attempt.
o Extend the egress error bits used to count toward transmit
  discard.
o When an egress error or send error occur, try to map back
  to a VL.
o Implement a SDMA engine to VL (back) map.
o Add per-VL port transmit counters

Reviewed-by: Mike Marciniszyn <mike.marciniszyn@intel.com>
Reviewed-by: Dennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: Dean Luick <dean.luick@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>

staging/rdma/hfi1: Fix missing firmware NULL dereference

The gen3 bump code must mark a firmware download failure as fatal.
Otherwise a later load attempt will fail with a NULL dereference.

Also:
o Only do a firmware back-off for RTL.  There are no alternates for
  FPGA or simulation.
o Rearrange OS firmware request order to match what is actually
  loaded.  This results in more coherent informational messages
  in the case of missing firmware.

Reviewed-by: Dennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: Dean Luick <dean.luick@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>

staging/rdma/hfi1: Support external device configuration requests from 8051

This patch implements support for turning on and off the clock data
recovery mechanisms implemented in QSFP cable on request by the DC 8051
on a per-lane basis.

Reviewed-by: Dean Luick <dean.luick@intel.com>
Signed-off-by: Easwar Hariharan <easwar.hariharan@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>

staging/rdma/hfi1: Get port type from configuration file

The current code employs a heuristic to guess the port type.
The canonical location to identify the port type of the
designed platform is from the platform configuration data.

This patch uses the previously fetched port type from the platform
configuration and removes the now obsolete heuristic routine
and its associated defines.

Reviewed-by: Arthur Kepner <arthur.kepner@intel.com>
Signed-off-by: Easwar Hariharan <easwar.hariharan@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>

staging/rdma/hfi1: Add active and optical cable support

This patch qualifies and tunes active and optical cables for optimal
bit error rate and signal integrity settings. These settings are
fetched from the platform configuration data.

Based on attributes of the QSFP cable as read from the SFF-8636
compliant memory map, we select the appropriate settings from the
platform configuration data (examples: TX/RX equalization, enabling
cable high power, enabling TX/RX clock data recovery mechanisms, and RX
amplitude control) and apply them to the SERDES and QSFP cable.

The platform configuration data also contains system parameters such
as maximum power dissipation supported, and the cables are qualified
based on these parameters. As part of qualifying the cables, the
correct OfflineDisabledReasons are set for the appropriate scenarios.

Reviewed-by: Dean Luick <dean.luick@intel.com>
Reviewed-by: Mike Marciniszyn <mike.marciniszyn@intel.com>
Reviewed-by: Brent R Rothermel <brent.r.rothermel@intel.com>
Signed-off-by: Easwar Hariharan <easwar.hariharan@intel.com>
Signed-off-by: Jubin John <jubin.john@intel.com>
Signed-off-by: Dennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>

staging/rdma/hfi1: Fix QSFP memory read/write across 128 byte boundary

The QSFP memory cache reads both lower and upper page 0H in one shot,
which leads to the address counter wrapping around to the beginning of
lower page 00H at byte 128, as defined by SFF-8636.
This patch fixes this by modifying the underlying QSFP read and writes
to avoid this wrap around.

Reviewed-by: Dean Luick <dean.luick@intel.com>
Reviewed-by: Ira Weiny <ira.weiny@intel.com>
Signed-off-by: Easwar Hariharan <easwar.hariharan@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>

staging/rdma/hfi1: cleanup messages on qsfp_read() failure

The ":" in "%s:" adds no value.

Reviewed-by: Dean Luick <dean.luick@intel.com>
Signed-off-by: Easwar Hariharan <easwar.hariharan@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>

staging/rdma/hfi1: HFI reports wrong offline disabled reason when cable removed

Removing QSFP cable should report 'No Local Media' instead of
'Transient' as reported by 'opaportinfo'.
Workaround is to change the state to
OPA_LINKDOWN_REASON_LOCAL_MEDIA_NOT_INSTALLED in cable handler.

With cable still removed, 'opaportinfo bounce' should not cause a
state change to Polling, as reported by 'opaportinfo'.
Resolution is to prevent physical state change from Offline->Polling.

Use a macro to mask lower nibble of OPA_LINKDOWN_REASON* as needed
for offline_disabled_reason.

Reviewed-by: Mike Marciniszyn <mike.marciniszyn@intel.com>
Reviewed-by: Easwar Hariharan <easwar.hariharan@intel.com>
Reviewed-by: Dean Luick <dean.luick@intel.com>
Reported-by: Todd Rimmer <todd.rimmer@intel.com>
Signed-off-by: Bryan Morgan <bryan.c.morgan@intel.com>
Signed-off-by: Jubin John <jubin.john@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>

staging/rdma/hfi1: Remove srq functionality

srq functionality is now in rdmavt. Remove it from the hfi1 driver.

Reviewed-by: Dennis Dalessandro <dennis.dalessandro@intel.com>
Reviewed-by: Harish Chegondi <harish.chegondi@intel.com>
Signed-off-by: Jubin John <jubin.john@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>

staging/rdma/hfi1: Remove hfi1_query_qp function

Rely on rvt_query_qp function defined in rdmavt

Reviewed-by: Dennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: Harish Chegondi <harish.chegondi@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>

staging/rdma/hfi1: Remove create and free mad agents

Get rid of create and free mad agent from the driver and use rdmavt
version.

Reviewed-by: Ira Weiny <ira.weiny@intel.com>
Signed-off-by: Dennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>

staging/rdma/hfi1: Use rdmavt device allocation function

No longer do drivers need to call into the IB core to allocate the verbs
device.  Use the functionality provided by rdmavt.

Reviewed-by: Ira Weiny <ira.weiny@intel.com>
Signed-off-by: Dennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>

staging/rdma/hfi1: Clean up register device

Now that rdmavt has solidified in its design we can clean up the driver
specific register device functions. This handles hfi1.

Reviewed-by: Ira Weiny <ira.weiny@intel.com>
Signed-off-by: Dennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>

staging/rdma/hfi1: Remove post_recv and use rdmavt version

This patch removes the simple post recv function in favor of using rdmavt.
The packet receive processing still lives in the driver though.

Reviewed-by: Harish Chegondi <harish.chegondi@intel.com>
Signed-off-by: Dennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>

staging/rdma/hfi1: Remove destroy qp verb

This removes the destroy qp verbs in favor of using rdmavt.

Reviewed-by: Harish Chegondi <harish.chegondi@intel.com>
Signed-off-by: Dennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>

staging/rdma/hfi1: Remove modify queue pair from hfi1

In addition to removing the modify queue pair verb from hfi1 we also
remove ancillary functions which existed only for modify queue pair and
are also already present in hfi1.

Reviewed-by: Mike Marciniszyn <mike.marciniszyn@intel.com>
Signed-off-by: Dennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>

staging/rdma/hfi1: Remove multicast verbs functions

Multicast is now supported by rdmavt. Remove the verbs multicast functions
and use that.

Reviewed-by: Mike Marciniszyn <mike.marciniszyn@intel.com>
Reviewed-by: Harish Chegondi <harish.chegondi@intel.com>
Signed-off-by: Dennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>

staging/rdma/hfi1: Use rdmavt version of post_send

This patch removes the post_send and post_one_send from the hfi1 driver.
The "posting" of sends will be done by rdmavt which will walk a WQE and
queue work. This patch will still provide the capability to schedule that
work as well as kick the progress. These are provided to the rdmavt layer.

Reviewed-by: Jubin John <jubin.john@intel.com>
Signed-off-by: Dean Luick <dean.luick@intel.com>
Signed-off-by: Harish Chegondi <harish.chegondi@intel.com>
Signed-off-by: Edward Mascarenhas <edward.mascarenhas@intel.com>
Signed-off-by: Mike Marciniszyn <mike.marciniszyn@intel.com>
Signed-off-by: Ira Weiny <ira.weiny@intel.com>
Signed-off-by: Dennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>

staging/rdma/hfi1: Clean up return handling

Return directly from rvt_resize_cq rather than use a goto/label.

Reviewed-by: Ira Weiny <ira.weiny@intel.com>
Reviewed-by: Harish Chegondi <harish.chegondi@intel.com>
Signed-off-by: Dennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>

staging/rdma/hfi1: Remove CQ data structures and functions from hfi1

The completion queue is not a complex data structure and it can be removed
at the same time as its functions. Unlike the more complicated queue pair
which was done in multiple patches. This single patch removes all traces
of hfi1 specific completeion queues from the hfi1 driver.

Reviewed-by: Ira Weiny <ira.weiny@intel.com>
Reviewed-by: Harish Chegondi <harish.chegondi@intel.com>
Signed-off-by: Dennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>

staging/rdma/hfi1: Remove query_device function

Removed hfi1 query_device function to use rdmavt rvt_query_device function
The rvt dev info device attributes still need to be filled in by the driver

Reviewed-by: Dennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: Harish Chegondi <harish.chegondi@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>

staging/rdma/hfi1: Remove create_qp functionality

Rely on rdmavt to provide queue pair creation.

Reviewed-by: Ira Weiny <ira.weiny@intel.com>
Reviewed-by: Harish Chegondi <harish.chegondi@intel.com>
Signed-off-by: Dennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>

staging/rdma/hfi1: Remove qpdev and qpn table from hfi1

Another change on the way to removing queue pair functionality from
hfi1. This patch removes the private queue pair structure and the table
which holds the queue pair numbers in favor of using what is provided
by rdmavt.

Reviewed-by: Ira Weiny <ira.weiny@intel.com>
Reviewed-by: Harish Chegondi <harish.chegondi@intel.com>
Signed-off-by: Dennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>

staging/rdma/hfi1: Use rdmavt send flags and recv flags

Use the definitions of the s_flags and r_flags which are now in rdmavt.

Reviewed-by: Ira Weiny <ira.weiny@intel.com>
Reviewed-by: Harish Chegondi <harish.chegondi@intel.com>
Signed-off-by: Dennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>

staging/rdma/hfi1: Remove user context allocation and de-alloction functions

IB user context alloc and dealloc functions have been added to rdmavt.
This patch removes them from hfi1.

Reviewed-by: Dennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: Harish Chegondi <harish.chegondi@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>

staging/rdma/hfi1: Use rdmavt pkey verbs function

No need to keep providing the query pkey function. This is now being
done in rdmavt. Remove support from hfi1. The allocation and
maintenance of the list still resides in the driver.

Reviewed-by: Mike Marciniszyn <mike.marciniszyn@intel.com>
Signed-off-by: Dennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>

staging/rdma/hfi1: Remove mmap from hfi1

Mmap data structure has already been moved to rdmavt and hfi1 supports
it. Now that the mmap functionality has also been moved to rdmavt its
time for hfi1 to use that as well.

Reviewed-by: Mike Marciniszyn <mike.marciniszyn@intel.com>
Signed-off-by: Dennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>

staging/rdma/hfi1: Remove ibport and use rdmavt version

Remove most of the ibport members from hfi1 and use the rdmavt version.
Also register the port with rdmavt.

Reviewed-by: Mike Marciniszyn <mike.marciniszyn@intel.com>
Reviewed-by: Harish Chegondi <harish.chegondi@intel.com>
Signed-off-by: Dennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>

staging/rdma/hfi1: Remove srq from hfi1

SRQ data structure has been moved to rdmavt. Make use of it.

Reviewed-by: Harish Chegondi <harish.chegondi@intel.com>
Signed-off-by: Dennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>

staging/rdma/hfi1: Remove hfi1 MR and hfi1 specific qp type

This patch does the actual removal of the queue pair from the hfi1 driver
along with a number of dependent data structures. These were moved to rvt.

It also removes the MR functions to use those in rdmavt.

These two pieces can not reasonably be split apart becuase they depend on
each other.

Reviewed-by: Mike Marciniszyn <mike.marciniszyn@intel.com>
Signed-off-by: Dennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>

staging/rdma/hfi1: Implement hfi1 support for AH notification

For OPA devices additional work is required to create an AH.
This patch adds support to set the VL correctly.

Reviewed-by: Mike Marciniszyn <mike.marciniszyn@intel.com>
Signed-off-by: Dennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>

staging/rdma/hfi1: Use address handle in rdmavt and remove from hfi1

Original patch from Kamal Heib <kamalh@mellanox.com>, split
apart from original and modified to accomodate recent changes
in rdmavt.

Remove AH from hfi1 and use rdmavt version.

Signed-off-by: Kamal Heib <kamalh@mellanox.com>
Signed-off-by: Dennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>

staging/rdma/hfi1: Use correct rdmavt header files after move.

Rdmavt split the header files to be based on ibta object. This patch
makes changes in hfi1 to account for the move.

Reviewed-by: Mike Marciniszyn <mike.marciniszyn@intel.com>
Signed-off-by: Dennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>

staging/rdma/hfi1: Add device specific info prints

Implement get_card_name and get_pci_dev helper functions for rdmavt
for hfi1.

Reviewed-by: Mike Marciniszyn <mike.marciniszyn@intel.com>
Reviewed-by: Ira Weiny <ira.weiny@intel.com>
Signed-off-by: Dennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>

staging/rdma/hfi1: Remove driver specific members from hfi1 qp type

In preparation for moving the queue pair data structure to rdmavt the
members of the driver specific queue pairs which are not common need to be
pushed off to a private driver structure. This structure will be available
in the queue pair once moved to rdmavt as a void pointer. This patch while
not adding a lot of value in and of itself is a prerequisite to move the
queue pair out of the drivers and into rdmavt.

The driver specific, private queue pair data structure should condense as
more of the send side code moves to rdmavt.

Reviewed-by: Mike Marciniszyn <mike.marciniszyn@intel.com>
Signed-off-by: Dennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>

staging/rdma/hfi1: Remove MR data structures from hfi1

Remove MR data structures from hfi1 and use the version in rdmavt

Reviewed-by: Dean Luick <dean.luick@intel.com>
Reviewed-by: Mike Marciniszyn <mike.marciniszyn@intel.com>
Signed-off-by: Dennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>

staging/rdma/hfi1: Use rdmavt protection domain

Remove protection domain from hfi1 and use rdmavt's version.

Reviewed-by: Ira Weiny <ira.weiny@intel.com>
Reviewed-by: Mike Marciniszyn <mike.marciniszyn@intel.com>
Signed-off-by: Dennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>

staging/rdma/hfi1: Consolidate dma ops for hfi1

Remove the dma.c file from hfi1 in favor of using that which is
present in rdmavt.

Reviewed-by: Ira Weiny <ira.weiny@intel.com>
Reviewed-by: Mike Marciniszyn <mike.marciniszyn@intel.com>
Signed-off-by: Dennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>

staging/rdma/hfi1: Add basic rdmavt capability flags for hfi1

Most functionality is still being done in the driver, set flags so that
rdmavt will let hfi1 continue to handle mr, qp, and cq init.

Reviewed-by: Mike Marciniszyn <mike.marciniszyn@intel.com>
Signed-off-by: Dennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>

staging/rdma/hfi1: Begin to use rdmavt for verbs

This patch begins to make use of rdmavt by registering with it and
providing access to the header files. This is just the beginning of
rdmavt support in hfi1.

Reviewed-by: Ira Weiny <ira.weiny@intel.com>
Reviewed-by: Mike Marciniszyn <mike.marciniszyn@intel.com>
Signed-off-by: Dennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>

IB/qib: Remove modify_port and port_immutable functions

Delete code from query_port which has been moved into rvt_query_port
Create a call back function to shut down a port which may be called from
rvt_modify_port

Signed-off-by: Harish Chegondi <harish.chegondi@intel.com>
Signed-off-by: Dennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>

IB/qib: Support query gid in rdmavt

Query gid is in rdmavt, but still relies on the driver to maintain the
guid table. Add the necessary driver call back and remove the existing
verb handler.

Reviewed-by: Harish Chegondi <harish.chegondi@intel.com>
Signed-off-by: Dennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>

IB/qib: Remove destroy queue pair code

Destroy QP functionality in rdmavt will be used instead.
Remove the remove_qp function being called exclusively by destroy qp code.

Reviewed-by: Dennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: Harish Chegondi <harish.chegondi@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>

IB/qib: Remove modify queue pair code

Modify queue pair functionality in rdmavt will be used instead.
Remove ancillary functions which are being used by modify QP code.

Reviewed-by: Dennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: Harish Chegondi <harish.chegondi@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>

IB/qib: Remove qib_lookup_qpn and use rvt_lookup_qpn instead

Add calls to rcu_read_lock()/rcu_read_unlock() as rvt_lookup_qpn callers
must hold the rcu_read_lock before calling and keep the lock until the
returned qp is no longer in use.

Remove lookaside qp and some qp refcount atomics in the sdma send code
that is redundant with the s_dma_busy refcount, which will also stall
the state processing to the reset state.

Change the qpn hash function to hash_32 which is hash function used
in rvt_lookup_qpn. qpn_hash function would be eliminated in later patches.

Reviewed-by: Ira Weiny <ira.weiny@intel.com>
Reviewed-by: Mike Marciniszyn <mike.marciniszyn@intel.com>
Reviewed-by: Dennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: Harish Chegondi <harish.chegondi@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>

IB/qib: Clean up register_ib_device

Remove some of the unnecessary code from qib_register_ib_device.

Reviewed-by: Ira Weiny <ira.weiny@intel.com>
Reviewed-by: Harish Chegondi <harish.chegondi@intel.com>
Signed-off-by: Dennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>

IB/qib: Remove srq functionality

srq functionality is now in rdmavt. Remove it from the qib driver.

Reviewed-by: Dennis Dalessandro <dennis.dalessandro@intel.com>
Reviewed-by: Harish Chegondi <harish.chegondi@intel.com>
Signed-off-by: Jubin John <jubin.john@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>

IB/rdmavt: Properly pass gfp to hw driver function

alloc_qpn must use GFP and the hardware drivers should use it as well.

Reviewed-by: Mike Marciniszyn <mike.marciniszyn@intel.com>
Signed-off-by: Ira Weiny <ira.weiny@intel.com>
Signed-off-by: Dennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>

IB/rdmavt: Add support for query_port, modify_port and get_port_immutable

rvt_query_port calls into the driver through a call back function
query_port_state to populate the rest of ib_port_attr elements.
rvt_modify_port calls into the driver if needed through a call back
function shut_down_port()

Signed-off-by: Harish Chegondi <harish.chegondi@intel.com>
Signed-off-by: Dennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>

IB/rdmavt: Add query gid support.

Addin query gid support. Rdmavt still relies on the driver to maintain
the gid table. Rdmavt simply calls into the driver to retrive the guid
for a particular port.

Reviewed-by: Harish Chegondi <harish.chegondi@intel.com>
Signed-off-by: Dennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>

IB/rdmavt: Clean up distinction between port number and index

IB core uses 1 relative indexing for ports. All of our data structures
use 0 based indexing. Add an inline function that we can use whenever we
need to validate a legal value and try to convert a port number to a
port index at the entrance into rdmavt.

Try to follow the policy that when we are talking about a port from IB
core point of view we refer to it as a port number. When port is an
index into our arrays refer to it as a port index.

Reviewed-by: Mike Marciniszyn <mike.marciniszyn@intel.com>
Reviewed-by: Harish Chegondi <harish.chegondi@intel.com>
Signed-off-by: Dennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>

IB/rdmavt: Add Mem affinity support

Change verbs memory allocations to the device numa node.  This keeps memory
close to the device for optimal performance.

Reviewed-by: Dean Luick <dean.luick@intel.com>
Reviewed-by: Mike Marciniszyn <mike.marciniszyn@intel.com>
Reviewed-by: Ira Weiny <ira.weiny@intel.com>
Reviewed-by: Dennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: Mitko Haralanov <mitko.haralanov@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>

IB/rdmavt: Add hardware driver send work request check

Some hardware drivers requires additional checks on send WRs.  Create an
optional call back to allow hardware drivers to reject a send WR.

Reviewed-by: Dennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: Ira Weiny <ira.weiny@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>

IB/rdmavt: Add srq functionality to rdmavt

Fill in srq function stubs with code derived from hfi1 and qib.
Move necessary functions and data structure members as well.

Reviewed-by: Dennis Dalessandro <dennis.dalessandro@intel.com>
Reviewed-by: Harish Chegondi <harish.chegondi@intel.com>
Signed-off-by: Jubin John <jubin.john@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>

IB/qib: Remove qib_query_qp function

Rely on rvt_query_qp function defined in rdmavt

Reviewed-by: Dennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: Harish Chegondi <harish.chegondi@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>

IB/qib: Remove qib multicast verbs functions

Multicast is now supported by rdmavt. Remove the verbs multicast functions
and use that.

Reviewed-by: Dennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: Harish Chegondi <harish.chegondi@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>

IB/qib: Remove qib_post_receive and use rdmavt version

This patch removes the simple post recv function in favor of using rdmavt.
The packet receive processing still lives in the driver though.

Reviewed-by: Dennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: Harish Chegondi <harish.chegondi@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>

IB/qib: Use rdmavt version of post_send

This patch removes the post_send and post_one_send from the qib driver.
The "posting" of sends will be done by rdmavt which will walk a WQE and
queue work. This patch will still provide the capability to schedule that
work as well as kick the progress. These are provided to the rdmavt layer.

Reviewed-by: Dennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: Harish Chegondi <harish.chegondi@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>

IB/qib: Remove completion queue data structures and functions from qib

Use the completion queue functionality provided by rdmavt.

Reviewed-by: Dennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: Harish Chegondi <harish.chegondi@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>