Operations with poll groups list must be protected by rtransport->lock.
Make rtranposrt->lock recursive to avoid unnecessary mutex operations when
the poll group is being destroyed within spdk_nvmf_rdma_poll_group_create
Change-Id: If0856429c10ad3bfcc9942da613796cc86d68d8d
Signed-off-by: Alexey Marchuk <alexeymar@mellanox.com>
Signed-off-by: Sasha Kotchubievsky <sashakot@mellanox.com>
Signed-off-by: Evgenii Kochetov <evgeniik@mellanox.com>
Reviewed-on: https://review.gerrithub.io/c/spdk/spdk/+/468992
Reviewed-by: Shuhei Matsumoto <shuhei.matsumoto.xt@hitachi.com>
Reviewed-by: Jim Harris <james.r.harris@intel.com>
Reviewed-by: Ben Walker <benjamin.walker@intel.com>
Tested-by: SPDK CI Jenkins <sys_sgci@intel.com>
This issue can be reproduced on fedora30.
Add assert here is enough to fix this kind of warning.
Error log:
rdma.c:3070:20: warning: Access to field 'data_buf_pool' results in a
dereference of a null pointer (loaded from field 'transport')
spdk_mempool_put(group->transport->data_buf_pool, buf);
^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1 warning generated.
This is to fix issue #965.
Change-Id: Ifb742ab914ee9a0381dca0bb769ba8aa564c816f
Signed-off-by: yidong0635 <dongx.yi@intel.com>
Reviewed-on: https://review.gerrithub.io/c/spdk/spdk/+/468908
Tested-by: SPDK CI Jenkins <sys_sgci@intel.com>
Community-CI: Broadcom SPDK FC-NVMe CI <spdk-ci.pdl@broadcom.com>
Reviewed-by: Seth Howell <seth.howell@intel.com>
Reviewed-by: Ben Walker <benjamin.walker@intel.com>
Reviewed-by: Jim Harris <james.r.harris@intel.com>
Update transaction length wrt to medata size
Change buffers handling in the case of enabled DIF - add function nvmf_rdma_fill_buffer_with_md_interleave to split SGL into several parts with metadata blocks between them in order to perform RDMA operation with appropriate offsets
Add DIF generation before executing bdev IO operation
Add parsing of DifInsertOrStrip config parameter.
Since there is a limitation on the number of entries in SG list (16), the current approach has a limitation on the max transaction size which depends on the data block size. E.g. if data block size is 512 bytes then the maximum transaction size will be 512 * 16 = 8192 bytes.
In adiition, the size of IO buffer (IOUnitSize conf param) must be aligned to metadata size for better perfromance since metadata is treated as part of this buffer. E.g. if the initiator uses transaction size = 4096, data block size on nvme disk is 512 then IO buffer size should be aligned to (512 + 8) which is 4160. In other case an extra IO buffer will be consumed which will increase the number of entries in SGL and in iov.
Change-Id: I7ad2270fe9dcceb114ece34675eac44e5783a0d5
Signed-off-by: Alexey Marchuk <alexeymar@mellanox.com>
Signed-off-by: Sasha Kotchubievsky <sashakot@mellanox.com>
Signed-off-by: Evgenii Kochetov <evgeniik@mellanox.com>
Reviewed-on: https://review.gerrithub.io/c/spdk/spdk/+/465248
Tested-by: SPDK CI Jenkins <sys_sgci@intel.com>
Reviewed-by: Ben Walker <benjamin.walker@intel.com>
Reviewed-by: Seth Howell <seth.howell@intel.com>
Reviewed-by: Shuhei Matsumoto <shuhei.matsumoto.xt@hitachi.com>
Reviewed-by: Changpeng Liu <changpeng.liu@intel.com>
Now code always return 0 , do this like nvme_rdma_mr_map_notify.
That callback can get the right return.
Change-Id: Ief2924e14321b2062f6001e7ae3f50d507206594
Signed-off-by: yidong0635 <dongx.yi@intel.com>
Reviewed-on: https://review.gerrithub.io/c/spdk/spdk/+/468663
Tested-by: SPDK CI Jenkins <sys_sgci@intel.com>
Reviewed-by: Ben Walker <benjamin.walker@intel.com>
Reviewed-by: Jim Harris <james.r.harris@intel.com>
Reviewed-by: Changpeng Liu <changpeng.liu@intel.com>
It is a very rare thing for a buffer to be split over two memory
regions. In fact, it is only possible in dpdk versions where
--match-allocations is not passed as a startup parameter to dpdk but
dynamic memory allocation is enabled.
By adding a small helper function, we avoid failing an I/O because it
was assigned one of these improperly aligned buffers. Also, we try to
remove the buffer from circulation so that it doesn't get picked up
again by another request.
Also, add a unit test to catch this case.
Change-Id: Ia09865c2f77160a960571665b29c4533b11758ae
Signed-off-by: Seth Howell <seth.howell@intel.com>
Reviewed-on: https://review.gerrithub.io/c/spdk/spdk/+/467446
Tested-by: SPDK CI Jenkins <sys_sgci@intel.com>
Reviewed-by: Ben Walker <benjamin.walker@intel.com>
Reviewed-by: Shuhei Matsumoto <shuhei.matsumoto.xt@hitachi.com>
Reviewed-by: Paul Luse <paul.e.luse@intel.com>
Just cleaning up a few things like variable names and ordering to make
the whole function more readable.
Change-Id: I1503cdb43ddd73e063d6e57e9ff0cf2a06e79728
Signed-off-by: Seth Howell <seth.howell@intel.com>
Reviewed-on: https://review.gerrithub.io/c/spdk/spdk/+/467445
Tested-by: SPDK CI Jenkins <sys_sgci@intel.com>
Reviewed-by: Alexey Marchuk <alexeymar@mellanox.com>
Reviewed-by: Ben Walker <benjamin.walker@intel.com>
Reviewed-by: Shuhei Matsumoto <shuhei.matsumoto.xt@hitachi.com>
Reviewed-by: Paul Luse <paul.e.luse@intel.com>
IB Architecture Specification vol.1 rel.13. in ch.10.3.1 "QUEUE PAIR
AND EE CONTEXT STATES" suggests the following destroy procedure for
QPs associated with SRQ:
- Put the QP in the Error State;
- wait for the Affiliated Asynchronous Last WQE Reached Event;
- either:
* drain the CQ by invoking the Poll CQ verb and either wait for CQ
to be empty or the number of Poll CQ operations has exceeded CQ
capacity size; or
* post another WR that completes on the same CQ and wait for this WR
to return as a WC;
- and then invoke a Destroy QP or Reset QP.
Without the drain step it is possible that LAST_WQE_REACHED event is
received and QP is destroyed before the last receive WR completion is
polled from the CQ.
In SPDK there is no risk of resource leakage in this case. So, instead
of draining we can destroy QP and then just ignore receive completions
without QP and post receive WRs back to SRQ.
Fixes#903
Signed-off-by: Evgeniy Kochetov <evgeniik@mellanox.com>
Signed-off-by: Sasha Kotchubievsky <sashakot@mellanox.com>
Signed-off-by: Alexey Marchuk <alexeymar@mellanox.com>
Change-Id: Ice6d3d5afc205c489f768e3b51c6cda8809bee9a
Reviewed-on: https://review.gerrithub.io/c/spdk/spdk/+/465747
Reviewed-by: Seth Howell <seth.howell@intel.com>
Reviewed-by: Ben Walker <benjamin.walker@intel.com>
Reviewed-by: Shuhei Matsumoto <shuhei.matsumoto.xt@hitachi.com>
Tested-by: SPDK CI Jenkins <sys_sgci@intel.com>
This unifies buffer management among transports further and is a
preparation to make buffer allocation asynchronous.
Signed-off-by: Shuhei Matsumoto <shuhei.matsumoto.xt@hitachi.com>
Change-Id: I8c588eeac4081f50fe32605feb7352f72c628d95
Reviewed-on: https://review.gerrithub.io/c/spdk/spdk/+/466847
Reviewed-by: Broadcom SPDK FC-NVMe CI <spdk-ci.pdl@broadcom.com>
Reviewed-by: Ben Walker <benjamin.walker@intel.com>
Reviewed-by: Anil Veerabhadrappa <anil.veerabhadrappa@broadcom.com>
Reviewed-by: Jim Harris <james.r.harris@intel.com>
Tested-by: SPDK CI Jenkins <sys_sgci@intel.com>
Most variables related with I/O buffer are in struct spdk_nvmf_request
now. So we can pass nvmf_request instead of nvmf_rdma_request to
nvmf_rdma_request_fill_buffers and do it in this patch.
Additionally, we use the cached pointer to nvmf_request in
spdk_nvmf_rdma_request_fill_iovs which is the caller to
nvmf_rdma_request_fill_buffers in this patch.
Signed-off-by: Shuhei Matsumoto <shuhei.matsumoto.xt@hitachi.com>
Change-Id: Ia7664e9688bd9fa157504b4f5075f79759d0e489
Reviewed-on: https://review.gerrithub.io/c/spdk/spdk/+/466212
Tested-by: SPDK CI Jenkins <sys_sgci@intel.com>
Reviewed-by: Ben Walker <benjamin.walker@intel.com>
Reviewed-by: Changpeng Liu <changpeng.liu@intel.com>
Reviewed-by: Broadcom SPDK FC-NVMe CI <spdk-ci.pdl@broadcom.com>
Use spdk_nvmf_request_get_buffers() and spdk_nvmf_request_free_buffers(),
and then remove spdk_nvmf_rdma_request_free_buffers() and
nvmf_rdma_request_get_buffers().
Set rdma_req->data_from_pool to false after
spdk_nvmf_request_free_buffers().
Signed-off-by: Shuhei Matsumoto <shuhei.matsumoto.xt@hitachi.com>
Change-Id: Ie1fc4c261c3197c8299761655bf3138eebcea3bc
Reviewed-on: https://review.gerrithub.io/c/spdk/spdk/+/465875
Tested-by: SPDK CI Jenkins <sys_sgci@intel.com>
Reviewed-by: Changpeng Liu <changpeng.liu@intel.com>
Reviewed-by: Broadcom SPDK FC-NVMe CI <spdk-ci.pdl@broadcom.com>
Reviewed-by: Ben Walker <benjamin.walker@intel.com>
This is a prepration to unify buffer management among transports.
Signed-off-by: Shuhei Matsumoto <shuhei.matsumoto.xt@hitachi.com>
Change-Id: I6b1c208207ae3679619239db4e6e9a77b33291d0
Reviewed-on: https://review.gerrithub.io/c/spdk/spdk/+/466002
Tested-by: SPDK CI Jenkins <sys_sgci@intel.com>
Reviewed-by: Changpeng Liu <changpeng.liu@intel.com>
Reviewed-by: Ben Walker <benjamin.walker@intel.com>
This is a preparation to unify buffer management among transports.
struct spdk_nvmf_request already has SPDK_NVMF_MAX_SGL_ENTRIES (16) * 2
iovecs. Hence incresing the number of buffers twice will be no problem.
Signed-off-by: Shuhei Matsumoto <shuhei.matsumoto.xt@hitachi.com>
Change-Id: Idb525abbf35dc9f4b8547b785b5dfa77d106d8c9
Reviewed-on: https://review.gerrithub.io/c/spdk/spdk/+/465873
Tested-by: SPDK CI Jenkins <sys_sgci@intel.com>
Reviewed-by: Changpeng Liu <changpeng.liu@intel.com>
Reviewed-by: Ben Walker <benjamin.walker@intel.com>
Reviewed-by: Broadcom SPDK FC-NVMe CI <spdk-ci.pdl@broadcom.com>
One of stop conditions in data WR release function was wrong. This
can cause release of uncompleted data WRs. Release of WRs that are
not yet completed leads to different side-effects, up to data
corruption.
The issue was introduced with send WR batching feature in commit
9d63933b7f.
This patch fixes stop condition and contains some refactoring to
simplify WR release function.
Signed-off-by: Evgeniy Kochetov <evgeniik@mellanox.com>
Signed-off-by: Sasha Kotchubievsky <sashakot@mellanox.com>
Signed-off-by: Alexey Marchuk <alexeymar@mellanox.com>
Change-Id: Ie79f64da345e38038f16a0210bef240f63af325b
Reviewed-on: https://review.gerrithub.io/c/spdk/spdk/+/466029
Reviewed-by: Shuhei Matsumoto <shuhei.matsumoto.xt@hitachi.com>
Reviewed-by: Broadcom SPDK FC-NVMe CI <spdk-ci.pdl@broadcom.com>
Reviewed-by: Seth Howell <seth.howell@intel.com>
Reviewed-by: Ben Walker <benjamin.walker@intel.com>
Tested-by: SPDK CI Jenkins <sys_sgci@intel.com>
From rdma_cma.h "Users must destroy any QP associated with an
rdma_cm_id before destroying the ID."
Signed-off-by: Jacek Kalwas <jacek.kalwas@intel.com>
Change-Id: I5ed0c25221c5401cdde8b31a4e217b9d79e7caaa
Reviewed-on: https://review.gerrithub.io/c/spdk/spdk/+/464290
Reviewed-by: Broadcom SPDK FC-NVMe CI <spdk-ci.pdl@broadcom.com>
Reviewed-by: Seth Howell <seth.howell@intel.com>
Reviewed-by: Changpeng Liu <changpeng.liu@intel.com>
Reviewed-by: Tomasz Zawadzki <tomasz.zawadzki@intel.com>
Reviewed-by: Ben Walker <benjamin.walker@intel.com>
Tested-by: SPDK CI Jenkins <sys_sgci@intel.com>
We should be checking directly against the base of the iov when doing
memory map translations. The current behavior is to check against the
starting address of the buffer which is a close address, but not exactly
the same.
Change-Id: I7f65224a6836a814708438f2866d84ae22882216
Signed-off-by: Seth Howell <seth.howell@intel.com>
Reviewed-on: https://review.gerrithub.io/c/spdk/spdk/+/463893
Reviewed-by: Broadcom SPDK FC-NVMe CI <spdk-ci.pdl@broadcom.com>
Reviewed-by: <jiandong.zheng@broadcom.com>
Reviewed-by: Changpeng Liu <changpeng.liu@intel.com>
Reviewed-by: Darek Stojaczyk <dariusz.stojaczyk@intel.com>
Tested-by: SPDK CI Jenkins <sys_sgci@intel.com>
In case of failure during resource allocation within poll_group_create
there is a lack of return statement which could lead to NULL ptr
dereference.
Signed-off-by: Jacek Kalwas <jacek.kalwas@intel.com>
Change-Id: I84abe64a1843117d76b97e62656bdfc4fe2b35d8
Reviewed-on: https://review.gerrithub.io/c/spdk/spdk/+/463195
Reviewed-by: Tomasz Zawadzki <tomasz.zawadzki@intel.com>
Reviewed-by: Ben Walker <benjamin.walker@intel.com>
Reviewed-by: Shuhei Matsumoto <shuhei.matsumoto.xt@hitachi.com>
Reviewed-by: Seth Howell <seth.howell@intel.com>
Reviewed-by: Changpeng Liu <changpeng.liu@intel.com>
Tested-by: SPDK CI Jenkins <sys_sgci@intel.com>
NVMf statistics functions use spdk_get_io_channel function to get a
poll group. It increases reference counter in io channel and causes
problems on application exit. spdk_put_io_channel calls were added to
release the channel.
Signed-off-by: Evgeniy Kochetov <evgeniik@mellanox.com>
Change-Id: I832d1eae346c3bc3858ed0ed063ff7a7a897a2f5
Reviewed-on: https://review.gerrithub.io/c/spdk/spdk/+/463389
Reviewed-by: Jim Harris <james.r.harris@intel.com>
Reviewed-by: Ben Walker <benjamin.walker@intel.com>
Tested-by: SPDK CI Jenkins <sys_sgci@intel.com>
This patch adds measurement of time request spends from the moment it
was polled till completion.
Signed-off-by: Evgeniy Kochetov <evgeniik@mellanox.com>
Change-Id: I1fcda68735f2210c5365dd06f26c10162e4ddf33
Reviewed-on: https://review.gerrithub.io/c/spdk/spdk/+/445291
Tested-by: SPDK CI Jenkins <sys_sgci@intel.com>
Reviewed-by: Jim Harris <james.r.harris@intel.com>
Reviewed-by: Ben Walker <benjamin.walker@intel.com>
This patch adds statistics for pending state in NVMf RDMA subsytem
which may help to detect lack of resources and adjust configuration
correctly.
Signed-off-by: Evgeniy Kochetov <evgeniik@mellanox.com>
Change-Id: I9560d931c0dfb469659be42e13b8302c52912420
Reviewed-on: https://review.gerrithub.io/c/spdk/spdk/+/452300
Tested-by: SPDK CI Jenkins <sys_sgci@intel.com>
Reviewed-by: Paul Luse <paul.e.luse@intel.com>
Reviewed-by: Jim Harris <james.r.harris@intel.com>
Reviewed-by: Ben Walker <benjamin.walker@intel.com>
RDMA polling statistics: number of polls and number of completion
entries returned.
Signed-off-by: Evgeniy Kochetov <evgeniik@mellanox.com>
Change-Id: Iabcf2cb6f6a35f595b89b58cdfcd177a637dda13
Reviewed-on: https://review.gerrithub.io/c/spdk/spdk/+/445289
Tested-by: SPDK CI Jenkins <sys_sgci@intel.com>
Reviewed-by: Jim Harris <james.r.harris@intel.com>
Reviewed-by: Ben Walker <benjamin.walker@intel.com>
This patch adds transport part to nvmf_get_stats RPC method and basic
infrastructure to report NVMf transport specific statistics.
Signed-off-by: Evgeniy Kochetov <evgeniik@mellanox.com>
Change-Id: Ie83b34f4ed932dd5f6d6e37897cf45228114bd88
Reviewed-on: https://review.gerrithub.io/c/spdk/spdk/+/452299
Tested-by: SPDK CI Jenkins <sys_sgci@intel.com>
Reviewed-by: Ben Walker <benjamin.walker@intel.com>
Reviewed-by: Jim Harris <james.r.harris@intel.com>
qpair structure is freed and an error code is returned to the caller in the case of failed qpair initialization in function spdk_nvmf_rdma_qpair_initialize (e.g. bad return value of rdma_create_qp).
The return code is handled by nvmf_tgt_poll_group_add function which destroys the qpair for the second time.
This patch fixes#857
Change-Id: I0773652ecccbbd634ad272106e0a93c1e591d7d2
Signed-off-by: Alexey Marchuk <alexeymar@mellanox.com>
Signed-off-by: Sasha Kotchubievsky <sashakot@mellanox.com>
Signed-off-by: Evgenii Kochetov <evgeniik@mellanox.com>
Reviewed-on: https://review.gerrithub.io/c/spdk/spdk/+/462011
Reviewed-by: Darek Stojaczyk <dariusz.stojaczyk@intel.com>
Reviewed-by: Shuhei Matsumoto <shuhei.matsumoto.xt@hitachi.com>
Reviewed-by: Lorne Li <lorneli@163.com>
Reviewed-by: Ben Walker <benjamin.walker@intel.com>
Tested-by: SPDK CI Jenkins <sys_sgci@intel.com>
Func spdk_nvmf_rdma_destroy_defunct_qpair is a "last chance option"
to destroy qp manually if some driver/hardware doesn't drain qp's
failed wr as expected.
There's a probability that ibv_poll_cq polls wr of the destoryed qp
after spdk_nvmf_rdma_destroy_defunct_qpair's execution. Although in
practice the risk of this situation is minimal(if not non-existent),
add a log here so that we could detect this situation easily.
Change-Id: Ifa9534397513bcea34c18fbb8168eef8f53599c1
Signed-off-by: lorneli <lorneli@163.com>
Reviewed-on: https://review.gerrithub.io/c/spdk/spdk/+/462441
Reviewed-by: Seth Howell <seth.howell@intel.com>
Reviewed-by: Changpeng Liu <changpeng.liu@intel.com>
Reviewed-by: Darek Stojaczyk <dariusz.stojaczyk@intel.com>
Reviewed-by: Ben Walker <benjamin.walker@intel.com>
Tested-by: SPDK CI Jenkins <sys_sgci@intel.com>
Currently rqpair will be destroyed directly in ibv_poll_cq path
if it has been drained, regardless of whether there are outstanding
I/Os issued to bdev layer. So after outstanding I/Os completing,
spdk_nvmf_rdma_close_qpair will be called from nvmf layer, accessing
a destroyed qp.
This path defers qp destruction in nvmf_rdma_destroy_drained_qpair
func until nvmf layer closes qp.
Fixes 851
Change-Id: I8bcce66f8053ddb105702ac603d5d73af54bdcfc
Signed-off-by: lorneli <lorneli@163.com>
Reviewed-on: https://review.gerrithub.io/c/spdk/spdk/+/461237
Reviewed-by: Changpeng Liu <changpeng.liu@intel.com>
Reviewed-by: Seth Howell <seth.howell@intel.com>
Reviewed-by: Darek Stojaczyk <dariusz.stojaczyk@intel.com>
Reviewed-by: Ben Walker <benjamin.walker@intel.com>
Tested-by: SPDK CI Jenkins <sys_sgci@intel.com>
spdk_dma_*malloc() is about to be deprecated.
Change-Id: I5bcac50baca785255eb068086e67c07d120b042f
Signed-off-by: Darek Stojaczyk <dariusz.stojaczyk@intel.com>
Reviewed-on: https://review.gerrithub.io/c/spdk/spdk/+/459432
Tested-by: SPDK CI Jenkins <sys_sgci@intel.com>
Reviewed-by: Changpeng Liu <changpeng.liu@intel.com>
Reviewed-by: Tomasz Zawadzki <tomasz.zawadzki@intel.com>
Reviewed-by: Shuhei Matsumoto <shuhei.matsumoto.xt@hitachi.com>
Reviewed-by: Ben Walker <benjamin.walker@intel.com>
No need to have it under lock. Additionally in case of failure
there was a lack of rdma_destroy_id(). This is addresed within this
change as well.
Signed-off-by: Jacek Kalwas <jacek.kalwas@intel.com>
Change-Id: Idbb36d51ad4ef7ef81051463f56efc87ef00c966
Reviewed-on: https://review.gerrithub.io/c/spdk/spdk/+/462054
Reviewed-by: Darek Stojaczyk <dariusz.stojaczyk@intel.com>
Reviewed-by: Changpeng Liu <changpeng.liu@intel.com>
Tested-by: SPDK CI Jenkins <sys_sgci@intel.com>
In case of failure during pd or map allocation freeing list of devices
was missing.
Signed-off-by: Jacek Kalwas <jacek.kalwas@intel.com>
Change-Id: If62f7b072f3894fd1a7e856c19b4ea51646dd20e
Reviewed-on: https://review.gerrithub.io/c/spdk/spdk/+/462079
Reviewed-by: Darek Stojaczyk <dariusz.stojaczyk@intel.com>
Reviewed-by: Changpeng Liu <changpeng.liu@intel.com>
Tested-by: SPDK CI Jenkins <sys_sgci@intel.com>
In case of pd allocation by nvmf hooks there is a lack of null
check as oposed to pd allocation by ibv_alloc_pd.
Signed-off-by: Jacek Kalwas <jacek.kalwas@intel.com>
Change-Id: Iead6e0332bdee3da4adb6e657af298215c4e2196
Reviewed-on: https://review.gerrithub.io/c/spdk/spdk/+/461576
Tested-by: SPDK CI Jenkins <sys_sgci@intel.com>
Reviewed-by: Darek Stojaczyk <dariusz.stojaczyk@intel.com>
Reviewed-by: Shuhei Matsumoto <shuhei.matsumoto.xt@hitachi.com>
Reviewed-by: Changpeng Liu <changpeng.liu@intel.com>
Vhost testing crashed from Nightly testing, because a member
access within null pointer of type 'struct ibv_send_wr'.
Change-Id: If8f34f23864883ea73516d2d1fe3b30137c04316
Signed-off-by: Hailiang Wang <hailiangx.e.wang@intel.com>
Reviewed-on: https://review.gerrithub.io/c/spdk/spdk/+/458913
Tested-by: SPDK CI Jenkins <sys_sgci@intel.com>
Reviewed-by: Shuhei Matsumoto <shuhei.matsumoto.xt@hitachi.com>
Reviewed-by: Ben Walker <benjamin.walker@intel.com>
Unsetting this flag will decrease the number of WRs retrieved during CQ polling and will decrease
the oeverall processing time. Since RDMA_WRITE operations are always paired with RDMA_SEND (response),
it is possible to track the number of outstanding WRs relying on the completed response WR.
Completed WRs of type RDMA_WR_TYPE_DATA are now always RDMA_READ operations.
The patch shows %2 better peformance for read operations on x86 machine. The performance was measured using perf with the following parameters:
-q 16 -o 4096 -w read -t 300 -c 2
with nvme null device, each measurement was done 4 times
avg IOPS (with patch): 865861.71
avg IOPS (master): 847958.77
avg latency (with patch): 18.46 [us]
avg latency (master): 18.85 [us]
Change-Id: Ifd3329fbd0e45dd5f27213b36b9444308660fc8b
Signed-off-by: Alexey Marchuk <alexeymar@mellanox.com>
Signed-off-by: Sasha Kotchubievsky <sashakot@mellanox.com>
Signed-off-by: Evgenii Kochetov <evgeniik@mellanox.com>
Reviewed-on: https://review.gerrithub.io/c/spdk/spdk/+/456469
Reviewed-by: Ben Walker <benjamin.walker@intel.com>
Reviewed-by: Seth Howell <seth.howell5141@gmail.com>
Reviewed-by: Darek Stojaczyk <dariusz.stojaczyk@intel.com>
Tested-by: SPDK CI Jenkins <sys_sgci@intel.com>
This matches the Linux kernel target. Users can
still decrease this default when creating the
transport (i.e. -p option for nvmf_create_transport
in rpc.py).
Signed-off-by: Jim Harris <james.r.harris@intel.com>
Change-Id: Icad59350a2cd35cfc4ad76d06399345191680c05
Reviewed-on: https://review.gerrithub.io/c/spdk/spdk/+/454820
Tested-by: SPDK CI Jenkins <sys_sgci@intel.com>
Reviewed-by: Darek Stojaczyk <dariusz.stojaczyk@intel.com>
Reviewed-by: Seth Howell <seth.howell5141@gmail.com>
Reviewed-by: Changpeng Liu <changpeng.liu@intel.com>
Reviewed-by: Shuhei Matsumoto <shuhei.matsumoto.xt@hitachi.com>
We could run into issues with this if we were using an arbitrarily large
amount of cores to run SPDK.
Change-Id: Ia7add027d7e6ef1ccb4a69ac328dbdf4f2751fd8
Signed-off-by: Seth Howell <seth.howell@intel.com>
Reviewed-on: https://review.gerrithub.io/c/spdk/spdk/+/452250
Tested-by: SPDK CI Jenkins <sys_sgci@intel.com>
Reviewed-by: Ben Walker <benjamin.walker@intel.com>
Reviewed-by: Shuhei Matsumoto <shuhei.matsumoto.xt@hitachi.com>
This will help us not iterate through the whole list of connections when
only some of them have pending recvs.
Change-Id: I681bc98befbdda4e77ef333b7a086c08b2708eb3
Signed-off-by: Seth Howell <seth.howell@intel.com>
Reviewed-on: https://review.gerrithub.io/c/spdk/spdk/+/449266
Tested-by: SPDK CI Jenkins <sys_sgci@intel.com>
Reviewed-by: Ben Walker <benjamin.walker@intel.com>
Reviewed-by: Jim Harris <james.r.harris@intel.com>
Reviewed-by: Shuhei Matsumoto <shuhei.matsumoto.xt@hitachi.com>
This will help save MMIO overhead. Especially in the SRQ case.
Change-Id: I6fb70cf6de4763450f97961f41ccdce3acec2e63
Signed-off-by: Seth Howell <seth.howell@intel.com>
Reviewed-on: https://review.gerrithub.io/c/spdk/spdk/+/449265
Tested-by: SPDK CI Jenkins <sys_sgci@intel.com>
Reviewed-by: Ben Walker <benjamin.walker@intel.com>
Reviewed-by: Jim Harris <james.r.harris@intel.com>
By creating a list of qpairs, we can avoid looping over every connected
qpair to process sends each time we poll.
Change-Id: If24bbc363176f52fbfb756d56719edd885a21a11
Signed-off-by: Seth Howell <seth.howell@intel.com>
Reviewed-on: https://review.gerrithub.io/c/spdk/spdk/+/449264
Tested-by: SPDK CI Jenkins <sys_sgci@intel.com>
Reviewed-by: Shuhei Matsumoto <shuhei.matsumoto.xt@hitachi.com>
Reviewed-by: Jim Harris <james.r.harris@intel.com>
By batching ibv sends each time we poll, we can reduce the number of
MMIO writes that we do.
Change-Id: Ia5a07b0037365abfa8732629c34d34a9ed49ac70
Signed-off-by: Seth Howell <seth.howell@intel.com>
Reviewed-on: https://review.gerrithub.io/c/spdk/spdk/+/449253
Tested-by: SPDK CI Jenkins <sys_sgci@intel.com>
Reviewed-by: Shuhei Matsumoto <shuhei.matsumoto.xt@hitachi.com>
Reviewed-by: Jim Harris <james.r.harris@intel.com>
There are cases where srq can be a detriment. Add a flag to allow users
to disable srq even if they have a piece of hardware that supports it.
Change-Id: Ia3be8e8c8e8463964e6ff1c02b07afbf4c3cc8f7
Signed-off-by: Seth Howell <seth.howell@intel.com>
Reviewed-on: https://review.gerrithub.io/c/spdk/spdk/+/452271
Reviewed-by: Jim Harris <james.r.harris@intel.com>
Reviewed-by: Shuhei Matsumoto <shuhei.matsumoto.xt@hitachi.com>
Reviewed-by: Ben Walker <benjamin.walker@intel.com>
Tested-by: SPDK CI Jenkins <sys_sgci@intel.com>
This restriction helps reduce the amount of padding when
printing out the event trace, allowing it to fit in a
small number of columns.
Signed-off-by: Jim Harris <james.r.harris@intel.com>
Change-Id: Ifa31e5a6967c7b9bc7028069effb71533f80596f
Reviewed-on: https://review.gerrithub.io/c/spdk/spdk/+/452736
Reviewed-by: Ben Walker <benjamin.walker@intel.com>
Reviewed-by: Shuhei Matsumoto <shuhei.matsumoto.xt@hitachi.com>
Reviewed-by: Darek Stojaczyk <dariusz.stojaczyk@intel.com>
Reviewed-by: Paul Luse <paul.e.luse@intel.com>
Tested-by: SPDK CI Jenkins <sys_sgci@intel.com>
This was not used by any of the trace register descriptions.
Let's remove it rather keeping it around if we don't need it.
Signed-off-by: Jim Harris <james.r.harris@intel.com>
Change-Id: Idda809e2911db5be555ff6aa13695484a14bf665
Reviewed-on: https://review.gerrithub.io/c/spdk/spdk/+/452734
Tested-by: SPDK CI Jenkins <sys_sgci@intel.com>
Reviewed-by: Ben Walker <benjamin.walker@intel.com>
Reviewed-by: Shuhei Matsumoto <shuhei.matsumoto.xt@hitachi.com>
Reviewed-by: Darek Stojaczyk <dariusz.stojaczyk@intel.com>
Reviewed-by: Paul Luse <paul.e.luse@intel.com>
We can use the rpoller->srq to check if a qpair is valid when processing
recv completions.
Change-Id: I6aa360adc48a3312ddcf79f10e2a65b502a7314f
Signed-off-by: Seth Howell <seth.howell@intel.com>
Reviewed-on: https://review.gerrithub.io/c/spdk/spdk/+/452247
Tested-by: SPDK CI Jenkins <sys_sgci@intel.com>
Reviewed-by: Ben Walker <benjamin.walker@intel.com>
Reviewed-by: Shuhei Matsumoto <shuhei.matsumoto.xt@hitachi.com>
Reviewed-by: Jim Harris <james.r.harris@intel.com>
Mempools are based off of a ring structure which allocates its elements
as a power of two. It also only exposes n-1 elements to the user. So
when we create a mempool with 2^n elements in it, we have to allocate a
ring with 2^n+1 entries. By decreasing the number of elements in these
key mempools by 1, we can save a decent amount of memory.
Change-Id: I942c9dd4cf59096969bc2559fb46fd2084a07f09
Signed-off-by: Seth Howell <seth.howell@intel.com>
Reviewed-on: https://review.gerrithub.io/c/spdk/spdk/+/448875
Tested-by: SPDK CI Jenkins <sys_sgci@intel.com>
Reviewed-by: Darek Stojaczyk <dariusz.stojaczyk@intel.com>
Reviewed-by: Ben Walker <benjamin.walker@intel.com>
Reviewed-by: Shuhei Matsumoto <shuhei.matsumoto.xt@hitachi.com>
Reviewed-by: Jim Harris <james.r.harris@intel.com>
It used to be that we would get async events very infrequently. However,
with the introduction of SRQ, this number has gone up tremendously.
Change the way we report our these events so that we don't spam/confuse
people running the target.
Change-Id: I33070281fa854cbc17784d61bbbb870196ca8780
Signed-off-by: Seth Howell <seth.howell@intel.com>
Reviewed-on: https://review.gerrithub.io/c/spdk/spdk/+/452159
Tested-by: SPDK CI Jenkins <sys_sgci@intel.com>
Reviewed-by: Ben Walker <benjamin.walker@intel.com>
Reviewed-by: Jim Harris <james.r.harris@intel.com>
If there are outstanding recvs for a qpair when it is destroyed, we need
to clear the qpair from it before reposting it. Otehrwise, we have a
potential heap-use-after-free of double free (depending on whether the
recv completion is in error state or not).
See github issues #730
Change-Id: Ic2009c761cbcc5e89174f62fbd0872d0489c67ca
Signed-off-by: Seth Howell <seth.howell@intel.com>
Reviewed-on: https://review.gerrithub.io/c/spdk/spdk/+/452122
Tested-by: SPDK CI Jenkins <sys_sgci@intel.com>
Reviewed-by: Jim Harris <james.r.harris@intel.com>
Reviewed-by: Changpeng Liu <changpeng.liu@intel.com>
We were setting this value in the target from our initiator, but it
turns out the rdma_conn_params struct is responsible for setting the
opposite side so we need to add it in the target side when accepting
connections.
Also, add a test to demonstrate target functionality when we overwhelm
the SRQ. It is useful to note that performance really tanks when you
start overwhelming the srq so it may be useful to use this test case to
check performance gains in edge cases over time.
Change-Id: Iac541bd9fc1d82eca9f21e7abc3f625663a6c460
Signed-off-by: Seth Howell <seth.howell@intel.com>
Reviewed-on: https://review.gerrithub.io/c/spdk/spdk/+/451678
Tested-by: SPDK CI Jenkins <sys_sgci@intel.com>
Reviewed-by: Ben Walker <benjamin.walker@intel.com>
Reviewed-by: Jim Harris <james.r.harris@intel.com>
We were incrementing over the end of the descriptor list and assigning
undefined values to the rsp opcode in SEND_WITH_INVAL case. We were only
hitting this error when mixing sgl and inline requests in the same
workload. We were just by chance hitting a four bit value that was set
to all 1s from the in capsule data from the last request.
Change-Id: Ied06356f3d22fa34a2cd869dfad6bdca8720791d
Signed-off-by: Seth Howell <seth.howell@intel.com>
Reviewed-on: https://review.gerrithub.io/c/spdk/spdk/+/450873
Tested-by: SPDK CI Jenkins <sys_sgci@intel.com>
Reviewed-by: Shuhei Matsumoto <shuhei.matsumoto.xt@hitachi.com>
Reviewed-by: Jim Harris <james.r.harris@intel.com>
