When using iSCSI PDU offload (cxgbei) on T6 adapters, a burst of
received PDUs can be reported via a single message to the driver.
Previously the driver passed these multi-PDU bursts up to the iSCSI
stack up as a single "large" PDU by rewriting the buffer offset, data
segment length, and DataSN fields in the iSCSI header. The DataSN
field in particular was rewritten so that each of the "large" PDUs
used consecutively increasing values. While this worked, the forged
DataSN values did not match the ExpDataSN value in the subsequent SCSI
Response PDU. The initiator does not currently verify this value, but
the forged DataSN values prevent adding a check.
To avoid this, allow a logical iSCSI PDU (struct icl_pdu) to describe
a burst of PDUs via a new 'ip_additional_pdus' field. Normally this
field is set to zero when 'struct icl_pdu' represents a single PDU.
If logical PDU represents a burst of on-the-wire PDUs, then 'ip_npdus'
contains the count of additional on-the-wire PDUs. The header of this
"large" PDU is still modified, but the DataSN field now contains the
DataSN value of the first on-the-wire PDU in the burst.
Reviewed by: mav
Sponsored by: Chelsio Communications
Differential Revision: https://reviews.freebsd.org/D31577
Similar to TSO, iSCSI segmentation offload permits the upper layers to
submit a "large" virtual PDU which is split up into multiple segments
(PDUs) on the wire. Similar to how the TCP/IP headers are used as
templates for TSO, the BHS at the start of a large PDU is used as a
template to construct the specific BHS at the start of each PDU. In
particular, the DataSN is incremented for each subsequent PDU, and the
'F' flag is only set on the last PDU.
struct icl_conn has a new 'ic_hw_isomax' field which defaults to 0,
but can be set to the largest virtual PDU a backend supports. If this
value is non-zero, the iSCSI target and initiator use this size
instead of 'ic_max_send_data_segment_length' to determine the maximum
size for SCSI Data-In and SCSI Data-Out PDUs. Note that since PDUs
can be constructed from multiple buffers before being dispatched, the
target and initiator must wait for the PDU to be fully constructed
before determining the number of DataSN values were consumed (and thus
updating the per-transfer DataSN value used for the start of the next
PDU).
The target generates large PDUs for SCSI Data-In PDUs in
cfiscsi_datamove_in(). The initiator generates large PDUs for SCSI
Data-Out PDUs generated in response to an R2T.
Reviewed by: mav
Sponsored by: Chelsio Communications
Differential Revision: https://reviews.freebsd.org/D31222
and remove repetetive code that calculates vnode locking type for write.
Reviewed by: khng, markj
Sponsored by: The FreeBSD Foundation
MFC after: 1 week
Differential revision: https://reviews.freebsd.org/D31405
Note that, as pointed out by scottl@, this code should really look
a bit different, in that the stack allocations should be replaced
with dynamic allocation, and the periph creation should be moved
to a context where one can use M_WAITOK. See the review for more
details. For now let's go with a minimal fix until we're done with
UMA CCBs.
Reviewed By: mav, imp
Sponsored by: NetApp, Inc.
Sponsored by: Klara, Inc.
Differential Revision: https://reviews.freebsd.org/D30298
This fixes a few bugs in iSCSI backends where the backends were using
the limits they advertised initially during the login phase as the
final values instead of the values negotiated with the other end.
Reported by: Jithesh Arakkan @ Chelsio
Reviewed by: mav
Differential Revision: https://reviews.freebsd.org/D30271
cxgbei stores state about a target transfer in the ctl_private[] array
of a ctl_io that is freed when a target transfer (represented by the
cdw) is freed. As such, freeing a ctl_io before a cdw that references
it can result in a use after free in cxgbei. Two of the four places
freed the cdw first, and the other two freed the ctl_io first. Fix
the latter two places to free the cdw first.
Reported by: Jithesh Arakkan @ Chelsio
Reviewed by: mav
Differential Revision: https://reviews.freebsd.org/D30270
With new ZFS prefetcher improvements it is no longer needed to fully
serialize reads to reach decent prediction hit rate. Softer variant
only creates small time window to reduce races instead of completely
blocking following reads while previous is running. It much less
hurts the performance in case of prediction miss.
MFC after: 1 month
Return while there are any I/Os in a queue may result in them stuck
indefinitely, since there is only one taskqueue task for all of them.
I think I've reproduced this by switching ha_role to secondary under
heavy load.
MFC after: 3 days
- Move ctl_get_cmd_entry() calls from every OOA traversal to when
the requests first inserted, storing seridx in struct ctl_scsiio.
- Move some checks out of the loop in ctl_check_ooa().
- Replace checks for errors that can not happen with asserts.
- Transpose ctl_serialize_table, so that any OOA traversal accessed
only one row (cache line). Compact it from enum to uint8_t.
- Optimize static branch predictions in hottest places.
Due to O(n) nature on deep LUN queues this can be the hottest code
path in CTL, and additional 20% of IOPS I see in some 4KB I/O tests
are good to have in reserve. About 50% of CPU time here according
to the profiles is now spent in two memory accesses per traversed
request in OOA.
Sponsored by: iXsystems, Inc.
MFC after: 2 weeks
This makes random read benchmarks look better on a wide ZFS pools.
I am not sure where the original value goes from, but it is there
for too long now.
MFC after: 1 week
- Make frontends call unified CTL core method ctl_datamove_done()
to report move completion. It allows to reduce code duplication
in differerent backends by accounting DMA time in common code.
- Add to ctl_datamove_done() and be_move_done() callback samethr
argument, reporting whether the callback is called in the same
context as ctl_datamove(). It allows for some cases like iSCSI
write with immediate data or camsim frontend write save one context
switch, since we know that the context is sleepable.
- Remove data_move_done() methods from struct ctl_backend_driver,
unused since forever.
MFC after: 1 month
Switch OOA queue from TAILQ to LIST and change its direction, so that
we traverse it forward, not backward. There is only one place where
we really need other direction, and it is not critical.
Use STAILQ_REMOVE_HEAD() instead of STAILQ_REMOVE() in backends.
Replace few impossible conditions with assertions.
MFC after: 1 month
Introduce new CTL core KPI ctl_run(), preprocessing I/Os in the caller
context instead of scheduling another thread just for that. This call
may sleep, that is not acceptable for some frontends like the original
CAM/FC one, but iSCSI already has separate sleepable per-connection RX
threads, and another thread scheduling is mostly just a waste of time.
IOCTL frontend actually waits for the I/O completion in the caller
thread, so the use of another thread for this has even less sense.
With this change I can measure ~5% IOPS improvement on 4KB iSCSI I/Os
to ZFS.
MFC after: 1 month
RFC 7143 (11.7.4):
The Target Transfer Tag values are not specified by this protocol,
except that the value 0xffffffff is reserved and means that the
Target Transfer Tag is not supplied.
MFC after: 1 month
Replace MAXPHYS by runtime variable maxphys. It is initialized from
MAXPHYS by default, but can be also adjusted with the tunable kern.maxphys.
Make b_pages[] array in struct buf flexible. Size b_pages[] for buffer
cache buffers exactly to atop(maxbcachebuf) (currently it is sized to
atop(MAXPHYS)), and b_pages[] for pbufs is sized to atop(maxphys) + 1.
The +1 for pbufs allow several pbuf consumers, among them vmapbuf(),
to use unaligned buffers still sized to maxphys, esp. when such
buffers come from userspace (*). Overall, we save significant amount
of otherwise wasted memory in b_pages[] for buffer cache buffers,
while bumping MAXPHYS to desired high value.
Eliminate all direct uses of the MAXPHYS constant in kernel and driver
sources, except a place which initialize maxphys. Some random (and
arguably weird) uses of MAXPHYS, e.g. in linuxolator, are converted
straight. Some drivers, which use MAXPHYS to size embeded structures,
get private MAXPHYS-like constant; their convertion is out of scope
for this work.
Changes to cam/, dev/ahci, dev/ata, dev/mpr, dev/mpt, dev/mvs,
dev/siis, where either submitted by, or based on changes by mav.
Suggested by: mav (*)
Reviewed by: imp, mav, imp, mckusick, scottl (intermediate versions)
Tested by: pho
Sponsored by: The FreeBSD Foundation
Differential revision: https://reviews.freebsd.org/D27225
Before this CTL always allocated MAXPHYS-sized buffers, even for 4KB I/O,
that is even more overkill for MAXPHYS of 1MB. This change limits maximum
allocation to 512KB if MAXPHYS is bigger, plus if one is above 128KB, adds
new 128KB UMA zone for smaller I/Os. The patch factors out alloc/free,
so later we could make it use more zones or malloc() if we'd like.
MFC after: 1 week
Sponsored by: iXsystems, Inc.
Foundation copyrights, approved by emaste@. It does not include
files which carry other people's copyrights; if you're one
of those people, feel free to make similar change.
Reviewed by: emaste, imp, gbe (manpages)
Differential Revision: https://reviews.freebsd.org/D26980
SAM-3 specification introduced concept of Task Priority, that was renamed
to Command Priority in SAM-4, and supported by all modern SCSI transports.
It provides 15 levels of relative priorities: 1 - highest, 15 - lowest and
0 - default. SAT specification for SATA devices translates priorities 1-3
into NCQ high priority.
This change adds new "priority" field into empty spots of struct ccb_scsiio
and struct ccb_accept_tio of CAM and struct ctl_scsiio of CTL. Respective
support is added into iscsi(4), isp(4), mpr(4), mps(4) and ocs_fc(4) drivers
for both initiator and where applicable target roles. Minimal support was
added to CTL to receive the priority value from different frontends, pass it
between HA controllers and report in few places.
This patch does not add consumers of this functionality, so nothing should
really change yet, since the field is still set to 0 (default) on initiator
and not actively used on target. Those are to be implemented separately.
I've confirmed priority working on WD Red SATA disks connected via mpr(4)
and properly transferred to CTL target via iscsi(4), isp(4) and ocs_fc(4).
While there, added missing tag_action support to ocs_fc(4) initiator role.
MFC after: 1 month
Relnotes: yes
Sponsored by: iXsystems, Inc.
It allows to report to initiator LU identifying information, preset via
"ident_info" and "text_ident_info" options.
Unfortunately it is impossible to implement SET IDENTIFYING INFORMATION,
since we have no persistent storage it requires, so the information is
read-only for initiator and has to be set out-of-band.
MFC after: 1 week
Sponsored by: iXsystems, Inc.
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.
ha_lso is a listening socket (unless bind() has failed), so should use
solisten_upcall_set(NULL, NULL), not soupcall_clear().
MFC after: 1 week
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
- Make ctl_add_lun() synchronous. Asynchronous addition was used by
Copan's proprietary code long ago and never for upstream FreeBSD.
- Move LUN enable/disable calls from backends to CTL core.
- Serialize LUN modification and partially removal to avoid double frees.
- Slightly unify backends code.
MFC after: 2 weeks
Sponsored by: iXsystems, Inc.
These flags have been unused for some time. Some of them were in the
CAM2 specification, but CAM has moved on a bit from that. Some were
used in the old Pluto VideoSpace (and AirSpace) systems which had the
video playback I/O scheduler in userspace, but have been unused since
then.
Reviewed by: chuck, ken
Differential Revision: https://reviews.freebsd.org/D24008
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
Filesystems which want to use it in limited capacity can employ the
VOP_UNLOCK_FLAGS macro.
Reviewed by: kib (previous version)
Differential Revision: https://reviews.freebsd.org/D21427
Combined with earlier nstart/nend removal it allows to remove several locks
from request path of GEOM and few other places. It would be cool if we had
more SMP-friendly statistics, but this helps too.
Sponsored by: iXsystems, Inc.
The current vnode layout is not smp-friendly by having frequently read data
avoidably sharing cachelines with very frequently modified fields. In
particular v_iflag inspected for VI_DOOMED can be found in the same line with
v_usecount. Instead make it available in the same cacheline as the v_op, v_data
and v_type which all get read all the time.
v_type is avoidably 4 bytes while the necessary data will easily fit in 1.
Shrinking it frees up 3 bytes, 2 of which get used here to introduce a new
flag field with a new value: VIRF_DOOMED.
Reviewed by: kib, jeff
Differential Revision: https://reviews.freebsd.org/D22715
