Add ICL_NOCOPY flag to icl_pdu_append_data(), specifying that the method
can just reference the data buffer instead of immediately copying it.
Extend the offload KPI with optional PDU queue method, allowing to specify
completion callback, called when all the data referenced by above has been
transferred and won't be accessed any more (the buffers can be freed).
Implement the above functionality in software iSCSI driver using mbufs
with external storage and reference counter. Note that some NICs (ixl(4))
may keep the mbuf in TX queue for a long time, so CTL has to be ready.
Add optional method to struct ctl_scsiio for buffer reference counting.
Implement it for CTL block backend, allowing to delay free of the struct
ctl_be_block_io and memory it references as needed. In first reincarnation
of the patch I tried to delay whole I/O as it is done for FibreChannel,
that was cleaner, but due to the above callback delays I had to rewrite
it this way to not leave LUN referenced potentially for hours or more.
All together on sequential read from ZFS ARC this saves about 30% of CPU
time and memory bandwidth by avoiding one of 3 memory copies (the other
two are from ZFS ARC to DMU cache and then from DMU cache to CTL buffers).
On tests with 2x Xeon Silver 4114 this allows to reach full line rate of
100GigE NIC. Tests with Gold CPUs and two 100GigE NICs are stil TBD,
but expectations to saturate them are pretty high. ;)
Discussed with: Chelsio
Sponsored by: iXsystems, Inc.
cs_cmdsn can be incremented with single atomic. expcmdsn/maxcmdsn set in
cfiscsi_pdu_prepare() based on cs_cmdsn are not required to be updated
synchronously, only monotonically, that is achieved with lock there.
MFC after: 2 weeks
Sponsored by: iXsystems, Inc.
We any way have per-I/O space in CTL_PRIV_FRONTEND, while for PDU private
fields I have better use ideas. Plus to me such use of PDU fields looked
a layering violation.
MFC after: 2 weeks
Sponsored by: iXsystems, Inc.
I don't see a point to copy io->scsiio.kern_total_len into the request
PDU private field. The io is going to stay with us till the end, and
kern_total_len field is not changed after being first initialized.
MFC after: 2 weeks
Sponsored by: iXsystems, Inc.
While there, remove ifdef around cs_target check in cfiscsi_ioctl_list().
I am not sure why this ifdef was added, but without this check code will
crash below on NULL dereference.
Submitted by: Aleksandr Fedorov <aleksandr.fedorov@itglobal.com>
MFC after: 1 week
Differential Revision: https://reviews.freebsd.org/D24587
r357614 added CTLFLAG_NEEDGIANT to make it easier to find nodes that are
still not MPSAFE (or already are but aren’t properly marked).
Use it in preparation for a general review of all nodes.
This is non-functional change that adds annotations to SYSCTL_NODE and
SYSCTL_PROC nodes using one of the soon-to-be-required flags.
Mark all obvious cases as MPSAFE. All entries that haven't been marked
as MPSAFE before are by default marked as NEEDGIANT
Approved by: kib (mentor, blanket)
Commented by: kib, gallatin, melifaro
Differential Revision: https://reviews.freebsd.org/D23718
of the fix/workaround for the "ctld hanging on reload" problem.
PR: 220175
Reported by: Eugene M. Zheganin <emz at norma.perm.ru>
Tested by: Eugene M. Zheganin <emz at norma.perm.ru>
Approved by: re (kib)
MFC after: 2 weeks
Sponsored by: playkey.net
for the "ctld hanging on reload" problem observed in same cases under
high load. I'm not 100% sure it's _the_ fix, as the issue is rather hard
to reproduce, but it was tested as part of a larger path and the problem
disappeared. It certainly shouldn't break anything.
Now, technically, it shouldn't be needed. Quoting mav@, "After
ct->ct_online == 0 there should be no new sessions attached to the target.
And if you see some problems abbout it, it may either mean that there are
some races where single cfiscsi_session_terminate(cs) call may be lost,
or as a guess while this thread was sleeping target was reenabbled and
redisabled again". Should such race be discovered and properly fixed
in the future, than this and the followup two commits can be backed out.
PR: 220175
Reported by: Eugene M. Zheganin <emz at norma.perm.ru>
Tested by: Eugene M. Zheganin <emz at norma.perm.ru>
Discussed with: mav
Approved by: re (gjb)
MFC after: 2 weeks
Sponsored by: playkey.net
race condition, due to a missing call to cfiscsi_target_release().
Discussed with: mav@
Tested by: Eugene M. Zheganin <emz at norma.perm.ru> (earlier version)
MFC after: 2 weeks
Sponsored by: playkey.net
ioctl frontend ports.
This revision introduces two changes to CTL:
- Changes the way options are passed to CTL_LUN_REQ and CTL_PORT_REQ ioctls.
Removes ctl_be_arg structure and associated logic and replaces it with
nv(3)-based logic for passing in and out arguments.
- Allows creating multiple ioctl frontend ports using either ctladm(8) or
ctld(8).
New frontend ports are represented by /dev/cam/ctl<pp>.<vp> nodes, eg /dev/cam/ctl5.3.
Those device nodes respond only to CTL_IO ioctl.
New command-line options for ctladm:
# creates new ioctl frontend port with using free pp and vp=0
ctladm port -c
# creates new ioctl frontend port with pp=10 and vp=0
ctladm port -c -O pp=10
# creates new ioctl frontend port with pp=11 and vp=12
ctladm port -c -O pp=11 -O vp=12
# removes port with number 4 (it's a "targ_port" number, not pp number)
ctladm port -r -p 4
New syntax for ctl.conf:
target ... {
port ioctl/<pp>
...
}
target ... {
port ioctl/<pp>/<vp>
...
Note: Most of this work was made by jceel@, thank you.
Submitted by: jceel
Reworked by: myself
Reviewed by: mav (earlier versions and recently during the rework)
Obtained from: FreeNAS and TrueOS
Relnotes: Yes
Sponsored by: iXsystems Inc.
Differential Revision: https://reviews.freebsd.org/D9299
The crash scenario goes like this: there's a thread waiting on "reinstate";
because it doesn't update the timeout counter it gets terminated by the
callout; at this point the maintenance thread starts the termination routine.
The first thread finishes waiting, proceeds to icl_conn_handoff(), and drops
the refcount, which allows the maintenance thread to free its resources. At
this point another thread receives a PDU. Boom.
PR: 222898, 219866
Reported by: Eugene M. Zheganin <emz at norma.perm.ru>
Tested by: Eugene M. Zheganin <emz at norma.perm.ru>
Reviewed by: mav@ (earlier version)
MFC after: 2 weeks
Sponsored by: playkey.net
Mainly focus on files that use BSD 2-Clause license, however the tool I
was using misidentified many licenses so this was mostly a manual - error
prone - task.
The Software Package Data Exchange (SPDX) group provides a specification
to make it easier for automated tools to detect and summarize well known
opensource licenses. We are gradually adopting the specification, noting
that the tags are considered only advisory and do not, in any way,
superceed or replace the license texts.
The goal of this work is to remove the explicit dependency for ctl(4)
on iscsi(4), so end-users without iscsi(4) support in the kernel can
use ctl(4) for its other functions.
This allows those without iscsi(4) support built into the kernel to use
ctl(4) as a test mechanism. As a sidenote, this was possible around the
10.0-RELEASE period, but made impossible for end-users without iscsi(4)
between 10.0-RELEASE and 11.0-RELEASE.
Automatically load cfiscsi(4) from ctladm(8) and ctld(8) for backwards
compatibility with previously releases. The automatic loading feature is
compiled into the beforementioned tools if MK_ISCSI == yes when building
world.
Add a manpage for cfiscsi(4) and refer to it in ctl(4).
Differential Revision: D10099
MFC after: 2 months
Relnotes: yes
Reviewed by: mav, trasz
Sponsored by: Dell EMC Isilon
Before this change MaxCmdSN was reported as CmdSN + delta, that made it
limit number of requests in transmission from the initiator to target,
that was pretty useless. After this change MaxCmdSN limits number of
requests queued to CTL, i.e. maximal queue depth for the initiator.
The default limit is 256 outstanding requests per initiator at a time.
This code uses existing cs_outstanding_ctl_pdus counter to track queue
depth. It's semantics doen't perfectly match, but close enough to not
add another counter. Just don't set the maxtags below 2.
MFC after: 2 weeks
It is only a first step and not perfect, but better then nothing.
The main blocker is CAM target frontend, that can not be unloaded,
since CAM does not have mechanism to unregister periph driver now.
MFC after: 2 weeks
It seems like kern_data_resid was never really implemented. This change
finally does it. Now frontends update this field while transferring data,
while CTL/backends getting it can more flexibly handle the result.
At this point behavior should not change significantly, still reporting
errors on write overrun, but that may be changed later, if we decide so.
CAM target frontend still does not properly handle overruns due to CAM API
limitations. We may need to add some fields to struct ccb_accept_tio to
pass information about initiator requested transfer size(s).
MFC after: 2 weeks
observed to fix any actual error, but it's the right thing to do
from the correctness point of view.
Tested by: Eugene M. Zheganin <emz at norma.perm.ru>
MFC after: 1 month
Decouple the send and receive limits on the amount of data in a single
iSCSI PDU. MaxRecvDataSegmentLength is declarative, not negotiated, and
is direction-specific so there is no reason for both ends to limit
themselves to the same min(initiator, target) value in both directions.
Allow iSCSI drivers to report their send, receive, first burst, and max
burst limits explicitly instead of using hardcoded values or trying to
derive all of them from the receive limit (which was the only limit
reported by the drivers prior to this change).
Display the send and receive limits separately in the userspace iSCSI
utilities.
Reviewed by: jpaetzel@ (earlier version), trasz@
Sponsored by: Chelsio Communications
Differential Revision: https://reviews.freebsd.org/D7279
CTL HA functionality was originally implemented by Copan many years ago,
but large part of the sources was never published. This change includes
clean room implementation of the missing code and fixes for many bugs.
This code supports dual-node HA with ALUA in four modes:
- Active/Unavailable without interlink between nodes;
- Active/Standby with second node handling only basic LUN discovery and
reservation, synchronizing with the first node through the interlink;
- Active/Active with both nodes processing commands and accessing the
backing storage, synchronizing with the first node through the interlink;
- Active/Active with second node working as proxy, transfering all
commands to the first node for execution through the interlink.
Unlike original Copan's implementation, depending on specific hardware,
this code uses simple custom TCP-based protocol for interlink. It has
no authentication, so it should never be enabled on public interfaces.
The code may still need some polishing, but generally it is functional.
Relnotes: yes
Sponsored by: iXsystems, Inc.
Its idea was to be a simple initiator and execute several commands from
kernel level, but FreeBSD never had consumer for that functionality,
while its implementation polluted many unrelated places..
- remove last remnants of never implemented multiple targets support;
- implement missing support for LUN mapping in this area.
Due to existing locking constraints LUN mapping code is practically
unlocked at this point. Hopefully it is not racy enough to live until
somebody get idea how to call sleeping fronend methods under lock also
taken by the same frontend in non-sleepable context. :(
target iSCSI offload. Add mechanism to query maximum receive data segment
size supported by chosen hardware offload module, and use it in ctld(8)
to determine the value to advertise to the other side.
MFC after: 1 month
Sponsored by: The FreeBSD Foundation
While ctld(8) still does not allow multiple portal groups per target
to be configured, kernel should now be able to handle it.
MFC after: 2 weeks
Sponsored by: iXsystems, Inc.
Replace iSCSI-specific LUN mapping mechanism with new one, working for any
ports. By default all ports are created without LUN mapping, exposing all
CTL LUNs as before. But, if needed, LUN mapping can be manually set on
per-port basis via ctladm. For its iSCSI ports ctld does it via ioctl(2).
The next step will be to teach ctld to work with FibreChannel ports also.
Respecting additional flexibility of the new mechanism, ctl.conf now allows
alternative syntax for LUN definition. LUNs can now be defined in global
context, and then referenced from targets by unique name, as needed. It
allows same LUN to be exposed several times via multiple targets.
While there, increase limit for LUNs per target in ctld from 256 to 1024.
Some initiators do not support LUNs above 255, but that is not our problem.
Discussed with: trasz
MFC after: 2 weeks
Relnotes: yes
Sponsored by: iXsystems, Inc.
While we don't support MCS, hole in received sequence numbers may mean
only PDU loss. While we don't support lost PDU recovery, terminate the
connection to avoid stuck commands.
While there, improve handling of sequence numbers wrap after 2^32 PDUs.
MFC after: 2 weeks
Make CTL core and block backend set success status before initiating last
data move for read commands. Make CAM target and iSCSI frontends detect
such condition and send command status together with data. New I/O flag
allows to skip duplicate status sending on later fe_done() call.
For Fibre Channel this change saves one of three interrupts per read command,
increasing performance from 126K to 160K IOPS. For iSCSI this change saves
one of three PDUs per read command, increasing performance from 1M to 1.2M
IOPS.
MFC after: 1 month
Sponsored by: iXsystems, Inc.
Old allocator created significant lock congestion protecting its lists
of preallocated I/Os, while UMA provides much better SMP scalability.
The downside of UMA is lack of reliable preallocation, that could guarantee
successful allocation in non-sleepable environments. But careful code
review shown, that only CAM target frontend really has that requirement.
Fix that making that frontend preallocate and statically bind CTL I/O for
every ATIO/INOT it preallocates any way. That allows to avoid allocations
in hot I/O path. Other frontends either may sleep in allocation context
or can properly handle allocation errors.
On 40-core server with 6 ZVOL-backed LUNs and 7 iSCSI client connections
this change increases peak performance from ~700K to >1M IOPS! Yay! :)
MFC after: 1 month
Sponsored by: iXsystems, Inc.
