Purpose: to make the timeout work for NVMe TCP transport,
we miss this for TCP transport.
Change-Id: Iab4af988cc4796b4d6d98430453f3dbce1fcf313
Signed-off-by: Ziye Yang <optimistyzy@gmail.com>
Reviewed-on: https://review.gerrithub.io/c/445117
Tested-by: SPDK CI Jenkins <sys_sgci@intel.com>
Reviewed-by: Changpeng Liu <changpeng.liu@intel.com>
Reviewed-by: Shuhei Matsumoto <shuhei.matsumoto.xt@hitachi.com>
Reviewed-by: Seth Howell <seth.howell5141@gmail.com>
Reviewed-by: Ben Walker <benjamin.walker@intel.com>
This patch refactors driver init and in doing so eliminates the mem
leak described in the GitHub issue. Also it is now consistent with
how the pending compression driver does init.
Fixes#633
Change-Id: Ia2d55d9e98fb9470ff8f9b34aeb4ee9f3d0478f5
Signed-off-by: paul luse <paul.e.luse@intel.com>
Reviewed-on: https://review.gerrithub.io/c/442896
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>
We should not add addtional check since we already have this
option in timeout_cb function, the addtional check is unnecessary.
Change-Id: I77c89303155e0c14072a1838994f9e76a0ffc0f4
Signed-off-by: Ziye Yang <optimistyzy@gmail.com>
Reviewed-on: https://review.gerrithub.io/c/445319
Tested-by: SPDK CI Jenkins <sys_sgci@intel.com>
Reviewed-by: Seth Howell <seth.howell5141@gmail.com>
Reviewed-by: Changpeng Liu <changpeng.liu@intel.com>
Reviewed-by: Ben Walker <benjamin.walker@intel.com>
This patch is used to implement this function.
Since we need to call nvme_tcp_req_complete in this
function, so we need to adjust the location of the
nvme_tcp_rep_complete funtion.
Change-Id: I5fc3693aec8dc166ac1eb03babcd2d73d7b00e63
Signed-off-by: Ziye Yang <optimistyzy@gmail.com>
Reviewed-on: https://review.gerrithub.io/c/444489
Tested-by: SPDK CI Jenkins <sys_sgci@intel.com>
Reviewed-by: Shuhei Matsumoto <shuhei.matsumoto.xt@hitachi.com>
Reviewed-by: Changpeng Liu <changpeng.liu@intel.com>
Reviewed-by: Ben Walker <benjamin.walker@intel.com>
Reviewed-by: Seth Howell <seth.howell5141@gmail.com>
In this patch series, spdk_bdev_scsi_read and spdk_bdev_scsi_write
became almost identical. Hence squash them into spdk_bdev_scsi_read_write.
Change-Id: Ibbaddf74c1bf2dac37a0133eac27086af650a061
Signed-off-by: Shuhei Matsumoto <shuhei.matsumoto.xt@hitachi.com>
Reviewed-on: https://review.gerrithub.io/c/444780
Tested-by: SPDK CI Jenkins <sys_sgci@intel.com>
Reviewed-by: Jim Harris <james.r.harris@intel.com>
Reviewed-by: Darek Stojaczyk <dariusz.stojaczyk@intel.com>
This is in a effort to consolidate SCSI read and write I/O
for the upcoming transparent DIF support.
Previously conversion of bytes and blocks are done both in
SCSI layer and BDEV layer. After the patch series, conversion is
consolidated into SCSI layer.
Change-Id: Ib964a41ec22757f2a09cea22f398903f78d0781f
Signed-off-by: Shuhei Matsumoto <shuhei.matsumoto.xt@hitachi.com>
Reviewed-on: https://review.gerrithub.io/c/444779
Tested-by: SPDK CI Jenkins <sys_sgci@intel.com>
Reviewed-by: Jim Harris <james.r.harris@intel.com>
Reviewed-by: Darek Stojaczyk <dariusz.stojaczyk@intel.com>
This is in a effort to consolidate SCSI read and write I/O
for the upcoming transparent DIF support.
Previously conversion of bytes and blocks are done both in
SCSI layer and BDEV layer. After the patch series, conversion is
consolidated into SCSI layer.
About conversion from bytes to blocks, we don't expose bdev API
spdk_bdev_bytes_to_blocks and but create private helper function
_bytes_to_blocks because we will use not block size but data
block size when we support transparent DIF feature.
Change-Id: I37169c673479c92e027e2507a0e54a1e414b43e1
Signed-off-by: Shuhei Matsumoto <shuhei.matsumoto.xt@hitachi.com>
Reviewed-on: https://review.gerrithub.io/c/444778
Tested-by: SPDK CI Jenkins <sys_sgci@intel.com>
Reviewed-by: Jim Harris <james.r.harris@intel.com>
Reviewed-by: Darek Stojaczyk <dariusz.stojaczyk@intel.com>
The last parameter xfer_len of spdk_bdev_scsi_read is not used,
and of spdk_bdev_scsi_write is used only to check task->transfer_len.
Hence remove the last parameter xfer_len from spdk_bdev_scsi_read/write
and extract the check operation from spdk_bdev_scsi_write and insert
it into spdk_bdev_scsi_read_write.
Additionally, remove a debug log because xfer_len is not passed to
spdk_bdev_scsi_write anymore. Hopufully, this will not degrade any
maintainability.
On top of this, factoring out the operation to convert byte to
block in spdk_bdev_scsi_read/write be done.
Change-Id: I35faca269a9c4a7f15d27e8e61b6a1b809a36b3f
Signed-off-by: Shuhei Matsumoto <shuhei.matsumoto.xt@hitachi.com>
Reviewed-on: https://review.gerrithub.io/c/444776
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: Darek Stojaczyk <dariusz.stojaczyk@intel.com>
Currently the task and buffer pools are hardcoded to
512 each. This effectively limits the queue depth
per channel to 256 since we need both source and
destination buffers from the buffer pool. So make
the pool sizes dynamic based on the user's queue depth
input.
Signed-off-by: Jim Harris <james.r.harris@intel.com>
Change-Id: I52f04308c0329fa99277746a0768eb1214e37a8e
Reviewed-on: https://review.gerrithub.io/c/445357
Tested-by: SPDK CI Jenkins <sys_sgci@intel.com>
Reviewed-by: Shuhei Matsumoto <shuhei.matsumoto.xt@hitachi.com>
Reviewed-by: Darek Stojaczyk <dariusz.stojaczyk@intel.com>
Reviewed-by: Changpeng Liu <changpeng.liu@intel.com>
Use the new spdk_ioat_flush interface to batch
doorbell writes - this significantly improves
descriptor throughput. For now, just set the
threshold for writing the descriptor to half of
the queue depth. We can always modify this later
to allow very specific thresholds, but for now
this simple change is sufficient.
Increases 512B descriptor throughput at QD=256 from
3.7M/s to 14.0M/s on my Skylake Xeon server.
Signed-off-by: Jim Harris <james.r.harris@intel.com>
Change-Id: I4504579e23cee5b6a1044849c49d33d58bdb51a9
Reviewed-on: https://review.gerrithub.io/c/445355
Tested-by: SPDK CI Jenkins <sys_sgci@intel.com>
Reviewed-by: Shuhei Matsumoto <shuhei.matsumoto.xt@hitachi.com>
Reviewed-by: Darek Stojaczyk <dariusz.stojaczyk@intel.com>
Reviewed-by: Changpeng Liu <changpeng.liu@intel.com>
This helps ensure it gets inlined in the spdk_vtophys
code path, now that spdk_vtophys is defined in the same
compilation module.
Signed-off-by: Jim Harris <james.r.harris@intel.com>
Change-Id: I0d0d9bba4295f0d9a7c0657834aa5d39f3b682d8
Reviewed-on: https://review.gerrithub.io/c/445354
Tested-by: SPDK CI Jenkins <sys_sgci@intel.com>
Reviewed-by: Shuhei Matsumoto <shuhei.matsumoto.xt@hitachi.com>
Reviewed-by: Darek Stojaczyk <dariusz.stojaczyk@intel.com>
Reviewed-by: Changpeng Liu <changpeng.liu@intel.com>
CPU profiling on workloads with intensive vtophys
operations (i.e. very small CB-DMA transfers) exposed
overhead introduce by spdk_vtophys having to call
spdk_mem_map_translate in a different compilation
unit. Let's just move the vtophys.c contents into
memory.c so that spdk_vtophys can inline
spdk_mem_map_translate and avoid this extra overhead.
This of course breaks the memory and vtophys unit
tests, so some additional changes are needed there
to keep everything linking correctly.
Signed-off-by: Jim Harris <james.r.harris@intel.com>
Change-Id: I295ed5f441d3eec7abdbc9d881c49d2174ec9f48
Reviewed-on: https://review.gerrithub.io/c/444975
Tested-by: SPDK CI Jenkins <sys_sgci@intel.com>
Reviewed-by: Shuhei Matsumoto <shuhei.matsumoto.xt@hitachi.com>
Reviewed-by: Darek Stojaczyk <dariusz.stojaczyk@intel.com>
Reviewed-by: Changpeng Liu <changpeng.liu@intel.com>
When called from scripts, sometimes the arguments passed
will be empty.
Change-Id: I2f9fa1daa173eecc7b20928cd359284bb6f74ff3
Signed-off-by: Ben Walker <benjamin.walker@intel.com>
Reviewed-on: https://review.gerrithub.io/c/444278
Tested-by: SPDK CI Jenkins <sys_sgci@intel.com>
Reviewed-by: Jim Harris <james.r.harris@intel.com>
Reviewed-by: Darek Stojaczyk <dariusz.stojaczyk@intel.com>
Previously, we can -p + hex value(e.g., 0x1) to assign the master core
and start the NVMe-oF or iSCSI target app.
However now it is not supported and prints error. I checked
the code, it only supports transformation with Decimal format,
so chaning the base to 0 to make it supporting other formats.
Change-Id: I82510ba0cef47b5593484b4fd3490f85c93cf6a5
Signed-off-by: Ziye Yang <optimistyzy@gmail.com>
Reviewed-on: https://review.gerrithub.io/c/444830
Reviewed-by: Shuhei Matsumoto <shuhei.matsumoto.xt@hitachi.com>
Reviewed-by: Darek Stojaczyk <dariusz.stojaczyk@intel.com>
Tested-by: SPDK CI Jenkins <sys_sgci@intel.com>
This function should be changed as well but was overlooked.
Fixing this here.
Change-Id: I2f7c8e475d89cfe177441e1843d3713397f6fe7b
Signed-off-by: Pawel Wodkowski <pawelx.wodkowski@intel.com>
Reviewed-on: https://review.gerrithub.io/c/444079
Tested-by: SPDK CI Jenkins <sys_sgci@intel.com>
Reviewed-by: Shuhei Matsumoto <shuhei.matsumoto.xt@hitachi.com>
Reviewed-by: Darek Stojaczyk <dariusz.stojaczyk@intel.com>
Reviewed-by: Ben Walker <benjamin.walker@intel.com>
Do not set the completion_update bit except on
the last descriptor built before the dmacount doorbell
is written. This allows much better batching of
completions (to match batching of the submissions).
Signed-off-by: Jim Harris <james.r.harris@intel.com>
Change-Id: Idd0281fb2e9e1ad2eb0f65f097c54fc051dfd935
Reviewed-on: https://review.gerrithub.io/c/444974
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>
Add spdk_ioat_build_copy and spdk_ioat_build_fill
which mirror the existing spdk_ioat_submit_copy
and spdk_ioat_submit_fill. These new functions
*only* build the descriptors in the ring - they do
not write the doorbell. This enables batching
which can significantly improve performance by
reducing the number of MMIO writes.
Signed-off-by: Jim Harris <james.r.harris@intel.com>
Change-Id: Ia3539f936924b7f833f4a7b963d06ffefa68379f
Reviewed-on: https://review.gerrithub.io/c/444973
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>
This will enable batching of doorbell writes in
future commits. For now, just make the API public.
This is the first in a series of patches that
drastically improves performance for high queue
depth CB-DMA workloads. Some basic tests on
my Xeon E5-v3 platform shows about 4x improvement
for 512B transfers.
Signed-off-by: Jim Harris <james.r.harris@intel.com>
Change-Id: Ia8d28a63f5020ae8644c1efdec7f68740bb6920c
Reviewed-on: https://review.gerrithub.io/c/444972
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>
To enable the timeout function.
Change-Id: Id5c40848957743683b6a5c2d085e7f777f14497d
Signed-off-by: Ziye Yang <optimistyzy@gmail.com>
Reviewed-on: https://review.gerrithub.io/c/444803
Tested-by: SPDK CI Jenkins <sys_sgci@intel.com>
Reviewed-by: Changpeng Liu <changpeng.liu@intel.com>
Reviewed-by: Shuhei Matsumoto <shuhei.matsumoto.xt@hitachi.com>
Reviewed-by: Ben Walker <benjamin.walker@intel.com>
Use NBD_SET_SOCK to check whether the nbd device is setup
by other process or whether nbd kernel module is ready
before other nbd ioctl operations. This can avoid bad
influence to the nbd device setup by other process.
Change-Id: Ic12acbfddb8c4388e25731c39159b1ce559b8f23
Signed-off-by: Xiaodong Liu <xiaodong.liu@intel.com>
Reviewed-on: https://review.gerrithub.io/c/444805
Tested-by: SPDK CI Jenkins <sys_sgci@intel.com>
Reviewed-by: Pawel Wodkowski <pawelx.wodkowski@intel.com>
Reviewed-by: Ben Walker <benjamin.walker@intel.com>
Reviewed-by: Jim Harris <james.r.harris@intel.com>
The ioctl NBD_SET_SOCK can return EBUSY on conditions not
only the kernel module hasn't loaded entirely yet, but
also the nbd device is setup by another process, which will
lead the poller's infinite polling.
This patch will wait only 1 second if device is busy.
Change-Id: I8b1cfab725cba180f774a57ced3fa4ba81da2037
Signed-off-by: Xiaodong Liu <xiaodong.liu@intel.com>
Reviewed-on: https://review.gerrithub.io/c/444804
Tested-by: SPDK CI Jenkins <sys_sgci@intel.com>
Reviewed-by: Darek Stojaczyk <dariusz.stojaczyk@intel.com>
Reviewed-by: Pawel Wodkowski <pawelx.wodkowski@intel.com>
Reviewed-by: Ben Walker <benjamin.walker@intel.com>
A new module switch which was missed at here.
Change-Id: If1784ace13657756d8034cd04e594af5b1799381
Signed-off-by: yidong0635 <dongx.yi@intel.com>
Reviewed-on: https://review.gerrithub.io/c/444820
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>
There is no need to lock g_ftl_bdev_lock when unregister a ftl_bdev.
Besides, the destructor of ftl_bdev will lock it again.
Change-Id: I99870483183879d9422584dbac6e154f605daea8
Signed-off-by: wuzhouhui <wuzhouhui@kingsoft.com>
Reviewed-on: https://review.gerrithub.io/c/444794
Tested-by: SPDK CI Jenkins <sys_sgci@intel.com>
Reviewed-by: Changpeng Liu <changpeng.liu@intel.com>
Reviewed-by: Konrad Sztyber <konrad.sztyber@intel.com>
Reviewed-by: Ben Walker <benjamin.walker@intel.com>
Added check before write submission to indicate if
LBA was update in meantime. In such case don't set band's
metadata and rwb entry cache bit. Previous implementation
invalidates such address during write completion and could
cause that inconsistent lba map was stored into disk.
Change-Id: I4353d9f96c53132ca384aeca43caef8d11f07fa4
Signed-off-by: Wojciech Malikowski <wojciech.malikowski@intel.com>
Reviewed-on: https://review.gerrithub.io/c/444403
Tested-by: SPDK CI Jenkins <sys_sgci@intel.com>
Reviewed-by: Konrad Sztyber <konrad.sztyber@intel.com>
Reviewed-by: Ben Walker <benjamin.walker@intel.com>
Reviewed-by: Jim Harris <james.r.harris@intel.com>
We assumed io_channel allocation always succeeds, but
that's not true. Doing I/O to any vhost session that
failed to allocate an io_channel would most likely
cause a crash.
We'll now detect io_channel allocation failure and
print a proper error message. The SCSI target for
which the channel allocation failed simply won't be
visible to the vhost master. All I/O to that target
will be rejected.
We should probably report the error to the upper
layer and either prevent the device from starting
or fail the SCSI target hotplug request. But for now
let's just prevent the crash.
Change-Id: I735dfb930d8905f70636a236b4fa94288d0aaf3a
Signed-off-by: Darek Stojaczyk <dariusz.stojaczyk@intel.com>
Reviewed-on: https://review.gerrithub.io/c/444874
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>
nvme_ctrlr_submit_admin_request() will access admin queue, and we
should hold ctrl->ctrlr_lock when access it.
Change-Id: Iff576fe5e14e854eb38dbc64d6c6d9ec1ba17056
Signed-off-by: wuzhouhui <wuzhouhui@kingsoft.com>
Reviewed-on: https://review.gerrithub.io/c/444793
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>
Also use the same style condition check for secondary process
with PCIE type.
Change-Id: I93c83126145255887914ef5efea1a493c8f7f767
Signed-off-by: Changpeng Liu <changpeng.liu@intel.com>
Reviewed-on: https://review.gerrithub.io/c/444492
Reviewed-by: Ben Walker <benjamin.walker@intel.com>
Reviewed-by: Jim Harris <james.r.harris@intel.com>
Tested-by: SPDK CI Jenkins <sys_sgci@intel.com>
The helper function spdk_get_data_out_buffer_size() is a little
confusing because it does only returning macro constant
SPDK_ISCSI_MAX_RECV_DATA_SEGMENT_LENGTH.
The macro constant will be configurable and so the helper function
is not sustainable.
Replace the helper function simply by the macro constant.
Change-Id: I4ec300f61783da7bb712512603c2dd80987ec702
Signed-off-by: Shuhei Matsumoto <shuhei.matsumoto.xt@hitachi.com>
Reviewed-on: https://review.gerrithub.io/c/444537
Tested-by: SPDK CI Jenkins <sys_sgci@intel.com>
Reviewed-by: Pawel Wodkowski <pawelx.wodkowski@intel.com>
Reviewed-by: Ben Walker <benjamin.walker@intel.com>
Reviewed-by: Jim Harris <james.r.harris@intel.com>
When hotplug feature is enabled by NVMe driver, users may
call delete_nvme_controller() RPC to delete one controller,
however, the hotplug monitor will probe this controller
automaticlly and attach it back to NVMe driver. We added
a skip list, for those user deleted controllers so that
NVMe driver will not attach it again.
Fix issue #602.
Change-Id: Ibbe21ff8a021f968305271acdae86207e6228e20
Signed-off-by: Changpeng Liu <changpeng.liu@intel.com>
Reviewed-on: https://review.gerrithub.io/c/444323
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>
Error logs in nvmf_rdma_dump_request lead to report error about
address points to the zero page, add judgement to return.
this issue occurs in heavy load fio testing.
Change-Id: I50302be88b3af53f718e3800aa16df7c506ca4e8
Signed-off-by: yidong0635 <dongx.yi@intel.com>
Reviewed-on: https://review.gerrithub.io/c/441110
Tested-by: SPDK CI Jenkins <sys_sgci@intel.com>
Reviewed-by: Ben Walker <benjamin.walker@intel.com>
Reviewed-by: Seth Howell <seth.howell5141@gmail.com>
Reviewed-by: Changpeng Liu <changpeng.liu@intel.com>
User can create a probe context to probe and attach controllers
asynchronously, the controllers will be added to the context list
for the first step, then users can poll the context until the list
becomes empty.
Change-Id: I3a96e2d8a9724332ff15542f78f9553fdab505e2
Signed-off-by: Changpeng Liu <changpeng.liu@intel.com>
Reviewed-on: https://review.gerrithub.io/c/442664
Tested-by: SPDK CI Jenkins <sys_sgci@intel.com>
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>
Existing NVMe driver uses a global list g_nvme_init_ctrlrs
to track the controllers during initialization, and internal
function will start each controller in the list one by one
until the list is empty. We introduce a probe context
and move the global list into the context, with the context
we can enable asynchronous probe API in the next patch, also
this can enable parallel probe feature.
Change-Id: I538537abe8c1a4a82fb168ca8055de42caa6e4f9
Signed-off-by: Changpeng Liu <changpeng.liu@intel.com>
Reviewed-on: https://review.gerrithub.io/c/426304
Tested-by: SPDK CI Jenkins <sys_sgci@intel.com>
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>
Previously, function spdk_nvme_probe_internal() will probe
NVMe controllers and then bring up probed controllers
into the ready state after that. Broke up original two parts
with probe and start stage, this will help us to introduce
a probe context in the next patch.
Change-Id: Ie0c55a6a5463fb437f84349b0b2b33a217ba63e0
Signed-off-by: Changpeng Liu <changpeng.liu@intel.com>
Reviewed-on: https://review.gerrithub.io/c/426303
Tested-by: SPDK CI Jenkins <sys_sgci@intel.com>
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>
