Original implementation creates pollers and CQs for all discovered
devices at poll group creation. Device (ibv_context) that has no
references, i.e. has no QPs, may be removed from the system and
ibv_context may be closed by rdma_cm. In this case we will have a CQ
that refers to closed ibv_context and it may crash in ibv_poll_cq.
With this patch pollers are created on demand when we create the first
QP for a device. When there are no more QPs on the poller, we destroy
the poller. This also helps to avoid polling CQs that don't have any
QPs attached.
Signed-off-by: Evgeniy Kochetov <evgeniik@nvidia.com>
Change-Id: I46dd2c8b9b2902168dba24e139c904f51bd1b101
Reviewed-on: https://review.spdk.io/gerrit/c/spdk/spdk/+/13692
Tested-by: SPDK CI Jenkins <sys_sgci@intel.com>
Community-CI: Broadcom CI <spdk-ci.pdl@broadcom.com>
Community-CI: Mellanox Build Bot
Reviewed-by: Shuhei Matsumoto <smatsumoto@nvidia.com>
Reviewed-by: Ben Walker <benjamin.walker@intel.com>
Both PCIE and VFIO-USER can use the same APIs to get IO queue
pair statistic data, so merge them here.
Change-Id: Iadf9ead2bd5abaf11d2ef5d1884acb67369f85bb
Signed-off-by: Changpeng Liu <changpeng.liu@intel.com>
Reviewed-on: https://review.spdk.io/gerrit/c/spdk/spdk/+/13538
Community-CI: Mellanox Build Bot
Tested-by: SPDK CI Jenkins <sys_sgci@intel.com>
Reviewed-by: Jim Harris <james.r.harris@intel.com>
Reviewed-by: Konrad Sztyber <konrad.sztyber@intel.com>
Reviewed-by: Aleksey Marchuk <alexeymar@nvidia.com>
When a secondary process exit without deleting allocated IO
queue pair, then a new secondary process will do cleanup for
previous allocated queue pair, then segment fault will happen
due to `stat` inside IO queue pair data strucutre can't be
accessed in this cleanup process.
Fix issue #2565.
Change-Id: I01a037642683901941b5268ac20d17b78b6c6350
Signed-off-by: Changpeng Liu <changpeng.liu@intel.com>
Reviewed-on: https://review.spdk.io/gerrit/c/spdk/spdk/+/13537
Community-CI: Mellanox Build Bot
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>
Reviewed-by: Aleksey Marchuk <alexeymar@nvidia.com>
That is possible to get/set registers from any thread,
during regs processing we are polling admin qpair to
get a completion. At the same time, another thread
can also poll admin qpair and that can lead to
undefined behavior.
This patch fixes an issue when bdev_nvme is configured
with io_timeout. If remote target becomes unresponsive
(e.g. due to link down), IO timeout occurs and bdev_nvme
tries to get csts registers in timeout_cb. At the same
time another thread can process adminq, so we may have
2 simultaneous adminq polls. If admin qpair is disconnecting
at that time (RDMA transport) we may destroy resources
twice from different threads.
We don't see a problem with set_regs function but it
won't be redundant to lock mutex in set_regs as well.
Signed-off-by: Alexey Marchuk <alexeymar@nvidia.com>
Change-Id: I7ec3984d25d0249061005533d13b22315b44ddf2
Reviewed-on: https://review.spdk.io/gerrit/c/spdk/spdk/+/13687
Tested-by: SPDK CI Jenkins <sys_sgci@intel.com>
Reviewed-by: Konrad Sztyber <konrad.sztyber@intel.com>
Reviewed-by: Jim Harris <james.r.harris@intel.com>
In the multi-process case, a process may call `spdk_nvme_ctrlr_free_io_qpair` on
a foreign I/O qpair (i.e. one that this process did not create) when that qpairs
process exits unexpectedly.
The variable `qpair->poll_group` isn't multi-process safe, we can't use it
in `spdk_nvme_ctrlr_free_io_qpair` and related transport poll group APIs.
Change-Id: Ic13a6a2c7d760477be5be5a56a45caa2b5518717
Signed-off-by: Changpeng Liu <changpeng.liu@intel.com>
Reviewed-on: https://review.spdk.io/gerrit/c/spdk/spdk/+/13573
Community-CI: Broadcom CI <spdk-ci.pdl@broadcom.com>
Community-CI: Mellanox Build Bot
Tested-by: SPDK CI Jenkins <sys_sgci@intel.com>
Reviewed-by: Aleksey Marchuk <alexeymar@nvidia.com>
Reviewed-by: Jim Harris <james.r.harris@intel.com>
Reviewed-by: Shuhei Matsumoto <smatsumoto@nvidia.com>
All of the callers immediately put the req right
after the nvme_rdma_req_complete call, so just move
the put into that function instead.
Signed-off-by: Jim Harris <james.r.harris@intel.com>
Change-Id: Ic370cf689850924e0c902a6071af8b3a7ed58c0b
Reviewed-on: https://review.spdk.io/gerrit/c/spdk/spdk/+/13527
Tested-by: SPDK CI Jenkins <sys_sgci@intel.com>
Reviewed-by: Changpeng Liu <changpeng.liu@intel.com>
Reviewed-by: Shuhei Matsumoto <smatsumoto@nvidia.com>
Community-CI: Broadcom CI <spdk-ci.pdl@broadcom.com>
This follows similar logic in the pcie and tcp
completion paths, including omitting error
messages when aborting aers by adding a print_on_error
parameter to the completion function.
Signed-off-by: Jim Harris <james.r.harris@intel.com>
Change-Id: Id558d0af2cdd705dfb60abb842bd567a0949ccce
Reviewed-on: https://review.spdk.io/gerrit/c/spdk/spdk/+/13525
Tested-by: SPDK CI Jenkins <sys_sgci@intel.com>
Reviewed-by: Changpeng Liu <changpeng.liu@intel.com>
Reviewed-by: Shuhei Matsumoto <smatsumoto@nvidia.com>
Community-CI: Broadcom CI <spdk-ci.pdl@broadcom.com>
By default, the SPDK nvmf target reports vid==INTEL,
which results in the SPDK nvme driver trying to enable
Intel vendor-specific log page. Fix this by trying to
enable those log pages only for PCIE transport
controllers.
Signed-off-by: Jim Harris <james.r.harris@intel.com>
Change-Id: I78ebf365d4fa6295d1f610697266c3ead765988d
Reviewed-on: https://review.spdk.io/gerrit/c/spdk/spdk/+/13524
Tested-by: SPDK CI Jenkins <sys_sgci@intel.com>
Reviewed-by: Changpeng Liu <changpeng.liu@intel.com>
Reviewed-by: Shuhei Matsumoto <smatsumoto@nvidia.com>
Community-CI: Broadcom CI <spdk-ci.pdl@broadcom.com>
We can only get to this code path if the controller
has vid==INTEL, so make that more clear by changing
the check to an assert.
Remove unit test that calls
nvme_ctrlr_construct_intel_support_log_page_list()
for a controller that is not VID==INTEL - this is
no longer valid.
Signed-off-by: Jim Harris <james.r.harris@intel.com>
Change-Id: I3b58451bc95992bf641e7452f0ac4c2bac9fe31c
Reviewed-on: https://review.spdk.io/gerrit/c/spdk/spdk/+/13523
Tested-by: SPDK CI Jenkins <sys_sgci@intel.com>
Reviewed-by: Changpeng Liu <changpeng.liu@intel.com>
Reviewed-by: Shuhei Matsumoto <smatsumoto@nvidia.com>
Community-CI: Broadcom CI <spdk-ci.pdl@broadcom.com>
This follows similar logic in the pcie completion
path, including omitting error messages when aborting
aers by adding a print_on_error parameter to the
completion function.
Signed-off-by: Jim Harris <james.r.harris@intel.com>
Change-Id: I96df72280bb8fcbee3847fdc27f38e14a1bf3251
Reviewed-on: https://review.spdk.io/gerrit/c/spdk/spdk/+/13522
Tested-by: SPDK CI Jenkins <sys_sgci@intel.com>
Reviewed-by: Changpeng Liu <changpeng.liu@intel.com>
Reviewed-by: Shuhei Matsumoto <smatsumoto@nvidia.com>
Community-CI: Broadcom CI <spdk-ci.pdl@broadcom.com>
nvme_tcp_req_complete_safe caches values on
the request, so that we can free the request *before*
completing it. This allows the recently completed
req to get reused in full queue depth workloads, if
the callback function submits a new I/O.
So do this nvme_tcp_req_complete as well, to make
all of the completion paths identical. The paths
that were calling nvme_tcp_req_complete previously
are all non-fast-path, so the extra overhead is
not important.
This allows us to call nvme_tcp_req_complete from
nvme_tcp_req_complete_safe to reduce code duplication,
so do that in this patch as well.
Signed-off-by: Jim Harris <james.r.harris@intel.com>
Change-Id: I876cea5ea20aba8ccc57d179e63546a463a87b35
Reviewed-on: https://review.spdk.io/gerrit/c/spdk/spdk/+/13521
Tested-by: SPDK CI Jenkins <sys_sgci@intel.com>
Reviewed-by: Changpeng Liu <changpeng.liu@intel.com>
Reviewed-by: Shuhei Matsumoto <smatsumoto@nvidia.com>
Community-CI: Broadcom CI <spdk-ci.pdl@broadcom.com>
All callers of nvme_tcp_req_complete call
nvme_tcp_req_put immediately afterwards, so move
this call into nvme_tcp_req_complete.
This will help enable some improvements in later
patches.
Note that nvme_tcp_req_complete_safe has this same
functionality open coded right now, but that will
get changed in the next patch. It calls
nvme_tcp_req_put immediately after the TAILQ_REMOVE,
so do that in nvme_tcp_req_complete as well.
Signed-off-by: Jim Harris <james.r.harris@intel.com>
Change-Id: I368122bc49a7f0772e3011e5427e3c43618380eb
Reviewed-on: https://review.spdk.io/gerrit/c/spdk/spdk/+/13520
Tested-by: SPDK CI Jenkins <sys_sgci@intel.com>
Reviewed-by: Changpeng Liu <changpeng.liu@intel.com>
Reviewed-by: Shuhei Matsumoto <smatsumoto@nvidia.com>
Community-CI: Broadcom CI <spdk-ci.pdl@broadcom.com>
nvme_qpair_abort_all_queued_reqs() aborts error injections, queued
requests, aborting queued requests, and outstanding requests. (Aborting
outstanding requests depends on transports.) However, it did not abort
queued aborts.
Include nvme_ctrlr_abort_queued_aborts() into
nvme_qpair_abort_all_queued_reqs() to do really the name of the
function indicates.
nvme_ctrlr_abort_queued_aborts() has been called in a few cases, but
we do not care duplication.
Signed-off-by: Shuhei Matsumoto <smatsumoto@nvidia.com>
Change-Id: I19102cc6603a72ce5c398a7947cb4d606b692991
Reviewed-on: https://review.spdk.io/gerrit/c/spdk/spdk/+/12849
Tested-by: SPDK CI Jenkins <sys_sgci@intel.com>
Reviewed-by: Konrad Sztyber <konrad.sztyber@intel.com>
Reviewed-by: Jim Harris <james.r.harris@intel.com>
Reviewed-by: Vasuki Manikarnike <vasuki.manikarnike@hpe.com>
Community-CI: Broadcom CI <spdk-ci.pdl@broadcom.com>
Community-CI: Mellanox Build Bot
In SPDK, declarations have the return type on the same line. Definitions
have the return type on a separate line. Astyle has an option for
enforcing this. Unfortunately, it seems to have two bugs:
1) It doesn't work correctly at all on C++ files.
2) It often fails on functions that return enums, or long type names
Deal with 1) by adjusting the check_format.sh script to only tell astyle
to fix return type line breaks for C files and not C++. Deal with 2) by
adding a few typedefs to work around the problem.
Change-Id: Idf28281466cab8411ce252d5f02ab384166790c6
Signed-off-by: Ben Walker <benjamin.walker@intel.com>
Reviewed-on: https://review.spdk.io/gerrit/c/spdk/spdk/+/13437
Tested-by: SPDK CI Jenkins <sys_sgci@intel.com>
Reviewed-by: Jim Harris <james.r.harris@intel.com>
Reviewed-by: Dong Yi <dongx.yi@intel.com>
Reviewed-by: Tomasz Zawadzki <tomasz.zawadzki@intel.com>
Community-CI: Broadcom CI <spdk-ci.pdl@broadcom.com>
We had not incremented ctrlr->outstanding_aborts when aborting a
request in the ctrlr->queued_aborts, and ctrlr->outstanding_aborts
became negative. Fix the bug in this patch. Additionally add assert
to check if ctrlr->outstanding_aborts is not negative.
Signed-off-by: Shuhei Matsumoto <smatsumoto@nvidia.com>
Change-Id: I58090286f070ba854bdea87f0f8ecb7810890338
Reviewed-on: https://review.spdk.io/gerrit/c/spdk/spdk/+/13452
Community-CI: Broadcom CI <spdk-ci.pdl@broadcom.com>
Community-CI: Mellanox Build Bot
Tested-by: SPDK CI Jenkins <sys_sgci@intel.com>
Reviewed-by: Konrad Sztyber <konrad.sztyber@intel.com>
Reviewed-by: Tomasz Zawadzki <tomasz.zawadzki@intel.com>
Fully asynchronous ctrlr detach (b6ecc3729) introduce a register
operation state machine that waits for operation to complete. When
controller failed to initialize, `nvme_ctrlr_fail` set qpair state to
`DISCONNECTED` immediately, causing qpair process completions to
never complete register operations therefore prevent async detach exit.
Signed-off-by: Sebastian Brzezinka <sebastian.brzezinka@intel.com>
Change-Id: I205c5157b8ea7b4535f98ff4052414310e421446
Reviewed-on: https://review.spdk.io/gerrit/c/spdk/spdk/+/12858
Community-CI: Broadcom CI <spdk-ci.pdl@broadcom.com>
Community-CI: Mellanox Build Bot
Tested-by: SPDK CI Jenkins <sys_sgci@intel.com>
Reviewed-by: Konrad Sztyber <konrad.sztyber@intel.com>
Reviewed-by: Jim Harris <james.r.harris@intel.com>
Reviewed-by: Ben Walker <benjamin.walker@intel.com>
When running perf test, sometimes after CONNECT req's resp was
received and processed, the qpair still failed to change from state
CONNECTING to CONNECTED. For when it goes to nvme_fabric_qpair_connect_poll
-> nvme_wait_for_completion_robust_lock_timeout_poll to process the
CONNECT req's resp, the req may have not been finished in sock_check_zcopy,
although its resp has been received and processed, which means the
tcp_req->ordering.bits.send_ack is still 0 and the status->done still
is false. And after the req is completed in sock_check_zcopy, we need
to poll this qpair again to make the state enter CONNECTED.
And if icreq's resp received and processed before nvme_tcp_send_icreq_complete
is called by _sock_check_zcopy, the qpair will be stuck in CONNECTING
and it never proceed to send the CONNECT req. We also need to put it
in pgroup->needs_poll to fix it.
I can reproduce this bug with the following configuration.
target: 16NVMe SSD, running on 20 cores;
initiator: randread test using nvme perf with 32 cpu cores and
zerocopy enabled.
The error doesn't always occur. CONNECT failure is about 1 failure in
ten with the following log. And icreq failure is less frequent with
only target side's "keep alive timeout" log.
Error reported in initiator side:
Initialization complete. Launching workers.
[2022-05-23 14:51:07.286794] nvme_qpair.c: 760:spdk_nvme_qpair_process_completions:
*ERROR*: CQ transport error -6 (No such device or address) on qpair id 2
ERROR: unable to connect I/O qpair.
ERROR: init_ns_worker_ctx() failed
And target side shows:
Disconnecting host from subsystem nqn.2016-06.io.spdk:cnode2 due to keep alive timeout
Change-Id: Id72c2ffd615ab73c5fc67d36c3ff8b730cebcef7
Signed-off-by: Richael Zhuang <richael.zhuang@arm.com>
Reviewed-on: https://review.spdk.io/gerrit/c/spdk/spdk/+/12975
Community-CI: Broadcom CI <spdk-ci.pdl@broadcom.com>
Community-CI: Mellanox Build Bot
Tested-by: SPDK CI Jenkins <sys_sgci@intel.com>
Reviewed-by: Aleksey Marchuk <alexeymar@nvidia.com>
Reviewed-by: Konrad Sztyber <konrad.sztyber@intel.com>
Many open source projects have moved to using SPDX identifiers
to specify license information, reducing the amount of
boilerplate code in every source file. This patch replaces
the bulk of SPDK .c, .cpp and Makefiles with the BSD-3-Clause
identifier.
Almost all of these files share the exact same license text,
and this patch only modifies the files that contain the
most common license text. There can be slight variations
because the third clause contains company names - most say
"Intel Corporation", but there are instances for Nvidia,
Samsung, Eideticom and even "the copyright holder".
Used a bash script to automate replacement of the license text
with SPDX identifier which is checked into scripts/spdx.sh.
Signed-off-by: Jim Harris <james.r.harris@intel.com>
Change-Id: Iaa88ab5e92ea471691dc298cfe41ebfb5d169780
Reviewed-on: https://review.spdk.io/gerrit/c/spdk/spdk/+/12904
Community-CI: Broadcom CI <spdk-ci.pdl@broadcom.com>
Community-CI: Mellanox Build Bot
Tested-by: SPDK CI Jenkins <sys_sgci@intel.com>
Reviewed-by: Aleksey Marchuk <alexeymar@nvidia.com>
Reviewed-by: Changpeng Liu <changpeng.liu@intel.com>
Reviewed-by: Dong Yi <dongx.yi@intel.com>
Reviewed-by: Konrad Sztyber <konrad.sztyber@intel.com>
Reviewed-by: Paul Luse <paul.e.luse@intel.com>
Reviewed-by: <qun.wan@intel.com>
Play it safe and add the same memory barrier in
nvme_pcie_qpair_process_completions() as for ppc64.
Signed-off-by: Heinrich Schuchardt <heinrich.schuchardt@canonical.com>
Change-Id: I7079b4769d30106387ef4549495a72b7fea6a77a
Reviewed-on: https://review.spdk.io/gerrit/c/spdk/spdk/+/12879
Community-CI: Broadcom CI <spdk-ci.pdl@broadcom.com>
Community-CI: Mellanox Build Bot
Tested-by: SPDK CI Jenkins <sys_sgci@intel.com>
Reviewed-by: Jim Harris <james.r.harris@intel.com>
Reviewed-by: Aleksey Marchuk <alexeymar@nvidia.com>
Prepare for the later patch, and make the later patch code clean
Signed-off-by: MengjinWu <mengjin.wu@intel.com>
Change-Id: I12b175c86a5245f38dc76fe2d3918ec4b30a475a
Reviewed-on: https://review.spdk.io/gerrit/c/spdk/spdk/+/12830
Community-CI: Mellanox Build Bot
Community-CI: Broadcom CI <spdk-ci.pdl@broadcom.com>
Tested-by: SPDK CI Jenkins <sys_sgci@intel.com>
Reviewed-by: Shuhei Matsumoto <smatsumoto@nvidia.com>
Reviewed-by: Aleksey Marchuk <alexeymar@nvidia.com>
Reviewed-by: Xiaodong Liu <xiaodong.liu@intel.com>
For a C2HTermReq PDU, there's no associated tcp_req, so we need to check
it for NULL before dereferencing it.
Also, while here, moved some of the assignments to the declarations to
reduce the number of boilerplate lines.
Signed-off-by: Konrad Sztyber <konrad.sztyber@intel.com>
Change-Id: Iac05ef0ba605e2f40d0026ad1b131c28d29f7314
Reviewed-on: https://review.spdk.io/gerrit/c/spdk/spdk/+/12845
Community-CI: Broadcom CI <spdk-ci.pdl@broadcom.com>
Community-CI: Mellanox Build Bot
Tested-by: SPDK CI Jenkins <sys_sgci@intel.com>
Reviewed-by: Jim Harris <james.r.harris@intel.com>
Reviewed-by: Aleksey Marchuk <alexeymar@nvidia.com>
This SGL type was missed in the original commit
that added the pretty printing.
Fixes: 4d9ab1e9a1 ("nvme: pretty print dptr")
Reported-by: Ramanjaneya Burugula <burugula@gmail.com>
Signed-off-by: Jim Harris <james.r.harris@intel.com>
Change-Id: Ibc655db4e65009071f39f55f691c94a094cea0bc
Reviewed-on: https://review.spdk.io/gerrit/c/spdk/spdk/+/12705
Tested-by: SPDK CI Jenkins <sys_sgci@intel.com>
Community-CI: Broadcom CI <spdk-ci.pdl@broadcom.com>
Community-CI: Mellanox Build Bot
Reviewed-by: Changpeng Liu <changpeng.liu@intel.com>
Reviewed-by: Aleksey Marchuk <alexeymar@nvidia.com>
Use the conventional huge-pages based spdk allocation scheme for the initiator
data-structures unconditionally.
Change-Id: I5baee7614e3ac9b5497b3d771dfddfbaa7fdf65b
Signed-off-by: Or Gerlitz <ogerlitz@nvidia.com>
Reviewed-on: https://review.spdk.io/gerrit/c/spdk/spdk/+/12687
Community-CI: Broadcom CI <spdk-ci.pdl@broadcom.com>
Reviewed-by: Aleksey Marchuk <alexeymar@nvidia.com>
Reviewed-by: Dong Yi <dongx.yi@intel.com>
Reviewed-by: Shuhei Matsumoto <smatsumoto@nvidia.com>
Tested-by: SPDK CI Jenkins <sys_sgci@intel.com>
A iterator function nvme_request_add_abort() covers not only a small
I/O request but also children of a large I/O.
However nvme_qpair_abort_queued_reqs_with_cbarg() did not check the
latter. check if cmd_cb_arg matches not only req->cb_arg but also
req->parent_cb_arg.
Signed-off-by: Shuhei Matsumoto <smatsumoto@nvidia.com>
Change-Id: I015e29b0a8f58920b9a13081330a94f9dd976a45
Reviewed-on: https://review.spdk.io/gerrit/c/spdk/spdk/+/12557
Community-CI: Broadcom CI <spdk-ci.pdl@broadcom.com>
Community-CI: Mellanox Build Bot
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>
For PCIe transport, we need to stop any activity of the controller
before deleting I/O qpair resource in a controller reset sequence.
However, we set I/O qpairs to failed before disabling a controller.
In the NVMe bdev module, this caused disconnected qpair callback to
delete I/O qpairs before disabling the controller.
Hence, change the code slightly to set I/O qpairs to failed only if
reset is synchronous to keep backward compatibility.
Signed-off-by: Shuhei Matsumoto <smatsumoto@nvidia.com>
Change-Id: Ica71aad0a1dabce45616dfdfff5f11b07131bbd1
Reviewed-on: https://review.spdk.io/gerrit/c/spdk/spdk/+/12736
Community-CI: Broadcom CI <spdk-ci.pdl@broadcom.com>
Tested-by: SPDK CI Jenkins <sys_sgci@intel.com>
Reviewed-by: Ben Walker <benjamin.walker@intel.com>
Reviewed-by: Michael Haeuptle <michaelhaeuptle@gmail.com>
Reviewed-by: Jim Harris <james.r.harris@intel.com>
Reviewed-by: Aleksey Marchuk <alexeymar@nvidia.com>
The following patches swaps the ordering of destrloying I/O qpairs
and disconnecting a controller for PCIe transport.
prepare_for_reset is a flag for PCIe transport.
Signed-off-by: Shuhei Matsumoto <smatsumoto@nvidia.com>
Change-Id: I3009de9fea089fc93ecf87adba42e85c9a77e715
Reviewed-on: https://review.spdk.io/gerrit/c/spdk/spdk/+/12582
Community-CI: Mellanox Build Bot
Community-CI: Broadcom CI <spdk-ci.pdl@broadcom.com>
Tested-by: SPDK CI Jenkins <sys_sgci@intel.com>
Reviewed-by: Aleksey Marchuk <alexeymar@nvidia.com>
Reviewed-by: Michael Haeuptle <michaelhaeuptle@gmail.com>
Reviewed-by: Ben Walker <benjamin.walker@intel.com>
As described in the previous patches, we need to delete all I/O
SQ/CQs before aborting trackers when disconnecting a controller.
The following patches reorder the operations. This patch changes
adminq disconnection to initiate a Controller Level Reset and
adminq completion processes it if ctrlr->is_disconnecting is true.
Signed-off-by: Shuhei Matsumoto <smatsumoto@nvidia.com>
Change-Id: I64f06bae2ce8a9127124029fd042db0028198e3c
Reviewed-on: https://review.spdk.io/gerrit/c/spdk/spdk/+/12560
Tested-by: SPDK CI Jenkins <sys_sgci@intel.com>
Community-CI: Mellanox Build Bot
Community-CI: Broadcom CI <spdk-ci.pdl@broadcom.com>
Reviewed-by: Michael Haeuptle <michaelhaeuptle@gmail.com>
Reviewed-by: Ben Walker <benjamin.walker@intel.com>
Reviewed-by: Aleksey Marchuk <alexeymar@nvidia.com>
Make this a transport-level decision instead. TCP and RDMA do want to
abort, but PCIe cannot because these commands may still be receiving DMA
operations from the device.
Change-Id: I305acddc3819c903eb3217e8f710d4216d0b3931
Signed-off-by: Ben Walker <benjamin.walker@intel.com>
Reviewed-on: https://review.spdk.io/gerrit/c/spdk/spdk/+/11509
Community-CI: Mellanox Build Bot
Community-CI: Broadcom CI <spdk-ci.pdl@broadcom.com>
Tested-by: SPDK CI Jenkins <sys_sgci@intel.com>
Reviewed-by: Shuhei Matsumoto <smatsumoto@nvidia.com>
Reviewed-by: Konrad Sztyber <konrad.sztyber@intel.com>
Reviewed-by: Aleksey Marchuk <alexeymar@nvidia.com>
Reviewed-by: Michael Haeuptle <michaelhaeuptle@gmail.com>
Previously, we did not do any Controller Level Reset when disconnecting
the admin qpair.
However, for PCIe transport, we need to stop any activity of the
controller, i.e., delete all I/O SQ and CQs before
nvme_transport_ctrlr_disconnect_qpair_done() calls
nvme_transport_qpair_abort_reqs() (i.e., nvme_pcie_qpair_abort_trackers()).
Otherwise, some corruption may occur because completed I/Os may still be
in progress on the NVMe device.
Not to change any public API, nvme_pcie_ctrlr_disconnect_qpair() is a
convenient place to initiate a Controller Level Reset because it is
called from spdk_nvme_ctrlr_disconnect(). Then
nvme_pcie_qpair_process_completions() can process it until completion.
However, necessary functions are not accessible from PCIe transport.
This patch adds two helper functions and guards us from some undesirable
behaviors because it was not assumed that nvme_ctrlr_process_init() is
called from the completion context and ends in the middle of transition.
Signed-off-by: Shuhei Matsumoto <smatsumoto@nvidia.com>
Change-Id: I3d986e94ba71b83beeff7e75cf92033b5fa6f075
Reviewed-on: https://review.spdk.io/gerrit/c/spdk/spdk/+/12559
Community-CI: Mellanox Build Bot
Community-CI: Broadcom CI <spdk-ci.pdl@broadcom.com>
Reviewed-by: Michael Haeuptle <michaelhaeuptle@gmail.com>
Reviewed-by: Ben Walker <benjamin.walker@intel.com>
Reviewed-by: Aleksey Marchuk <alexeymar@nvidia.com>
Tested-by: SPDK CI Jenkins <sys_sgci@intel.com>
It simplifies code and removes cast of nvme_qpair
to rdma_qpair
Signed-off-by: Alexey Marchuk <alexeymar@mellanox.com>
Change-Id: I363246cf9d8c9cbafd48b26facdb5cc37fdd8e67
Reviewed-on: https://review.spdk.io/gerrit/c/spdk/spdk/+/12701
Community-CI: Broadcom CI <spdk-ci.pdl@broadcom.com>
Community-CI: Mellanox Build Bot
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>
When qpair is attached to a poll group, disconnect
process is async - we are waiting for the DISCONNECTED
event from rdmacm to destroy rdma resources. However
the user (nvme_perf) can destroy qpair immediatelly,
so memory allocated for qpair is freed but rdma
resouces are still allocated. That means that we may
receive rdmacm event (DISCONNECTED) for the destroyed qpair,
that leads to use-after-free.
To fix this problem, add a check for internal qpair state
when qpair is destroyed, if disconnect is not finished, then
we forcefully destroy rdma resources.
Fixes issue #2515
Signed-off-by: Alexey Marchuk <alexeymar@mellanox.com>
Reported-by: Or Gerlitz <ogerlitz@nvidia.com>
Change-Id: I7bfa53c9f6fe6ed787323a8941f1f2db17ea0c20
Reviewed-on: https://review.spdk.io/gerrit/c/spdk/spdk/+/12700
Community-CI: Broadcom CI <spdk-ci.pdl@broadcom.com>
Community-CI: Mellanox Build Bot
Tested-by: SPDK CI Jenkins <sys_sgci@intel.com>
Reviewed-by: Shuhei Matsumoto <smatsumoto@nvidia.com>
Reviewed-by: Ben Walker <benjamin.walker@intel.com>
Reviewed-by: Jim Harris <james.r.harris@intel.com>
In the following patches, nvme_ctrlr_process_init() will be used to
disable the controller when disconnecting the admin qpair for PCIe
transport. In this case, we will have to exit nvme_ctrlr_process_init()
after CSTS.RDY is 0. However, spdk_nvme_ctrlr_reset() and
spdk_nvme_ctrlr_reconnect_poll_async() have to continue
nvme_ctrlr_process_init() until the controller becomes ready.
To differentiate stop and continue clearly, add a new state
NVME_CTRLR_STATE_DISABLED to enum nvme_ctrlr_state.
Signed-off-by: Shuhei Matsumoto <smatsumoto@nvidia.com>
Change-Id: Ic0a5fb7114d4eeb1cefec28bc404184768fb0a96
Reviewed-on: https://review.spdk.io/gerrit/c/spdk/spdk/+/12613
Community-CI: Mellanox Build Bot
Community-CI: Broadcom CI <spdk-ci.pdl@broadcom.com>
Tested-by: SPDK CI Jenkins <sys_sgci@intel.com>
Reviewed-by: Michael Haeuptle <michaelhaeuptle@gmail.com>
Reviewed-by: Ben Walker <benjamin.walker@intel.com>
Reviewed-by: Tomasz Zawadzki <tomasz.zawadzki@intel.com>
We see reports that Huawei SSDs can't handle hardware
SGL properly, it requires additional alignment, so add
a quirk here to force Huawei SSDs use PRP instead.
Fix#2489.
Change-Id: I20a57e754bc6ff8666d681191994818f2192decc
Signed-off-by: Changpeng Liu <changpeng.liu@intel.com>
Reviewed-on: https://review.spdk.io/gerrit/c/spdk/spdk/+/12405
Community-CI: Mellanox Build Bot
Community-CI: Broadcom CI <spdk-ci.pdl@broadcom.com>
Tested-by: SPDK CI Jenkins <sys_sgci@intel.com>
Reviewed-by: wanghailiang <hailiangx.e.wang@intel.com>
Reviewed-by: Aleksey Marchuk <alexeymar@nvidia.com>
Reviewed-by: GangCao <gang.cao@intel.com>
Reviewed-by: Dong Yi <dongx.yi@intel.com>
Reviewed-by: Jim Harris <james.r.harris@intel.com>
When sending the first part of a fuse command, we set the
first_fused_submitted flag so that we don't ring the doorbell
immediately. When the second part is sent, we ring the doorbell for
both commands.
However, this doesn't work well when we use the option to delay ringing
the doorbell. We send both parts, then later when we try to ring the
doorbell, we don't because of the first_fused_submitted flag from the
first command.
Replace this mechanism by keeping track of the last submitted fuse.
Change-Id: Ia4ac9b3ce9c319ee4c7e42f86eadda93dac85fca
Signed-off-by: Alex Michon <amichon@kalrayinc.com>
Reviewed-on: https://review.spdk.io/gerrit/c/spdk/spdk/+/12182
Community-CI: Broadcom CI <spdk-ci.pdl@broadcom.com>
Community-CI: Mellanox Build Bot
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>
Reviewed-by: Konrad Sztyber <konrad.sztyber@intel.com>
Now controller initialization with RDMA
transport is fully async
Signed-off-by: Alexey Marchuk <alexeymar@mellanox.com>
Change-Id: I26e857740d3137d0b0e987facc81fc5f6ef81f2b
Reviewed-on: https://review.spdk.io/gerrit/c/spdk/spdk/+/10756
Community-CI: Broadcom CI <spdk-ci.pdl@broadcom.com>
Tested-by: SPDK CI Jenkins <sys_sgci@intel.com>
Reviewed-by: Jim Harris <james.r.harris@intel.com>
Reviewed-by: Shuhei Matsumoto <smatsumoto@nvidia.com>
Transport specific qpair_abort_reqs() set SC to SC_ABORTED_SQ_DELETION.
However, nvme_qpair_abort_queued_reqs() set SC to SC_ABORTED_BY_REQUEST
even if its call is not requested by the upper layer.
Change nvme_qpair_abort_queued_reqs() to set SC to SC_ABORTED_SQ_DELETION
for consistency.
nvme_qpair_abort_queued_reqs() is used to abort queued requests that
were sent while adminq was connecting. SC_ABORTED_SQ_DELETION will not
be so bad even for the case.
This change is required for the NVMe bdev module to be resilient for I/O
error. The NVMe bdev module does not retry I/O if SC is
SC_ABORTED_BY_REQUEST.
SC is set to SC_INTERNAL_DEVICE_ERROR if a request is failed to submit
to qpair by a generic qpair layer. We can change it to
SC_ABORTED_SQ_DELETION as well but we keep this for now.
SC_INTERNAL_DEVICE_ERROR is also retriable for the NVMe bdev module.
Signed-off-by: Shuhei Matsumoto <smatsumoto@nvidia.com>
Change-Id: I7d8d5e97b222fe9275afc4fed024c1654c9579a2
Reviewed-on: https://review.spdk.io/gerrit/c/spdk/spdk/+/12121
Community-CI: Broadcom CI <spdk-ci.pdl@broadcom.com>
Tested-by: SPDK CI Jenkins <sys_sgci@intel.com>
Reviewed-by: Jim Harris <james.r.harris@intel.com>
Reviewed-by: Aleksey Marchuk <alexeymar@mellanox.com>
The value of ack_timeout is calculated according to
the formula 2^(transport_ack_timeout) msec.
Signed-off-by: zhangduan <zhangd28@chinatelecom.cn>
Change-Id: I5a938635d70693ddd405fa5907555bb745b4df0f
Reviewed-on: https://review.spdk.io/gerrit/c/spdk/spdk/+/12215
Community-CI: Broadcom CI <spdk-ci.pdl@broadcom.com>
Community-CI: Mellanox Build Bot
Tested-by: SPDK CI Jenkins <sys_sgci@intel.com>
Reviewed-by: Shuhei Matsumoto <smatsumoto@nvidia.com>
Reviewed-by: Aleksey Marchuk <alexeymar@mellanox.com>
Reviewed-by: Jim Harris <james.r.harris@intel.com>
Now that IO qpairs can be created asynchronously, we need to make sure
that all the create IO CQ/SQ commands can be executed simultaneously.
It is pretty common to create multiple IO qpairs at the same time, e.g.
adding an NVMe bdev to an nvmf subsystem will create an IO qpair on each
poll group. In that case, if the number of cores exceed the size of the
admin queue (actually it can be even lower due to outstanding AERs), we
might run out nvme_requests on the admin queue.
The chosen minimum value for the admin queue size, 256, should be enough
to cover most cases.
Fixes#2465
Signed-off-by: Konrad Sztyber <konrad.sztyber@intel.com>
Change-Id: I55c59aef64f3fdb33f7b4824d3e9beb403602633
Reviewed-on: https://review.spdk.io/gerrit/c/spdk/spdk/+/12270
Community-CI: Broadcom CI <spdk-ci.pdl@broadcom.com>
Community-CI: Mellanox Build Bot
Tested-by: SPDK CI Jenkins <sys_sgci@intel.com>
Reviewed-by: Shuhei Matsumoto <smatsumoto@nvidia.com>
Reviewed-by: Tomasz Zawadzki <tomasz.zawadzki@intel.com>
The RDMA transport can disconnect qpair asynchronously now.
Previously, we tried to release the resource of the qpair after disconnected.
However it did not work because it was done when deleting the qpair.
The admin qpair was not deleted in a ctrlr reset sequence.
This patch tries to satisfy the same aim again but by a different way.
Previously, we released the resource of the qpair before starting
actual disconnection process. This patch release the resource of the
qpair after the qpair is actually disconnected.
The related patches are:
b9518a5540eb09178a59
Signed-off-by: Shuhei Matsumoto <smatsumoto@nvidia.com>
Change-Id: Id6a814895a35b1589b781a91744ef872b42aaa69
Reviewed-on: https://review.spdk.io/gerrit/c/spdk/spdk/+/11783
Community-CI: Broadcom CI <spdk-ci.pdl@broadcom.com>
Community-CI: Mellanox Build Bot
Tested-by: SPDK CI Jenkins <sys_sgci@intel.com>
Reviewed-by: Aleksey Marchuk <alexeymar@mellanox.com>
Reviewed-by: Ben Walker <benjamin.walker@intel.com>
The code to handle the lingering qpair when deleting it was really
complicated.
The RDMA transport can connect or disconnect qpair asynchronously.
Then we can include the code to handle the lingering qpair into the
code to disconnect qpair now.
If the disconnected qpair is still busy, defer completion of the
disconnection until qpair becomes idle.
If poll group is not used, we can complete disconnection immediately
because cq is already destroyed.
The related data and unit test cases are not necessary anymore.
So delete them in this patch.
Signed-off-by: Shuhei Matsumoto <smatsumoto@nvidia.com>
Change-Id: Ic8f81143fcad0714ac9b7db862313aa8094eeefb
Reviewed-on: https://review.spdk.io/gerrit/c/spdk/spdk/+/11778
Community-CI: Broadcom CI <spdk-ci.pdl@broadcom.com>
Community-CI: Mellanox Build Bot
Tested-by: SPDK CI Jenkins <sys_sgci@intel.com>
Reviewed-by: Aleksey Marchuk <alexeymar@mellanox.com>
Reviewed-by: Ben Walker <benjamin.walker@intel.com>
Include delayed disconnect/connect retries with finite times into
the state machine of asynchronous qpair connnection.
We do not need to call back to the common transport layer but
we need to do the following, clear rqpair->cq before starting disconnection
if qpair uses poll group, and clear qpair->transport_failure_reason after
disconnected.
Additionally locate the new state STALE_CONN before INITIALIZING
because cq is not ready to use for admin qpair when the state is
STALE_CONN.
Signed-off-by: Shuhei Matsumoto <smatsumoto@nvidia.com>
Change-Id: Ibc779a2b772be9506ffd8226d5f64d6d12102ff2
Reviewed-on: https://review.spdk.io/gerrit/c/spdk/spdk/+/11690
Community-CI: Broadcom CI <spdk-ci.pdl@broadcom.com>
Community-CI: Mellanox Build Bot
Tested-by: SPDK CI Jenkins <sys_sgci@intel.com>
Reviewed-by: Ben Walker <benjamin.walker@intel.com>
Reviewed-by: Aleksey Marchuk <alexeymar@mellanox.com>
TCP transport already does it but was not documented clearly.
RDMA and PCIe transports follow it and document it clearly.
Then we can check each qpair's state if
spdk_nvme_poll_group_process_completions() returns -ENXIO before
disconnected_qpair_cb() is called.
Signed-off-by: Shuhei Matsumoto <smatsumoto@nvidia.com>
Change-Id: I2afe920cfd06c374251fccc1c205948fb498dd33
Reviewed-on: https://review.spdk.io/gerrit/c/spdk/spdk/+/11328
Community-CI: Mellanox Build Bot
Community-CI: Broadcom CI <spdk-ci.pdl@broadcom.com>
Tested-by: SPDK CI Jenkins <sys_sgci@intel.com>
Reviewed-by: Aleksey Marchuk <alexeymar@mellanox.com>
Reviewed-by: Ben Walker <benjamin.walker@intel.com>
Add three states, INITIALIZING, EXITING, and EXITED to the rqpair
state.
Add async parameter to nvme_rdma_ctrlr_create_qpair() and set it
to opts->async_mode for I/O qpair and true for admin qpair.
Replace all nvme_rdma_process_event() calls by
nvme_rdma_process_event_start() calls.
nvme_rdma_ctrlr_connect_qpair() sets rqpair->state to INITIALIZING
when starting to process CM events.
nvme_rdma_ctrlr_connect_qpair_poll() calls
nvme_rdma_process_event_poll() with ctrlr->ctrlr_lock if qpair is
not admin qpair.
nvme_rdma_ctrlr_disconnect_qpair() returns if qpair->async is true
or qpair->poll_group is not NULL before polling CM events, or polls
CM events until completion otherwise. Add comments to clarify why
we do like this.
nvme_rdma_poll_group_process_completions() does not process submission
for any qpair which is still connecting.
Change-Id: Ie04c3408785124f2919eaaba7b2bd68f8da452c9
Signed-off-by: Shuhei Matsumoto <smatsumoto@nvidia.com>
Reviewed-on: https://review.spdk.io/gerrit/c/spdk/spdk/+/11442
Community-CI: Broadcom CI <spdk-ci.pdl@broadcom.com>
Community-CI: Mellanox Build Bot
Tested-by: SPDK CI Jenkins <sys_sgci@intel.com>
Reviewed-by: Aleksey Marchuk <alexeymar@mellanox.com>
Reviewed-by: Ben Walker <benjamin.walker@intel.com>
We can use lib/vfio-user API to setup BAR0 doorbells,
existing implementation is redundant.
Change-Id: Ib880d167c84c6b8482bf1a35559a34c939f6a02d
Signed-off-by: Changpeng Liu <changpeng.liu@intel.com>
Reviewed-on: https://review.spdk.io/gerrit/c/spdk/spdk/+/12211
Community-CI: Broadcom CI <spdk-ci.pdl@broadcom.com>
Community-CI: Mellanox Build Bot
Tested-by: SPDK CI Jenkins <sys_sgci@intel.com>
Reviewed-by: John Levon <levon@movementarian.org>
Reviewed-by: Jim Harris <james.r.harris@intel.com>
Reviewed-by: Tomasz Zawadzki <tomasz.zawadzki@intel.com>
The process of matching qpair to poll group is split into
two distinct parts that occur on different threads.
See spdk_nvmf_tgt_new_qpair().
This results in a race condition for TCP between spdk_sock_map_lookup()
and spdk_sock_map_insert(), which are called in spdk_nvmf_get_optimal_poll_group()
and spdk_nvmf_poll_group_add() respectively.
Fixes#2113
This patch picks a hint from nvmf_tcp for next poll group,
which is then passed down to spdk_sock_map_lookup().
When matching placement_id exists, but does not have
a poll group assigned - the hint will be used.
Signed-off-by: Tomasz Zawadzki <tomasz.zawadzki@intel.com>
Change-Id: I4abde2bc9c39225c9f5dd7c3654fa2639bb0a27f
Reviewed-on: https://review.spdk.io/gerrit/c/spdk/spdk/+/10271
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>
