Depending on the OS page-table configurations, applications
may request MRs which has page size alignment other than 4K

Underlying provider driver needs to adjust its PBL boundaries
according to the incoming page boundaries in the PA list.

Adding a capability to register MRs having pages-sizes other
than 4K (Hugepages).
Signed-off-by: NSomnath Kotur <somnath.kotur@broadcom.com>
Signed-off-by: NSelvin Xavier <selvin.xavier@broadcom.com>
Signed-off-by: NDevesh Sharma <devesh.sharma@broadcom.com>
Signed-off-by: NDoug Ledford <dledford@redhat.com>

The device now reports firmware version thus, removing
the hard coded values of the FW version string and
redundant fw_rev hook from sysfs. Adding code to query
firmware version from underlying device and report it
through the kernel verb to get firmware version string.
Signed-off-by: NSelvin Xavier <selvin.xavier@broadcom.com>
Signed-off-by: NDevesh Sharma <devesh.sharma@broadcom.com>
Signed-off-by: NDoug Ledford <dledford@redhat.com>

RoCE can be used by virtual functions (VFs) as well. Adding
code changes to allow resource reservation, initialization
and avail the resources to the RDMA applications running on
those VFs.

Currently, fifty percent of the total available resources
are reserved for PF and remaining are equally divided among
active VFs.
Signed-off-by: NSelvin Xavier <selvin.xavier@broadcom.com>
Signed-off-by: NDevesh Sharma <devesh.sharma@broadcom.com>
Signed-off-by: NDoug Ledford <dledford@redhat.com>

The iWARP Exception Queue (IEQ) resources are not freed when a QP is
destroyed. Fix this by freeing IEQ resources when freeing QP resources.

Fixes: d3749841 ("i40iw: add files for iwarp interface")
Signed-off-by: NMustafa Ismail <mustafa.ismail@intel.com>
Signed-off-by: NShiraz Saleem <shiraz.saleem@intel.com>
Signed-off-by: NJason Gunthorpe <jgg@mellanox.com>

Remove setting of rem_addr.len before calling iw_rdma_write,
iw_inline_rdma_write and rdma_read. rem_addr.len is not used in those
functions.
Signed-off-by: NMustafa Ismail <mustafa.ismail@intel.com>
Signed-off-by: NShiraz Saleem <shiraz.saleem@intel.com>
Signed-off-by: NJason Gunthorpe <jgg@mellanox.com>

Currently, if the number of processed Asynchronous Event Queue (AEQ)
entries exceeds 255, they are not returned to HW for re-use. During
scale-up, the unreturned AEQ entries can grow to the max AEQ size and
cause the HW to report an AEQ overflow.

Remove the check which limits the number of processed AEQ entries returned
to HW.

Fixes: 86dbcd0f ("RDMA/i40iw: add file to handle cqp calls")
Signed-off-by: NSindhu Devale <sindhu.devale@intel.com>
Signed-off-by: NShiraz Saleem <shiraz.saleem@intel.com>
Signed-off-by: NJason Gunthorpe <jgg@mellanox.com>

If the application invalidates the MR before the FMR WR, HW parses the
consumer key portion of the stag and returns an invalid stag key
Asynchronous Event (AE) that tears down the QP.

Fix this by zeroing-out the consumer key portion of the allocated stag
returned to application for FMR.

Fixes: ee855d3b93f3 ("RDMA/i40iw: Add base memory management extensions")
Signed-off-by: NShiraz Saleem <shiraz.saleem@intel.com>
Signed-off-by: NJason Gunthorpe <jgg@mellanox.com>

Remove redundant estimate SD function call.  sd_needed should already be
updated at the end of the do while resource reduction loop.
Signed-off-by: NShiraz Saleem <shiraz.saleem@intel.com>
Signed-off-by: NJason Gunthorpe <jgg@mellanox.com>

This patch assign a guid(Global Unique identifer) value to the hip08
device.
Signed-off-by: NLijun Ou <oulijun@huawei.com>
Signed-off-by: NYixian Liu <liuyixian@huawei.com>
Signed-off-by: NWei Hu (Xavier) <xavier.huwei@huawei.com>
Reviewed-by: NLeon Romanovsky <leonro@mellanox.com>
Signed-off-by: NJason Gunthorpe <jgg@mellanox.com>

If the port is a RoCEv2 port, the remote port address and QP information
which returned for UD will be modified.
Signed-off-by: NLijun Ou <oulijun@huawei.com>
Signed-off-by: NYixian Liu <liuyixian@huawei.com>
Signed-off-by: NWei Hu (Xavier) <xavier.huwei@huawei.com>
Signed-off-by: NJason Gunthorpe <jgg@mellanox.com>

Because pkey is fixed for hip08 RoCE, it needs to assign zero for
pkey_index of wc. otherwise, it will happen an error when establishing
connection by communication management mechanism.
Signed-off-by: NLijun Ou <oulijun@huawei.com>
Signed-off-by: NYixian Liu <liuyixian@huawei.com>
Signed-off-by: NWei Hu (Xavier) <xavier.huwei@huawei.com>
Signed-off-by: NJason Gunthorpe <jgg@mellanox.com>

This patch mainly configure the fields of sq wqe of ud type when posting
wr of gsi qp type.
Signed-off-by: NLijun Ou <oulijun@huawei.com>
Signed-off-by: NYixian Liu <liuyixian@huawei.com>
Signed-off-by: NWei Hu (Xavier) <xavier.huwei@huawei.com>
Signed-off-by: NJason Gunthorpe <jgg@mellanox.com>

It needs to Assign the values for some fields in qp context when qp type
is gsi qp type in hip08.
Signed-off-by: NLijun Ou <oulijun@huawei.com>
Signed-off-by: NYixian Liu <liuyixian@huawei.com>
Signed-off-by: NWei Hu (Xavier) <xavier.huwei@huawei.com>
Signed-off-by: NJason Gunthorpe <jgg@mellanox.com>

The gsi qp and rc qp use the same qp context structure and the created
flow, only differentiate them by qpn and qp type.
Signed-off-by: NLijun Ou <oulijun@huawei.com>
Signed-off-by: NYixian Liu <liuyixian@huawei.com>
Signed-off-by: NWei Hu (Xavier) <xavier.huwei@huawei.com>
Signed-off-by: NJason Gunthorpe <jgg@mellanox.com>

When modifying qp from init to init, it need to assign the cqn of send cq
for tx cqn field of qp context. Otherwise, it will cause a mistake when
the send and recv cq sizes are different.
Signed-off-by: NLijun Ou <oulijun@huawei.com>
Signed-off-by: NJason Gunthorpe <jgg@mellanox.com>

gcc-8 reports

drivers/infiniband/core/cma_configfs.c: In function 'make_cma_dev':
./include/linux/string.h:245:9: warning: '__builtin_strncpy' specified
bound 64 equals destination size [-Wstringop-truncation]

We need to use strlcpy() to make sure the string is nul-terminated.
Signed-off-by: NXiongfeng Wang <xiongfeng.wang@linaro.org>
Reviewed-by: NLeon Romanovsky <leonro@mellanox.com>
Signed-off-by: NJason Gunthorpe <jgg@mellanox.com>

iWARP does not use rdma_ah_attr_type, and for this reason we do not have a
RDMA_AH_ATTR_TYPE_IWARP. rdma_ah_find_type should not even be called on iwarp
ports and for clarity it shouldn't have a special test for iWarp.

This changes the result from RDMA_AH_ATTR_TYPE_ROCE to RDMA_AH_ATTR_TYPE_IB
when wrongly called on an iWarp port.

Fixes: 44c58487 ("IB/core: Define 'ib' and 'roce' rdma_ah_attr types")
Signed-off-by: NParav Pandit <parav@mellanox.com>
Signed-off-by: NLeon Romanovsky <leon@kernel.org>
Signed-off-by: NJason Gunthorpe <jgg@mellanox.com>

The change of slid from u16 to u32 results in sizeof(struct ib_wc)
cross 64B boundary, which causes more cache misses. This patch
rearranges the fields and remain the size to 64B.

Pahole output before this change:

struct ib_wc {
        union {
                u64                wr_id;                /*           8 */
                struct ib_cqe *    wr_cqe;               /*           8 */
        };                                               /*     0     8 */
        enum ib_wc_status          status;               /*     8     4 */
        enum ib_wc_opcode          opcode;               /*    12     4 */
        u32                        vendor_err;           /*    16     4 */
        u32                        byte_len;             /*    20     4 */
        struct ib_qp *             qp;                   /*    24     8 */
        union {
                __be32             imm_data;             /*           4 */
                u32                invalidate_rkey;      /*           4 */
        } ex;                                            /*    32     4 */
        u32                        src_qp;               /*    36     4 */
        int                        wc_flags;             /*    40     4 */
        u16                        pkey_index;           /*    44     2 */

        /* XXX 2 bytes hole, try to pack */

        u32                        slid;                 /*    48     4 */
        u8                         sl;                   /*    52     1 */
        u8                         dlid_path_bits;       /*    53     1 */
        u8                         port_num;             /*    54     1 */
        u8                         smac[6];              /*    55     6 */

        /* XXX 1 byte hole, try to pack */

        u16                        vlan_id;              /*    62     2 */
        /* --- cacheline 1 boundary (64 bytes) --- */
        u8                         network_hdr_type;     /*    64     1 */

        /* size: 72, cachelines: 2, members: 17 */
        /* sum members: 62, holes: 2, sum holes: 3 */
        /* padding: 7 */
        /* last cacheline: 8 bytes */
};

Pahole output after this change:

struct ib_wc {
        union {
                u64                wr_id;                /*           8 */
                struct ib_cqe *    wr_cqe;               /*           8 */
        };                                               /*     0     8 */
        enum ib_wc_status          status;               /*     8     4 */
        enum ib_wc_opcode          opcode;               /*    12     4 */
        u32                        vendor_err;           /*    16     4 */
        u32                        byte_len;             /*    20     4 */
        struct ib_qp *             qp;                   /*    24     8 */
        union {
                __be32             imm_data;             /*           4 */
                u32                invalidate_rkey;      /*           4 */
        } ex;                                            /*    32     4 */
        u32                        src_qp;               /*    36     4 */
        u32                        slid;                 /*    40     4 */
        int                        wc_flags;             /*    44     4 */
        u16                        pkey_index;           /*    48     2 */
        u8                         sl;                   /*    50     1 */
        u8                         dlid_path_bits;       /*    51     1 */
        u8                         port_num;             /*    52     1 */
        u8                         smac[6];              /*    53     6 */

        /* XXX 1 byte hole, try to pack */

        u16                        vlan_id;              /*    60     2 */
        u8                         network_hdr_type;     /*    62     1 */

        /* size: 64, cachelines: 1, members: 17 */
        /* sum members: 62, holes: 1, sum holes: 1 */
        /* padding: 1 */
};

Cc: <stable@vger.kernel.org> # v4.13
Fixes: 7db20ecd ("IB/core: Change wc.slid from 16 to 32 bits")
Signed-off-by: NBodong Wang <bodong@mellanox.com>
Reviewed-by: NParav Pandit <parav@mellanox.com>
Signed-off-by: NLeon Romanovsky <leon@kernel.org>
Signed-off-by: NJason Gunthorpe <jgg@mellanox.com>

Allocating steerable UD QPs depends on having at least one IB port,
while releasing those QPs does not.

As a result, when there are only ETH ports, the IB (RoCE) driver
requests releasing a qp range whose base qp is zero, with
qp count zero.

When SR-IOV is enabled, and the VF driver is running on a VM over
a hypervisor which treats such qp release calls as errors
(rather than NOPs), we see lines in the VM message log like:

 mlx4_core 0002:00:02.0: Failed to release qp range base:0 cnt:0

Fix this by adding a check for a zero count in mlx4_release_qp_range()
(which thus treats releasing 0 qps as a nop), and eliminating the
check for device managed flow steering when releasing steerable UD QPs.
(Freeing ib_uc_qpns_bitmap unconditionally is also OK, since it
remains NULL when steerable UD QPs are not allocated).

Cc: <stable@vger.kernel.org>
Fixes: 4196670b ("IB/mlx4: Don't allocate range of steerable UD QPs for Ethernet-only device")
Signed-off-by: NJack Morgenstein <jackm@dev.mellanox.co.il>
Signed-off-by: NLeon Romanovsky <leon@kernel.org>
Signed-off-by: NJason Gunthorpe <jgg@mellanox.com>

The double swap matches what user space rdma-core does to imm_data.

wc->imm_data is not used in the kernel so this change has no practical
impact.
Acked-by: NMichal Kalderon <michal.kalderon@cavium.com>
Signed-off-by: NJason Gunthorpe <jgg@mellanox.com>

This matches the changes made recently to the userspace hns
driver when it was made sparse clean.

See rdma-core commit bffd380cfe56 ("libhns: Make the provider sparse
clean")

wc->imm_data is not used in the kernel so this change has no practical
impact.
Signed-off-by: NJason Gunthorpe <jgg@mellanox.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>

Resolving DMAC for RoCE is applicable to only Connected mode QPs.
So resolve DMAC for only for Connected mode QPs.
Signed-off-by: NParav Pandit <parav@mellanox.com>
Reviewed-by: NDaniel Jurgens <danielj@mellanox.com>
Signed-off-by: NLeon Romanovsky <leon@kernel.org>
Signed-off-by: NJason Gunthorpe <jgg@mellanox.com>

Instead of returning 0 (success) for RoCE scenarios where DMAC should
not be resolved, avoid such attempt and make code consistent with
ib_create_user_ah().
Signed-off-by: NParav Pandit <parav@mellanox.com>
Reviewed-by: NDaniel Jurgens <danielj@mellanox.com>
Signed-off-by: NLeon Romanovsky <leon@kernel.org>
Signed-off-by: NJason Gunthorpe <jgg@mellanox.com>

Currently ah_attr is initialized by the ib_cm layer for rdma_cm
based applications. For RoCE transport ah_attr.roce.dmac is already
initialized by ib_cm, rdma_cm either from wc, path record, route
resolve, explicit path record setting depending on active or passive
side QP. Therefore avoid resolving DMAC for QP of kernel consumers.
Signed-off-by: NParav Pandit <parav@mellanox.com>
Reviewed-by: NDaniel Jurgens <danielj@mellanox.com>
Signed-off-by: NLeon Romanovsky <leon@kernel.org>
Signed-off-by: NJason Gunthorpe <jgg@mellanox.com>

Currently qp->port stores the port number whenever IB_QP_PORT
QP attribute mask is set (during QP state transition to INIT state).
This port number should be stored for the real QP when XRC target QP
is used.

Follow the ib_modify_qp() implementation and hide the access to ->real_qp.

Fixes: a512c2fb ("IB/core: Introduce modify QP operation with udata")
Signed-off-by: NParav Pandit <parav@mellanox.com>
Reviewed-by: NDaniel Jurgens <danielj@mellanox.com>
Signed-off-by: NLeon Romanovsky <leon@kernel.org>
Signed-off-by: NJason Gunthorpe <jgg@mellanox.com>

Use correct parameter names and formatting in function kernel-doc notation
to eliminate warnings from scripts/kernel-doc.

../drivers/infiniband/sw/rdmavt/mr.c:784: warning: Excess function parameter 'ibmfr' description in 'rvt_map_phys_fmr'
../drivers/infiniband/sw/rdmavt/vt.c:234: warning: Excess function parameter 'intex' description in 'rvt_query_pkey'
../drivers/infiniband/sw/rdmavt/vt.c:266: warning: Excess function parameter 'index' description in 'rvt_query_gid'
../drivers/infiniband/sw/rdmavt/vt.c:306: warning: Excess function parameter 'data' description in 'rvt_alloc_ucontext'
../drivers/infiniband/sw/rdmavt/cq.c:65: warning: Excess function parameter 'sig' description in 'rvt_cq_enter'
../drivers/infiniband/sw/rdmavt/qp.c:279: warning: Excess function parameter 'qpt' description in 'rvt_free_all_qps'
../drivers/infiniband/sw/rdmavt/mcast.c:282: warning: Excess function parameter 'igd' description in 'rvt_attach_mcast'
../drivers/infiniband/sw/rdmavt/mcast.c:345: warning: Excess function parameter 'igd' description in 'rvt_detach_mcast'
Signed-off-by: NRandy Dunlap <rdunlap@infradead.org>
Cc: Doug Ledford <dledford@redhat.com>
Cc: Jason Gunthorpe <jgg@mellanox.com>
Cc: linux-doc@vger.kernel.org
Acked-by: NDennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: NJason Gunthorpe <jgg@mellanox.com>

Use correct parameter name and description in kernel-doc notation to
eliminate a kernel-doc warning.

../drivers/infiniband/ulp/opa_vnic/opa_vnic_vema.c:730: warning: Excess function parameter 'cport' description in 'opa_vnic_vema_send_trap'
Signed-off-by: NRandy Dunlap <rdunlap@infradead.org>
Cc: Doug Ledford <dledford@redhat.com>
Cc: Jason Gunthorpe <jgg@mellanox.com>
Cc: linux-doc@vger.kernel.org
Acked-by: NDennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: NJason Gunthorpe <jgg@mellanox.com>

Fix kernel-doc warning for ib_fmr_pool_map_phys() and also format it
with function description and text spacing.

../drivers/infiniband/core/fmr_pool.c:404: warning: Excess function parameter 'pool' description in 'ib_fmr_pool_map_phys'
Signed-off-by: NRandy Dunlap <rdunlap@infradead.org>
Cc: Doug Ledford <dledford@redhat.com>
Cc: Jason Gunthorpe <jgg@mellanox.com>
Cc: linux-doc@vger.kernel.org
Signed-off-by: NJason Gunthorpe <jgg@mellanox.com>

Change function parameter name in kernel-doc notation and other comments
to eliminate a kernel-doc warning.

../drivers/infiniband/core/verbs.c:1790: warning: Excess function parameter 'wq_init_attr' description in 'ib_create_wq'
Signed-off-by: NRandy Dunlap <rdunlap@infradead.org>
Cc: Doug Ledford <dledford@redhat.com>
Cc: Jason Gunthorpe <jgg@mellanox.com>
Cc: linux-doc@vger.kernel.org
Signed-off-by: NJason Gunthorpe <jgg@mellanox.com>

The 'if' logic in ucma_query_path was broken with OPA was introduced
and started to treat RoCE paths as as OPA paths. Invert the logic
of the 'if' so only OPA paths are treated as OPA paths.

Otherwise the path records returned to rdma_cma users are mangled
when in RoCE mode.

Fixes: 57520751 ("IB/SA: Add OPA path record type")
Signed-off-by: NParav Pandit <parav@mellanox.com>
Reviewed-by: NMark Bloch <markb@mellanox.com>
Signed-off-by: NLeon Romanovsky <leon@kernel.org>
Signed-off-by: NJason Gunthorpe <jgg@mellanox.com>

rdma_set_ib_path() missed setting path record fields for RoCE
transport when RoCE support was added.

This results in setting incorrect ndev, destination mac address,
incorrect GID type etc errors when user space attempts to set a raw
IB path using the roce IB path compatibility mapping from userspace.

Fixes: 3c86aa70 ("RDMA/cm: Add RDMA CM support for IBoE devices")
Signed-off-by: NParav Pandit <parav@mellanox.com>
Reviewed-by: NMark Bloch <markb@mellanox.com>
Signed-off-by: NLeon Romanovsky <leon@kernel.org>
Signed-off-by: NJason Gunthorpe <jgg@mellanox.com>

Since 2006 there has been no user of rdmacm based application to make use
of setting multiple path records using rdma_set_ib_paths API.

Therefore code is simplified to allow setting one path record entry.
Now that it sets only single path, it is renamed to reflect the same.
Signed-off-by: NParav Pandit <parav@mellanox.com>
Reviewed-by: NMark Bloch <markb@mellanox.com>
Signed-off-by: NLeon Romanovsky <leon@kernel.org>
Signed-off-by: NJason Gunthorpe <jgg@mellanox.com>

Introduce a helper function to set path record L2 fields for RoCE.
This includes setting GID type, destination mac address and netdev
ifindex.
Signed-off-by: NParav Pandit <parav@mellanox.com>
Reviewed-by: NMark Bloch <markb@mellanox.com>
Signed-off-by: NLeon Romanovsky <leon@kernel.org>
Signed-off-by: NJason Gunthorpe <jgg@mellanox.com>

The net namespace is set in addr during create_rdma_id(),
cma_resolve_iboe_route() should use that instead of the
init namespace.

The original code was added in commit fa20105e ("IB/cma: Add support
for network namespaces"), but this path wasn't in use back then.

This patch updates the code to use right namespace, as preparation
for improving namespace support.
Signed-off-by: NParav Pandit <parav@mellanox.com>
Reviewed-by: NMark Bloch <markb@mellanox.com>
Signed-off-by: NLeon Romanovsky <leon@kernel.org>
Signed-off-by: NJason Gunthorpe <jgg@mellanox.com>

There is a need to increase number of possible char devices to support
large number of SR-IOV instances. The current limit is in the range of
64-128 devices/ports. Increase it to support up to 1024.

The patch performs the following steps to refactor the code:
1. Removes the split bitmap for fixed and overflow dev numbers.
2. Pre-allocates the non-legacy major number range during driver
   initialization, choosen for simplicity.
3. Add new define (RDMA_MAX_PORTS) that is shared between all drivers.
   This is the maximum total number of ports on all struct ib_devices.
4. Set RDMA_MAX_PORTS to 1024.
Signed-off-by: NHuy Nguyen <huyn@mellanox.com>
Signed-off-by: NLeon Romanovsky <leon@kernel.org>
Signed-off-by: NJason Gunthorpe <jgg@mellanox.com>

Remove the locks that protect character device bitmaps of
uverbs, umad and issm.

The character device bitmaps are accessed in "client->add" and
"client->remove" calls from ib_register_device and ib_unregister_device
respectively. These calls are already protected by the "device_mutex"
mutex. Thus, the spinlocks are not needed.
Signed-off-by: NHuy Nguyen <huyn@mellanox.com>
Signed-off-by: NLeon Romanovsky <leon@kernel.org>
Signed-off-by: NJason Gunthorpe <jgg@mellanox.com>

The following sequence:
* Change queue pair state into IB_QPS_ERR.
* Post a work request on the queue pair.

Triggers the following race condition in the rdma_rxe driver:
* rxe_qp_error() triggers an asynchronous call of rxe_completer(), the function
  that examines the QP send queue.
* rxe_post_send() posts a work request on the QP send queue.

If rxe_completer() runs prior to rxe_post_send(), it will drain the send
queue and the driver will assume no further action is necessary.
However, once we post the send to the send queue, because the queue is
in error, no send completion will ever happen and the send will get
stuck.  In order to process the send, we need to make sure that
rxe_completer() gets run after a send is posted to a queue pair in an
error state.  This patch ensures that happens.
Signed-off-by: NBart Van Assche <bart.vanassche@wdc.com>
Cc: Moni Shoua <monis@mellanox.com>
Cc: <stable@vger.kernel.org> # v4.8
Signed-off-by: NDoug Ledford <dledford@redhat.com>

The initial assignment to mdev is redundant as mdev is re-assigned
later and the first assigned value is never read. Remove this
redundant assignment.

Cleans up clang warning:
drivers/infiniband/hw/mlx5/main.c:359:24: warning: Value stored
to 'mdev' during its initialization is never read
Signed-off-by: NColin Ian King <colin.king@canonical.com>
Acked-by: NLeon Romanovsky <leonro@mellanox.com>
Signed-off-by: NDoug Ledford <dledford@redhat.com>

In xmit, there is a skb_clone. This function copies the struct sk_buff.
And some parameters are changed to the new skb. Then the new skb is sent
while the old skb is freed.

While the function skb_clone is removed, the parameter changes are made on
the old skb, then the old skb is sent. It can also work well.

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.

As 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 result on Server.
Before:
8192000 bytes in 0.88 seconds = 74.36 Mbit/sec
1000 iters in 0.88 seconds = 881.30 usec/iter

After:
8192000 bytes in 0.81 seconds = 81.15 Mbit/sec
1000 iters in 0.81 seconds = 807.62 usec/iter

The throughput is enhanced and the latency is reduced.

CC: Srinivas Eeda <srinivas.eeda@oracle.com>
CC: Joe Jin <joe.jin@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>