Patch 55f947933 ("bdev: remove spdk_bdev_ext_io_opts from spdk_bdev_io")
changed the way bdev_nvme submits IO to the NVMe driver causing
performance degradation for requests with iovcnt = 1, as they also had
to go through the path that executes the reset_sgl/next_sge callbacks.
This patch reverts those changes back to the original code checking
iovcnt and using the non-SGL functions if possible.
Suggested-by: Jim Harris <james.r.harris@intel.com>
Signed-off-by: Konrad Sztyber <konrad.sztyber@intel.com>
Change-Id: I5e7c6620d38b7690ff862d8cd0075afacc578217
Reviewed-on: https://review.spdk.io/gerrit/c/spdk/spdk/+/16961
Tested-by: SPDK CI Jenkins <sys_sgci@intel.com>
Reviewed-by: Tomasz Zawadzki <tomasz.zawadzki@intel.com>
Reviewed-by: Aleksey Marchuk <alexeymar@nvidia.com>
The spdk_bdev_ext_io_opts structure is used to pass extra options when
submitting a bdev IO request, without having to modify/add functions to
handle new options. Additionally, the structure has a size field to
allow adding new fields without breaking the ABI (and thus having to
bump up the major version of a library).
It is also a part of spdk_bdev_io and there are several reasons for
removing it from that structure:
1. The size field only makes sense in structures that are passed
through pointers. And spdk_bdev_ext_io_opts is indeed passed as a
pointer to spdk_bdev_{readv,writev}_blocks_ext(), however it is
also embedded in spdk_bdev_io (internal.ext_opts_copy), which is
also part of the API. It means that each time a new field is added
to spdk_bdev_ext_io_opts, the size of spdk_bdev_io will also
change, so we will need to bump the major version of libspdk_bdev
anyway, thus making spdk_bdev_ext_io_opts.size useless.
2. The size field also makes internal.ext_opts cumbersome to use, as
each time one of its fields is accessed, we need to check the size.
Currently the code doesn't do that, because all of the existing
spdk_bdev_ext_io_opts fields were present when this structure was
initially introduced, but we'd need to do check the size before
accessing any new fields.
3. spdk_bdev_ext_io_opts has a metadata field, while spdk_bdev_io
already has u.bdev.md_buf, which means that we store the same thing
in several different places in spdk_bdev_io (u.bdev.md_buf,
u.bdev.ext_opts->metadata, internal.ext_opts->metadata).
Therefore, this patch removes all references to spdk_bdev_ext_io_opts
from spdk_bdev_io and replaces them with fields (memory_domain,
memory_domain_ctx) that were missing in spdk_bdev_io. Unfortunately,
this change breaks the API and requires changes in bdev modules that
supported spdk_bdev_io.u.bdev.ext_opts.
Signed-off-by: Konrad Sztyber <konrad.sztyber@intel.com>
Change-Id: I49b7524eb84d1d4d7f12b7ab025fec36da1ee01f
Reviewed-on: https://review.spdk.io/gerrit/c/spdk/spdk/+/16773
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>
The io paths' stat will get lost when they are destroyed. Record
the stat in the nvme_ns structure.
Change-Id: I12fc0b04fac0d59e7465fe543ee733f2822a9cdb
Signed-off-by: Richael Zhuang <richael.zhuang@arm.com>
Reviewed-on: https://review.spdk.io/gerrit/c/spdk/spdk/+/14744
Reviewed-by: Jim Harris <james.r.harris@intel.com>
Reviewed-by: Aleksey Marchuk <alexeymar@nvidia.com>
Reviewed-by: Shuhei Matsumoto <smatsumoto@nvidia.com>
Tested-by: SPDK CI Jenkins <sys_sgci@intel.com>
Currently we have stat per bdev I/O channel, but for NVMe bdev
multipath, we don't have stat per I/O path. Especially for
active-active mode, we may want to observe each path's statistics.
This patch support IO stat for nvme_io_path. Record each nvme_io_path
stat using structure spdk_bdev_io_stat.
The following is the comparison of bdevperf test.
Test on Arm server with the following basic configuration.
1 Null bdev: block size: 4K, num_blocks:16k
run bdevperf with io size=4k, qdepth=1/32/128, rw type=randwrite/mixed with 70% read/randread
Each time run 30 seconds, each item run for 16 times and get the average.
The result is as follows.
qdepth type IOPS(default) IOPS(this patch) diff
1 randwrite 7795157.27 7859909.78 0.83%
1 mix(70% r) 7418607.08 7404026.54 -0.20%
1 randread 8053560.83 8046315.44 -0.09%
32 randwrite 15409191.3 15327642.11 -0.53%
32 mix(70% r) 13760145.97 13714666.28 -0.33%
32 randread 16136922.98 16038855.39 -0.61%
128 randwrite 14815647.56 14944902.74 0.87%
128 mix(70% r) 13414858.59 13412317.46 -0.02%
128 randread 15508642.43 15521752.41 0.08%
Change-Id: I4eb5673f49d65d3ff9b930361d2f31ab0ccfa021
Signed-off-by: Richael Zhuang <richael.zhuang@arm.com>
Reviewed-on: https://review.spdk.io/gerrit/c/spdk/spdk/+/14743
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>
Reviewed-by: Shuhei Matsumoto <smatsumoto@nvidia.com>
Support to specify rr_min_io for multipath round-robin policy,
which makes I/O switches to another io path after rr_min_io I/Os are
rounted to current io path.
Change-Id: I09f0d8d24271c0178ff816fa63ce8576b6e8ae47
Signed-off-by: Richael Zhuang <richael.zhuang@arm.com>
Reviewed-on: https://review.spdk.io/gerrit/c/spdk/spdk/+/15445
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>
Reviewed-by: Shuhei Matsumoto <smatsumoto@nvidia.com>
Support selecting io path according to number of outstanding io of
each path in a channel. It's optional, and can be set by calling
RPC "bdev_nvme_set_multipath_policy -s queue_depth".
Change-Id: I82cdfbd69b3e105c973844c4f34dc98f0dca2faf
Signed-off-by: Richael Zhuang <richael.zhuang@arm.com>
Reviewed-on: https://review.spdk.io/gerrit/c/spdk/spdk/+/14734
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: Jim Harris <james.r.harris@intel.com>
When an I/O gets an I/O error, the I/O path to which the I/O was
submitted may be still available. In this case, the I/O should be
retried to the same I/O path. However, a new I/O path was always
selected for an I/O retry.
For the active/passive policy, the same I/O path was selected naturally.
However, for the active/active policy, it was very likely that a
different I/O path was selected.
To use the same I/O path for an I/O retry, add a helper function
bdev_nvme_retry_io() into bdev_nvme_retry_ios() and replace
bdev_nvme_submit_request() by bdev_nvme_retry_io(). bdev_nvme_retry_io()
checks if nbdev_io->io_path is not NULL and is available. Then, call
_bdev_nvme_submit_request() if true, or call bdev_nvme_submit_request()
otherwise. For I/O path error, clear nbdev_io->io_path for
clarification. Add unit test to verify this change.
Linux kernel native NVMe multipath already takes this approach. Hence,
this change will be reasonable.
Signed-off-by: Shuhei Matsumoto <smatsumoto@nvidia.com>
Change-Id: I7022aafd8b1cdd5830c4f743d64b080aa970cf8d
Reviewed-on: https://review.spdk.io/gerrit/c/spdk/spdk/+/16015
Tested-by: SPDK CI Jenkins <sys_sgci@intel.com>
Reviewed-by: Aleksey Marchuk <alexeymar@nvidia.com>
Reviewed-by: Richael <richael.zhuang@arm.com>
Reviewed-by: Jim Harris <james.r.harris@intel.com>
The following patches will change I/O retry to use the same io_path if
it is still available. However, bdev_nvme_submit_request() always calls
bdev_nvme_find_io_path() first. For I/O retry, if possible, we want to
skip calling bdev_nvme_find_io_path() and use nbdev_io->io_path instead.
To reuse the code as much as possible and not to touch the fast code
path, factor out request submit functions from
bdev_nvme_submit_request() into _bdev_nvme_submit_request().
While developing this patch, a bug/mismatch was found such that
bdev_io->internal.ch was different from ch of
bdev_nvme_submit_request(). Fix it together in this patch.
Signed-off-by: Shuhei Matsumoto <smatsumoto@nvidia.com>
Change-Id: Id003e033ecde218d1902bca5706c772edef5d5e5
Reviewed-on: https://review.spdk.io/gerrit/c/spdk/spdk/+/16013
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>
Error counters for NVMe error was added in the generic bdev layer but
we want to know more detailed information for some use cases.
Add NVMe error counters per type and per code as module specific
statistics.
For status codes, the first idea was to have different named member
for each status code value. However, it was bad and too hard to test,
review, and maintain.
Instead, we have just two dimensional uint32_t arrays, and increment
one of these uint32_t values based on the status code type and status
code. Then, when dump the JSON, we use spdk_nvme_cpl_get_status_string()
and spdk_nvme_cpl_get_status_type_string().
This idea has one potential downside. This idea consumes 4 (types) *
256 (codes) * 4 (counter) = 4KB per NVMe bdev. We can make this smarter
if memory allocation is a problem. Hence we add an option
nvme_error_stat to enable this feature only if the user requests.
Additionally, the string returned by spdk_nvme_cpl_get_status_string()
or spdk_nvme_cpl_get_status_type_string() has uppercases, spaces, and
hyphens. These should not be included in JSON strings. Hence, convert
these via spdk_strcpy_replace().
Signed-off-by: Shuhei Matsumoto <smatsumoto@nvidia.com>
Change-Id: I07b07621e777bdf6556b95054abbbb65e5f9ea3e
Reviewed-on: https://review.spdk.io/gerrit/c/spdk/spdk/+/15370
Reviewed-by: Jim Harris <james.r.harris@intel.com>
Reviewed-by: Konrad Sztyber <konrad.sztyber@intel.com>
Reviewed-by: Aleksey Marchuk <alexeymar@nvidia.com>
Tested-by: SPDK CI Jenkins <sys_sgci@intel.com>
Community-CI: Mellanox Build Bot
Found with misspell-fixer.
Signed-off-by: Michal Berger <michal.berger@intel.com>
Change-Id: If062df0189d92e4fb2da3f055fb981909780dc04
Reviewed-on: https://review.spdk.io/gerrit/c/spdk/spdk/+/15207
Reviewed-by: Tomasz Zawadzki <tomasz.zawadzki@intel.com>
Reviewed-by: Shuhei Matsumoto <smatsumoto@nvidia.com>
Tested-by: SPDK CI Jenkins <sys_sgci@intel.com>
Added io_outstanding in struct nvme_io_path to record outstanding
I/O number in each path, which will be used by multipath to select
I/O path.
io_outstanding gets updated for I/O sent to a namespace and not get
updated if sent to a controller.
For FLUSH case, it calls bdev_nvme_io_complete() directly and
io_outstanding is not updated for this case.
spdk_bdev_io_get_buf() is executed in the generic bdev layer.
Hence, we do not update io_outstanding for spdk_bdev_io_get_buf().
Change-Id: I47b515e0f254e5daa7e1e88799a832032b23ff34
Signed-off-by: Richael Zhuang <richael.zhuang@arm.com>
Reviewed-on: https://review.spdk.io/gerrit/c/spdk/spdk/+/15032
Tested-by: SPDK CI Jenkins <sys_sgci@intel.com>
Community-CI: Mellanox Build Bot
Reviewed-by: Jim Harris <james.r.harris@intel.com>
Reviewed-by: Aleksey Marchuk <alexeymar@nvidia.com>
Reviewed-by: Shuhei Matsumoto <smatsumoto@nvidia.com>
Test uniqueness of generated UUIDs in two cases:
1) For a single controller with two namespaces.
Serial number remains the same, but different
namespace IDs should contribute enough to get
a unique UUID.
2) For two controllers with one namespace each.
Serial numbers differ only by one character and
namespace IDs are the same. Single different
character in serial number should make UUID
unique.
Also validate UUID generated from serial number
containing only space characters.
Change-Id: I16e39e269ced4d8405fb5b3af6aa8bf98ecfd7ba
Signed-off-by: Krzysztof Karas <krzysztof.karas@intel.com>
Reviewed-on: https://review.spdk.io/gerrit/c/spdk/spdk/+/15362
Tested-by: SPDK CI Jenkins <sys_sgci@intel.com>
Reviewed-by: Tomasz Zawadzki <tomasz.zawadzki@intel.com>
Reviewed-by: Jim Harris <james.r.harris@intel.com>
Community-CI: Mellanox Build Bot
per Intel policy to include file commit date using git cmd
below. The policy does not apply to non-Intel (C) notices.
git log --follow -C90% --format=%ad --date default <file> | tail -1
and then pull just the 4 digit year from the result.
Intel copyrights were not added to files where Intel either had
no contribution ot the contribution lacked substance (ie license
header updates, formatting changes, etc). Contribution date used
"--follow -C95%" to get the most accurate date.
Note that several files in this patch didn't end the license/(c)
block with a blank comment line so these were added as the vast
majority of files do have this last blank line. Simply there for
consistency.
Signed-off-by: paul luse <paul.e.luse@intel.com>
Change-Id: Id5b7ce4f658fe87132f14139ead58d6e285c04d4
Reviewed-on: https://review.spdk.io/gerrit/c/spdk/spdk/+/15192
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>
Community-CI: Mellanox Build Bot
This is consistent with the use of terms in other parts of SPDK and fits
with the code living under module/
Change-Id: If182f7cf2d160d57443a1b5f24e0065f191b59b2
Signed-off-by: Ben Walker <benjamin.walker@intel.com>
Reviewed-on: https://review.spdk.io/gerrit/c/spdk/spdk/+/13919
Tested-by: SPDK CI Jenkins <sys_sgci@intel.com>
Community-CI: Mellanox Build Bot
Reviewed-by: Aleksey Marchuk <alexeymar@nvidia.com>
Reviewed-by: Jim Harris <james.r.harris@intel.com>
After a controller was hot-removed, if a reset sequence started to
the controller, spdk_nvme_ctrlr_disconnect() failed and caused core
dump in debug mode.
When implemented, how to cause the failure and how to process the
failure were not clear. Hence assert was added to detect the
failure.
We know how we cause the failure now. Let's handle the failure
appropriately.
If spdk_nvme_ctrlr_disconnect() fails, we are on the nvme_ctrlr->thread.
Hence call bdev_nvme_reset_complete() with failure immediately.
Even if spdk_nvme_ctrlr_disconnect() completes synchronously, the
completion callback is executed asynchronously when polling an adminq.
Hence set the completion callback only if spdk_nvme_ctrlr_disconnect()
succeeds.
Fixes issue #2632
Signed-off-by: Shuhei Matsumoto <smatsumoto@nvidia.com>
Change-Id: I11f61853aba9eca2515592f964a291e59def7247
Reviewed-on: https://review.spdk.io/gerrit/c/spdk/spdk/+/13892
Tested-by: SPDK CI Jenkins <sys_sgci@intel.com>
Reviewed-by: Ben Walker <benjamin.walker@intel.com>
Reviewed-by: Aleksey Marchuk <alexeymar@nvidia.com>
Reviewed-by: Krzysztof Karas <krzysztof.karas@intel.com>
Reviewed-by: Jim Harris <james.r.harris@intel.com>
These checked parameters are necessary themselves even for single path
configuration.
Signed-off-by: Shuhei Matsumoto <smatsumoto@nvidia.com>
Change-Id: Ie1eb2f51eeec1dbc634c6bae462a41d4c209d6ac
Reviewed-on: https://review.spdk.io/gerrit/c/spdk/spdk/+/12052
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: Dong Yi <yidong0635@126.com>
Community-CI: Mellanox Build Bot
The word engine was both used (interchangeably with module) to refer to
the things that plug into the framework and to the framework itself.
This patch eliminates all use of the word engine that meant the
framework. It leaves uses of the word that meant "module".
Change-Id: I6b9b50e2f045ac39f2a74d0152ee8d6269be4bd1
Signed-off-by: Ben Walker <benjamin.walker@intel.com>
Reviewed-on: https://review.spdk.io/gerrit/c/spdk/spdk/+/13918
Tested-by: SPDK CI Jenkins <sys_sgci@intel.com>
Community-CI: Broadcom CI <spdk-ci.pdl@broadcom.com>
Reviewed-by: Shuhei Matsumoto <smatsumoto@nvidia.com>
Reviewed-by: Paul Luse <paul.e.luse@intel.com>
Reviewed-by: Jim Harris <james.r.harris@intel.com>
Reviewed-by: Aleksey Marchuk <alexeymar@nvidia.com>
bdev_nvme_create_bdev_channel_cb() did not initialized the multipath
policy of the newly created channel. 0 was active-passive and hence
multipath policy was always initialized to active-passive.
Fix the bug and add unit tests for verification.
Signed-off-by: Shuhei Matsumoto <smatsumoto@nvidia.com>
Change-Id: I6e44108740da4b9ff72311ae4b5500558c65c5c3
Reviewed-on: https://review.spdk.io/gerrit/c/spdk/spdk/+/14225
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: Aleksey Marchuk <alexeymar@nvidia.com>
Community-CI: Broadcom CI <spdk-ci.pdl@broadcom.com>
Admin passthrough supported retry and failover as same as I/O by
using the bdev_retry_count. However, doing retry or failover for
admin passthrough may have unexpected side effects and its value
is not clear. The safest way is to limit retry and failover for I/O.
If we need to support retry and failover for admin passthrough,
restore the code and add a new option bdev_admin_retry_count.
Signed-off-by: Shuhei Matsumoto <smatsumoto@nvidia.com>
Change-Id: I680513a40a80041f6ea6f546c74c672f2a81812d
Reviewed-on: https://review.spdk.io/gerrit/c/spdk/spdk/+/14227
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: Ben Walker <benjamin.walker@intel.com>
Reviewed-by: Jim Harris <james.r.harris@intel.com>
Reviewed-by: Aleksey Marchuk <alexeymar@nvidia.com>
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>
Calculation of the ANA log page size should use the
identify ctrl MNAN field
(maximum number of allowed namespaces)
not the NN (maximum valid nsid value).
An ANA-enabled controller must have a non-zero MNAN value,
see NVMe Base Specification, Figure 251,
therefore nvme_ctrlr_init_ana_log_page() may safely use MNAN.
Since NN might be much higher than MNAN,
ANA log size based on NN may results in a very large
log page and cause a failure to get ANA log,
e.g. if it is larger than the controller's MDTS.
Fix: replace cdata->nn with cdata->mnan
in nvme_ctrlr_init_ana_log_page()
Signed-off-by: Anton Eidelman <anton@lightbitslabs.com>
Change-Id: I2a522dca833a27dddad25848d7688efa23d23091
Reviewed-on: https://review.spdk.io/gerrit/c/spdk/spdk/+/13039
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: Shuhei Matsumoto <smatsumoto@nvidia.com>
Reviewed-by: Jacek Kalwas <jacek.kalwas@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>
During ctrlr reconnection, spdk_nvme_ctrlr_reconnect_poll_async()
is executed by a non-timed poller.
We should poll adminq more often during ctrlr disconnection too.
Signed-off-by: Shuhei Matsumoto <smatsumoto@nvidia.com>
Change-Id: Ib1f5b41015aed20deda8df6f2c837981ac233c04
Reviewed-on: https://review.spdk.io/gerrit/c/spdk/spdk/+/12615
Reviewed-by: Dong Yi <dongx.yi@intel.com>
Reviewed-by: Jim Harris <james.r.harris@intel.com>
Reviewed-by: Aleksey Marchuk <alexeymar@nvidia.com>
Reviewed-by: Michael Haeuptle <michaelhaeuptle@gmail.com>
Reviewed-by: Ben Walker <benjamin.walker@intel.com>
Community-CI: Broadcom CI <spdk-ci.pdl@broadcom.com>
Tested-by: SPDK CI Jenkins <sys_sgci@intel.com>
As described in the NVMe specification, a controller level reset
includes the following actions:
- the controller stops processing any outstanding admin or I/O
commands;
- all I/O SQs and CQs are deleted.
In a full controller reset sequence for a PCIe controller, if we do
a controller level reset first, we can abort outstanding commands
after the hardware has actually been stopped.
For NVMe-oF controller, each I/O qpair is an independent network
connection and is disconnected safely. We do not want to change
NVMe-oF controller.
Fixes the issue #2360
Signed-off-by: Shuhei Matsumoto <smatsumoto@nvidia.com>
Change-Id: If05febac74705bfd3df5abd15064c1203126e027
Reviewed-on: https://review.spdk.io/gerrit/c/spdk/spdk/+/12447
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@nvidia.com>
Reviewed-by: Michael Haeuptle <michaelhaeuptle@gmail.com>
Reviewed-by: Ben Walker <benjamin.walker@intel.com>
By default, failback to the preferred I/O path is done automatically
if it is restored. Some users may want to keep using the backup I/O
path even if the preferred I/O path is restored. In this case,
bdev_nvme_set_preferred_path can be used to do manual failback.
We may be able to clear/fill I/O path cache more strictly but it will
be complicated and have bugs. This patch does the minimal change,
just skips an apparent case.
Signed-off-by: Shuhei Matsumoto <smatsumoto@nvidia.com>
Change-Id: I78fe5faee6ff04e88ae3d7c6be6da1c20637c912
Reviewed-on: https://review.spdk.io/gerrit/c/spdk/spdk/+/12431
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: Jim Harris <james.r.harris@intel.com>
Reviewed-by: Ben Walker <benjamin.walker@intel.com>
The NVMe bdev module supported active-passive policy for multipath mode
first. By this patch, the NVMe bdev module supports active-active policy
for multipath node next. Following the Linux kernel native NVMe multipath,
the NVMe bdev module supports round robin algorithm for active-active
policy.
The multipath policy, active-passive or active-active, is managed per
nvme_bdev. The multipath policy is copied to all corresponding
nvme_bdev_channels.
Different from active-passive, active-active caches even non_optimized
path to provide load balance across multiple paths.
Signed-off-by: Shuhei Matsumoto <smatsumoto@nvidia.com>
Change-Id: Ie18b24db60d3da1ce2f83725b6cd3079f628f95b
Reviewed-on: https://review.spdk.io/gerrit/c/spdk/spdk/+/12001
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: Ben Walker <benjamin.walker@intel.com>
Reviewed-by: Aleksey Marchuk <alexeymar@nvidia.com>
If we specify a preferred path manually for each NVMe bdev, we will
be able to realize a simple static load balancing and make the failover
more controllable in the multipath mode.
The idea is to move I/O path to the NVMe-oF controller to the head of
the list and then clear the I/O path cache for each NVMe bdev channel.
We can set the I/O path to the I/O path cache directly but it must be
conditional and make the code very complex. Hence, let find_io_path() do
that.
However, a NVMe bdev channel may be acquired after setting the preferred
path. To cover such case, sort the nvme_ns list of the NVMe bdev too.
This feature supports only multipath mode. The NVMe bdev module supports
failover mode too. However, to support the latter, the new RPC needs to
have trid as parameters and the code and the usage will be come very
complex. Add a note for such limitation.
To verify one by one exactly, add unit test.
Signed-off-by: Shuhei Matsumoto <smatsumoto@nvidia.com>
Change-Id: Ia51c74f530d6d7dc1f73d5b65f854967363e76b0
Reviewed-on: https://review.spdk.io/gerrit/c/spdk/spdk/+/12262
Community-CI: Mellanox Build Bot
Community-CI: Broadcom CI <spdk-ci.pdl@broadcom.com>
Reviewed-by: <tanl12@chinatelecom.cn>
Reviewed-by: GangCao <gang.cao@intel.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>
When sending a fused compare and write command, we pass a callback
bdev_nvme_comparev_and_writev_done that we expect to be called twice
before marking the io as completed. In order to detect if a call to
bdev_nvme_comparev_and_writev_done is the first or the second one, we
currently rely on the opcode in cdw0. However, cdw0 may be set to 0,
especially when aborting the command. This may cause use-after-free
issues and this may call the user callbacks twice instead of once.
Use a bit in the nvme_bdev_io instead to keep track of the number of
calls to bdev_nvme_comparev_and_writev_done.
Signed-off-by: Alex Michon <amichon@kalrayinc.com>
Change-Id: I0474329e87648e44b08998d0552b2a9dd5d34ac2
Reviewed-on: https://review.spdk.io/gerrit/c/spdk/spdk/+/12180
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>
Previously, if a namespace is in ANA inaccessible state, I/O had been
queued infinitely. Fix this issue according to the NVMe spec.
Add a temporary poller anatt_timer and a flag ana_transition_timedout for
each nvme_ns.
Start anatt_timer if the nvme_ns enters ANA transition. If anatt_timer
is expired, set ana_transition_timedout to true. Cancel anatt_timer or
clear ana_transition_timedout if the nvme_ns exits ANA transition.
nvme_io_path_become_available() returns false if ana_transition_timedout
is true.
Add unit test case to verify these addition.
Signed-off-by: Shuhei Matsumoto <smatsumoto@nvidia.com>
Change-Id: Ic76933242046b3e8e553de88221b943ad097c91c
Reviewed-on: https://review.spdk.io/gerrit/c/spdk/spdk/+/12194
Tested-by: SPDK CI Jenkins <sys_sgci@intel.com>
Community-CI: Broadcom CI <spdk-ci.pdl@broadcom.com>
Reviewed-by: Aleksey Marchuk <alexeymar@mellanox.com>
Reviewed-by: Jim Harris <james.r.harris@intel.com>
Reviewed-by: Monica Kenguva <monica.kenguva@intel.com>
there is no group
The real implementation handles this by returning -ENOENT, so do the
same in the test.
Signed-off-by: Ben Walker <benjamin.walker@intel.com>
Change-Id: I405b6f60bf4dcdb22c57e48bbaf66d57522a49c5
Reviewed-on: https://review.spdk.io/gerrit/c/spdk/spdk/+/11508
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: Jim Harris <james.r.harris@intel.com>
Reviewed-by: Michael Haeuptle <michaelhaeuptle@gmail.com>
Count disconnecting a queue pair as activity so that the unit test
poll_threads() calls don't bail out until the disconnectedd_qpair_cb is
called at least once.
Signed-off-by: Ben Walker <benjamin.walker@intel.com>
Change-Id: Idc437d6c589dbf133bfcbb5edba1087f928a718c
Reviewed-on: https://review.spdk.io/gerrit/c/spdk/spdk/+/11507
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: Jim Harris <james.r.harris@intel.com>
Reviewed-by: Michael Haeuptle <michaelhaeuptle@gmail.com>
This was neither set nor used.
Signed-off-by: Ben Walker <benjamin.walker@intel.com>
Change-Id: I3119135843c5fc0b8724e593db40df46e6b5bdb0
Reviewed-on: https://review.spdk.io/gerrit/c/spdk/spdk/+/12097
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: Michael Haeuptle <michaelhaeuptle@gmail.com>
Reviewed-by: Aleksey Marchuk <alexeymar@mellanox.com>
Reviewed-by: Jim Harris <james.r.harris@intel.com>
Bdev modules must not access internal bdev_io
structure, so add a new pointer in a public
section. Pointer in internal section will be
used in next patch
Signed-off-by: Alexey Marchuk <alexeymar@mellanox.com>
Change-Id: Ib631563015b3e5fa9300d22b7ae59d8db43c8275
Reviewed-on: https://review.spdk.io/gerrit/c/spdk/spdk/+/10421
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: Tomasz Zawadzki <tomasz.zawadzki@intel.com>
Reviewed-by: Shuhei Matsumoto <smatsumoto@nvidia.com>
Reviewed-by: Jim Harris <james.r.harris@intel.com>
For RDMA transport, adminq will find transport error first because
usually only adminq polls CM events.
Change-Id: I7b22cc8883bf02198f1a90d2654c1de6f2e736e6
Signed-off-by: Shuhei Matsumoto <smatsumoto@nvidia.com>
Reviewed-on: https://review.spdk.io/gerrit/c/spdk/spdk/+/11331
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>
If qpair is disconnected asynchronously, it takes time from detecting
transport error to actually disconnected. We should avoid using the
path as soon as possible after detecting any transport error.
Poll group clears I/O path cache if it finds transport error and avoid
using the path which had transport error.
These changes will reduce the failover time.
Signed-off-by: Shuhei Matsumoto <smatsumoto@nvidia.com>
Change-Id: I00580159a84372a115ed5e62a6ce13eed4368999
Reviewed-on: https://review.spdk.io/gerrit/c/spdk/spdk/+/11329
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>
spdk_nvme_ctrlr_disconnect() will be made asynchronous in the
following patches and so we will need to have some changes.
spdk_nvme_ctrlr_disconnect() disconnects adminq and ctrlr synchronously
now.
If spdk_nvme_ctrlr_disconnect() is made asynchronous,
spdk_nvme_ctrlr_process_admin_completions() will complete to disconnect
adminq and ctrlr, and will return -ENXIO only if adminq is disconnected.
However even now spdk_nvme_ctrlr_process_admin_completions() returns
-ENXIO if adminq is disconnected.
So as a preparation, set a callback before calling spdk_nvme_ctrlr_disconnect()
and call the callback if it is set and spdk_nvme_ctrlr_process_admin_completions()
returns -ENXIO.
Besides, fix the return value of bdev_nvme_poll_adminq() in this patch.
Change-Id: I2559f86bb8cf9a92b5b386ed816c00b08c9832df
Signed-off-by: Shuhei Matsumoto <smatsumoto@nvidia.com>
Reviewed-on: https://review.spdk.io/gerrit/c/spdk/spdk/+/10950
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>
As we do when deleting ctrlr_channel, disconnect and then free I/O
qpair in a ctrlr reset sequence.
Deleting ctrlr_channel and resetting ctrlr_channel may cause conflicts.
This patch processes such conflicts correctly.
If destroy_ctrlr_channel_cb() is executed between pending and executing
reset_destroy_qpair(), reset_destroy_qpair() is not executed because
ctrlr_channel is not found. In this case, destroy_qpair_channel()
starts disconnecting qpair and deletes ctrlr_channel. Then
disconnected_qpair_cb() releases a reference to poll group.
If destroy_ctrlr_channel_cb() is excuted between executing reset_destroy_qpair()
and disconnected_qpair_cb(), destroy_ctrlr_channel_cb() skips
ctrlr_channel for a reset sequence.
Change-Id: I1f49f74b94aefbea178680aa53ded3a12876c676
Signed-off-by: Shuhei Matsumoto <smatsumoto@nvidia.com>
Reviewed-on: https://review.spdk.io/gerrit/c/spdk/spdk/+/10766
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>
spdk_nvme_poll_group has followed spdk_nvme_qpair about how to
process I/O qpair deletion inside of a completion context.
spdk_nvme_qpair_process_completions() accesses qpair after
returning from nvme_transport_qpair_process_completions().
So this is reasonable.
On the other hand, if spdk_nvme_poll_group_process_completions()
can execute spdk_nvme_ctrlr_free_io_qpair() inside of a completion
context, the target qpair is ensured to be deleted after returning
from spdk_nvme_ctrlr_free_io_qpair(). Then the target qpair is
not accessed anymore in spdk_nvme_poll_group_process_completions().
Remove two variables, in_completion_context and num_qpairs_to_delete,
of spdk_nvme_transport_poll_group and the related code.
This change is really necessary to support the following case.
In the NVMe bdev module, a nvme_qpair has a qpair and a poll_group
channel. disconnected_qpair_cb calls spdk_nvme_ctrlr_free_io_qpair()
for the qpair and spdk_put_io_channel() to the poll_group_channel.
spdk_nvme_ctrlr_free_io_qpair() is executed after unwinding stack
but spdk_put_io_channel() is executed now. The callback to
spdk_put_io_channel() calls spdk_nvme_poll_group_destroy(). However,
spdk_nvme_ctrlr_free_io_qpair() is not executed. Hence
spdk_nvme_poll_group_destroy() fails.
Update the corresponding stub in unit test together.
Signed-off-by: Shuhei Matsumoto <smatsumoto@nvidia.com>
Change-Id: Icd1f1daf049c6c7ffb28790fe87989a1060f8952
Reviewed-on: https://review.spdk.io/gerrit/c/spdk/spdk/+/11496
Community-CI: Broadcom CI <spdk-ci.pdl@broadcom.com>
Community-CI: Mellanox Build Bot
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>
This is another preparation to disconnect qpair asynchronously.
Add nvme_qpair object and move the qpair and poll_group pointers and
the io_path_list list from nvme_ctrlr_channel to nvme_qpair. nvme_qpair
is allocated dynamically when creating nvme_ctrlr_channel, and
nvme_ctrlr_channel points to nvme_qpair.
We want to keep the times of references at I/O path. Change nvme_io_path
to point nvme_qpair instead of nvme_ctrlr_channel, and add
nvme_ctrlr_channel pointer to nvme_qpair.
nvme_ctrlr_channel may be freed earlier than nvme_qpair. nvme_poll_group
lists nvme_qpair instead of nvme_ctrlr_channel and nvme_qpair has a
pointer to nvme_ctrlr.
By using the nvme_ctrlr pointer of the nvme_qpair, a helper function
nvme_ctrlr_channel_get_ctrlr() is not necessary any more. Remove it.
Change-Id: Ib3f579d3441f31b9db7d3844ec56c49e2bb53a5d
Signed-off-by: Shuhei Matsumoto <smatsumoto@nvidia.com>
Reviewed-on: https://review.spdk.io/gerrit/c/spdk/spdk/+/11832
Community-CI: Broadcom CI <spdk-ci.pdl@broadcom.com>
Community-CI: Mellanox Build Bot
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>
In the following patches, spdk_nvme_ctrlr_disconnect_io_qpair() will
be changed to be asynchronous, spdk_nvme_ctrlr_disconnect_io_qpair()
will be called first and then spdk_nvme_ctrlr_free_io_qpair() after
the qpair is actually disconnected.
We will not be able to keep the current bdev_nvme_destroy_qpair()
function.
As a preparation, inline bdev_nvme_destroy_qpair() and remove it.
Additionally, this patch has the following changes.
Previously I/O qpair was freed and then I/O path caches were cleared.
Both are SPDK thread local. So there is no dependency for the ordering
of these two operations. However, it will reduce the size of the
following patches if we clear I/O path caches before freeing I/O qpair
when the qpair is disconnected. Hence we clear I/O path caches and then
free I/O qpair.
Remove DTRACE for bdev_nvme_destroy_qpair() for now.
It will be restored in the following patches.
Furthermore, fix potential NULL pointer acces in
bdev_nvme_create_qpair().
Change-Id: I0ab78ccb0d240e56b95b53179341afcd909a31f6
Signed-off-by: Shuhei Matsumoto <smatsumoto@nvidia.com>
Reviewed-on: https://review.spdk.io/gerrit/c/spdk/spdk/+/10746
Community-CI: Broadcom CI <spdk-ci.pdl@broadcom.com>
Community-CI: Mellanox Build Bot
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 following patches will enable us to specify I/O error resiliency
options per nvme_ctrlr as global options. To do it easier, move
per controller options about I/O error resiliency into struct nvme_ctrlr_opts.
prchk_flags is not exactly for resiliency but move it into struct
nvme_ctrlr_opts too.
Signed-off-by: Shuhei Matsumoto <smatsumoto@nvidia.com>
Change-Id: I85fd1738bb6e293cd804b086ade82274485f213d
Reviewed-on: https://review.spdk.io/gerrit/c/spdk/spdk/+/11829
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>
If 0 - UINT32_MAX or UINT32_MAX - 0 is substituted into a int variable,
we cannot get any expected result.
Fix the bug and add unit test case to verify the fix.
Signed-off-by: Shuhei Matsumoto <smatsumoto@nvidia.com>
Change-Id: Ib045273238753e16755328805b38569909c8b83a
Reviewed-on: https://review.spdk.io/gerrit/c/spdk/spdk/+/11836
Tested-by: SPDK CI Jenkins <sys_sgci@intel.com>
Reviewed-by: Aleksey Marchuk <alexeymar@mellanox.com>
Reviewed-by: Jim Harris <james.r.harris@intel.com>
Reviewed-by: Ben Walker <benjamin.walker@intel.com>
Community-CI: Broadcom CI <spdk-ci.pdl@broadcom.com>
Community-CI: Mellanox Build Bot
This patch aligns namespace comparison with Linux kernel
implementation:
- UUID is optional and may be NULL
- command set (CSI) should be the same
Signed-off-by: Evgeniy Kochetov <evgeniik@nvidia.com>
Change-Id: I8f889989f24cd51b104057217f87eb303b30fa68
Reviewed-on: https://review.spdk.io/gerrit/c/spdk/spdk/+/11312
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: Jim Harris <james.r.harris@intel.com>
Reviewed-by: Aleksey Marchuk <alexeymar@mellanox.com>
