Tyrel Datwyler <tyreld@linux.ibm.com> says:
This series adds NVMe/FC initiator support to the ibmvfc driver, enabling
IBM POWER virtual machines to discover and use NVMe namespaces presented by
the IBM Virtual I/O Server (VIOS) over the existing NPIV transport.
The ibmvfc driver communicates with the VIOS via a CRQ-based protocol. With
this series the VIOS can present both SCSI/FCP and NVMe/FC targets through
parallel sets of protocol-specific MAD opcodes, fabric login flows, and
sub-CRQ channels.
The series is organized into three phases:
Patches 1-5: Bug fixes and preparatory refactoring
Four pre-existing bugs are fixed before any NVMe/FC work is introduced:
a deadlock in the MAD send-failure path (locked done variant called with
host_lock already held), a race during driver teardown where
rport_add_work_q work items can outlive the FC host, a NULL event
dereference in ibmvfc_tgt_implicit_logout_and_del, and an allocator
mismatch where mempool-allocated ibmvfc_target structs are freed via kfree
rather than mempool_free. Patch 5 moves the target list and count from
struct ibmvfc_host into struct ibmvfc_channels as the structural
prerequisite for independent per-protocol target tracking.
Patches 6-27: Protocol interface and driver scaffolding
Patch 6 extends ibmvfc.h with NVMe/FC protocol definitions: MAD opcodes,
capability flags, the v3 command layout, the fabric login MAD, async
sub-CRQ event format, and updated channel enquiry/setup fields. Patch 7
splits ibmvfc.c into ibmvfc-core.c and the new ibmvfc-nvme.c/h, registers
an nvme_fc_port_template with stub callbacks, and adds NVMe module
parameters. The remaining patches build the NVMe/FC plumbing: NVMe
channel-group initialization, sub-CRQ lifecycle management, protocol-
specific fabric login flow, target discovery, PLOGI/PRLI/query-target,
implicit logout, move-login, protocol-driven target allocation, NVMe
target deletion, state machine updates, and local/remote port registration
with the NVMe-FC transport layer.
Patches 28-33: NVMe-FC LLDD callbacks and I/O path
Implements the full nvme_fc_port_template: create_queue/delete_queue map
NVMe controller queues to sub-CRQ handles; ls_req submits FC-LS frames via
the ibmvfc passthru MAD; fcp_io builds and submits NVMe FCP commands via
the NVMe sub-CRQ and completes them via nvme_fc_rcv_fcp_rsp(); ls_abort
and fcp_abort cancel outstanding requests via NVMF cancel MADs. Patch 33
extends the purge path to fail outstanding NVMe FCP and LS requests during
host reset and link-down events.
Link: https://patch.msgid.link/20260723000149.969416-1-tyreld@linux.ibm.com
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
The current purge code for flushing outstanding commands during a
transport reset only deals with SCSI commands. Rename the
ibmvfc_scsi_eh_done completion handler to ibmvfc_vfc_eh_done and wire it
to correctly complete nvme fcp and ls commands when flushing the
inflight command list during a reset.
Signed-off-by: Tyrel Datwyler <tyreld@linux.ibm.com>
Link: https://patch.msgid.link/20260723000149.969416-34-tyreld@linux.ibm.com
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
Implement the NVMe-FC FCP abort callback by issuing an NVMF cancel MAD
on the same submission queue used by the original FCP request.
Use the original request event stored in abort_req->private to recover
the associated ibmvfc queue, then allocate a new event from that queue
so the cancel is sent on the matching nvme_scrq. Factor the TMF setup
into a dedicated helper, mirroring the LS abort path, and populate the
cancel key, task tag, target WWPN, and association ID needed for the
VIOS NVMF abort request.
The abort path sends the cancel synchronously, waits for completion,
frees the temporary event, and logs non-zero MAD status values for
debugging.
Signed-off-by: Tyrel Datwyler <tyreld@linux.ibm.com>
Link: https://patch.msgid.link/20260723000149.969416-33-tyreld@linux.ibm.com
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
Implement the NVMe FC-LS abort callback by issuing an ibmvfc cancel MAD
to the VIOS for the outstanding link-service request.
Use the saved event pointer from the original FC-LS request to identify
the command to cancel, submit the cancel operation, and complete the
abort request based on the returned status.
Signed-off-by: Tyrel Datwyler <tyreld@linux.ibm.com>
Link: https://patch.msgid.link/20260723000149.969416-32-tyreld@linux.ibm.com
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
Add helpers to initialize an ibmvfc command from an nvmefc_fcp_req,
map request scatterlists into either an inline descriptor or an external
DMA pool list, and submit the request on the selected NVMe hardware
queue. On completion, translate ibmvfc status into the NVMe-FC response
format, including transferred length and CQE handling for no-DMA
responses.
Also store the NVMe request pointer in struct ibmvfc_event so the
completion path can finish the original request.
Signed-off-by: Tyrel Datwyler <tyreld@linux.ibm.com>
Link: https://patch.msgid.link/20260723000149.969416-31-tyreld@linux.ibm.com
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
NVMe FC Link Service commands are required to use the ibmvfc_passthru
MAD. Initialize a pssthru mad for the target port including the DMA
addresses for the FC4_LS request and response as well as the max length
of each IU as provided in the nvmefc_ls_req struct. FC4_LS commands are
sent via the primary CRQ. Further, store the assoc_id during a create
association request as this is a required field in our vfc_cmd struct
for nvme_fcp_io commands.
Signed-off-by: Tyrel Datwyler <tyreld@linux.ibm.com>
Link: https://patch.msgid.link/20260723000149.969416-30-tyreld@linux.ibm.com
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
Implement the NVMe-FC queue create and delete callbacks and map NVMe
controller queues onto ibmvfc hardware queues.
Use qidx of NVMe controller queue to map onto a ibmvfc_queue channel.
The Admin queue is always qidx 0 and general practice among other
drivers is to map both the Admin queue and first IO queue to the same HW
queue. Add a new ibmvfc_nvme_qhandle struct that will be used as the
opaque queue handle by the NVMe-FC layer when issuing fcp IO.
Signed-off-by: Tyrel Datwyler <tyreld@linux.ibm.com>
Link: https://patch.msgid.link/20260723000149.969416-29-tyreld@linux.ibm.com
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
Some common functions will require visibility by both SCSI and NVMe
protocols. Make common ibmvfc helper routines available to the NVMe
support code.
Remove static from the core event allocation, event initialization,
event free, target release, command error, and event send helpers, and
declare them in ibmvfc.h. This allows ibmvfc-nvme.c to reuse the
existing event and target infrastructure.
No functional change is intended.
Signed-off-by: Tyrel Datwyler <tyreld@linux.ibm.com>
Link: https://patch.msgid.link/20260723000149.969416-28-tyreld@linux.ibm.com
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
Export ibmvfc_debug so the NVMe support code can use the
existing ibmvfc_dbg logging macro.
The debug control variable is currently file-local to the core driver,
which prevents protocol-specific code in ibmvfc-nvme.c from using the
shared debug infrastructure. Make the variable global within the module
and declare it in ibmvfc-nvme.h.
Signed-off-by: Tyrel Datwyler <tyreld@linux.ibm.com>
Link: https://patch.msgid.link/20260723000149.969416-27-tyreld@linux.ibm.com
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
Add an NVMe-specific remote-port add helper and update the rport worker
thread to walk the NVMe target list, register new NVMe remote ports, and
rescan existing ones through the NVMe-FC midlayer. Also handle delete
and delete-with-logout transitions for NVMe remote ports in the same
worker context used for SCSI rports.
This keeps remote-port registration serialized in the existing worker
model while allowing NVMe targets to participate in the common target
state machine.
Signed-off-by: Tyrel Datwyler <tyreld@linux.ibm.com>
Link: https://patch.msgid.link/20260723000149.969416-26-tyreld@linux.ibm.com
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
Register the local NVMe/FC port only after fabric login has completed.
The VIOS returns the client port ID in the fabric login response, and
that port ID is required to populate the local-port information passed
to the NVMe-FC midlayer. Delay local-port registration until that data
is available and update the registration helper accordingly.
Signed-off-by: Tyrel Datwyler <tyreld@linux.ibm.com>
Link: https://patch.msgid.link/20260723000149.969416-25-tyreld@linux.ibm.com
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
Implement the initial NVMe/FC local-port and remote-port registration
functions that notify the NVMe-FC midlayr of port discovery and loss.
Register the local port with the NVMe-FC transport, register discovered
remote ports against that local port, and add matching unregister paths
that wait for the NVMe-FC core to complete asynchronous deletion before
dropping driver references. Also store driver-private host and target
pointers in the registered NVMe-FC port objects.
Signed-off-by: Tyrel Datwyler <tyreld@linux.ibm.com>
Link: https://patch.msgid.link/20260723000149.969416-24-tyreld@linux.ibm.com
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
Update the host work loop and target state-machine helpers to process
NVMe targets in addition to SCSI targets.
Check both protocol-specific target lists when determining whether there
is initialization or logout work pending, and extend the query, target
init, and target delete phases to dispatch work for NVMe targets using
the same common state-machine callbacks.
This allows the existing discovery and login state machine to drive
NVMe/FC targets through query, login, and deletion without duplicating
the control flow.
Signed-off-by: Tyrel Datwyler <tyreld@linux.ibm.com>
Link: https://patch.msgid.link/20260723000149.969416-23-tyreld@linux.ibm.com
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
Extend target deletion paths to process NVMe targets as well as SCSI
targets.
Update link-down, and host reinitialization flows to walk both the SCSI
and NVMe target lists when marking targets for deletion. This ensures
that protocol-specific target state stays consistent across adapter
resets and fabric events.
Rename ibmvfc_relogin to ibmvfc_scsi_relogin as it acts on a scsi
command.
Signed-off-by: Tyrel Datwyler <tyreld@linux.ibm.com>
Link: https://patch.msgid.link/20260723000149.969416-22-tyreld@linux.ibm.com
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
Allocate discovered targets onto the channel-group list that matches
their protocol.
When a target is created, use the discovered protocol type to decide
which list it belongs on. This keeps protocol-specific discovery
results isolated and allows later state-machine and remote-port code to
walk the correct target set.
Signed-off-by: Tyrel Datwyler <tyreld@linux.ibm.com>
Link: https://patch.msgid.link/20260723000149.969416-21-tyreld@linux.ibm.com
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
Extend PRLI handling code to support NVMe/FC targets.
When the target protocol is NVMe/FC, issue the NVMe process login MAD,
set the NVMe FC-4 type, and populate NVMe-specific service parameters.
On completion, decode the returned PRLI service parameters and derive
the appropriate remote-port roles for NVMe initiator, target, and
discovery ports.
Keep the existing SCSI PRLI flow unchanged while allowing the common
target state machine to complete login for NVMe/FC targets.
Signed-off-by: Tyrel Datwyler <tyreld@linux.ibm.com>
Link: https://patch.msgid.link/20260723000149.969416-19-tyreld@linux.ibm.com
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
Expand the target login path to issue the NVMe/FC-specific port login
MAD for NVMe targets.
Select the correct PLOGI MAD opcode based on the target protocol and
include the protocol name in success and failure logging. The rest of
the target login flow remains shared with the existing SCSI
implementation.
Signed-off-by: Tyrel Datwyler <tyreld@linux.ibm.com>
Link: https://patch.msgid.link/20260723000149.969416-18-tyreld@linux.ibm.com
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
Add protocol-specific handling for implicit logout and move-login
operations on NVMe/FC targets.
Select the NVMe/FC-specific implicit logout opcode when operating on an
NVMe target and update the associated logging so protocol-specific
operations are visible in debug output. This extends the existing target
relogin and migration-related flows to work with NVMe targets as well as
SCSI targets.
These changes are needed so target reauthentication and target movement
continue to work once NVMe/FC targets are added to the driver's state
machine.
Signed-off-by: Tyrel Datwyler <tyreld@linux.ibm.com>
Link: https://patch.msgid.link/20260723000149.969416-17-tyreld@linux.ibm.com
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
Extend target discovery to send protocol-specific discover-target MADs
for NVMe/FC.
Use the protocol-aware discovery helper to build an NVMe discover-target
request, submit it when NVMe/FC support is active, and process the
returned target count using the NVMe channel group's discovery buffer.
This allows the driver to discover NVMe/FC targets in parallel with the
existing SCSI discovery flow while keeping protocol-specific target data
separate.
Signed-off-by: Tyrel Datwyler <tyreld@linux.ibm.com>
Link: https://patch.msgid.link/20260723000149.969416-16-tyreld@linux.ibm.com
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
Refactor discover-target event creation so it can be shared by both SCSI
and NVMe/FC discovery.
Introduce a helper that takes a protocol-specific channel group, selects
the correct discover-target opcode, and maps the corresponding discovery
buffer into the MAD.
This is a preparatory cleanup for issuing protocol-specific discovery
MADs in later patches.
Signed-off-by: Tyrel Datwyler <tyreld@linux.ibm.com>
Link: https://patch.msgid.link/20260723000149.969416-13-tyreld@linux.ibm.com
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
Add ibmvfc_channels_to_vhost() to recover the parent struct ibmvfc_host
from a protocol-specific struct ibmvfc_channels.
Later patches need to operate on either the SCSI or NVMe channel group
and still access host-wide state such as the primary CRQ, device, and
logging context. Centralize that mapping in a helper instead of open-
coding container lookups at each call site.
Signed-off-by: Tyrel Datwyler <tyreld@linux.ibm.com>
Link: https://patch.msgid.link/20260723000149.969416-12-tyreld@linux.ibm.com
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
Add support for the protocol-specific fabric login flow introduced by
the updated client/VIOS interface.
After NPIV login, a VIOS that advertises protocol-specific support
requires separate fabric login MADs for SCSI and NVMe/FC. Track whether
SCSI and NVMe/FC fabric login are needed, extend channel enquiry/setup
handling to negotiate both SCSI and NVMe queue counts, and issue the
appropriate fabric login MADs before target discovery begins.
Also update command layout selection so the driver uses the v3 command
format when the VIOS advertises NVMe/FC-capable framing.
Signed-off-by: Tyrel Datwyler <tyreld@linux.ibm.com>
Link: https://patch.msgid.link/20260723000149.969416-11-tyreld@linux.ibm.com
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
Allocate, register, deregister, and release NVMe subordinate CRQs
alongside the existing SCSI sub-CRQs.
Update the CRQ reset and re-enable paths to tear down and recreate NVMe
sub-queues, extend sub-CRQ initialization to allocate NVMe channels when
enabled, and release NVMe channel resources during adapter teardown.
This keeps the NVMe queue lifecycle aligned with the existing SCSI queue
lifecycle so both protocols are reset consistently across probe, remove,
and connection recovery.
Signed-off-by: Tyrel Datwyler <tyreld@linux.ibm.com>
Link: https://patch.msgid.link/20260723000149.969416-10-tyreld@linux.ibm.com
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
Initialize the host's NVMe channel-group state during probe.
Set up the NVMe channel list head, desired queue count, maximum queue
count, protocol identifier, and enablement state alongside the existing
SCSI channel-group initialization in ibmvfc_probe().
This prepares the driver with a NVMe/FC channel group that can will be
used by later patches for NVMe queue allocation, discovery buffers,
target management, and IO submission.
Signed-off-by: Tyrel Datwyler <tyreld@linux.ibm.com>
Link: https://patch.msgid.link/20260723000149.969416-9-tyreld@linux.ibm.com
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
Rename ibmvfc.c to ibmvfc-core.c as first step in decoupling each
protocol from the core driver logic. Add ibmvfc-nvme.[ch] files, and
register an nvme_fc_port_template with empty callback stubs.
Add empty registration functions definitions for local and remote ports.
No functional NVMe/FC support is added yet.
Signed-off-by: Tyrel Datwyler <tyreld@linux.ibm.com>
Link: https://patch.msgid.link/20260723000149.969416-8-tyreld@linux.ibm.com
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
Add the protocol definitions for client-VIOS interface updates needed to
support NVMe/FC over the ibmvfc NPIV transport.
Extend the ibmvfc interface with:
- NVMe/FC-specific capability bits and opcodes
- protocol-specific channel and queue definitions
- updated channel enquiry/setup fields for NVMe queues
- v3 command layout support for protocol-specific payloads
These changes provide the common header and interface plumbing needed by
later patches that add NVMe/FC login, discovery, remote-port handling,
and I/O submission.
Signed-off-by: Tyrel Datwyler <tyreld@linux.ibm.com>
Link: https://patch.msgid.link/20260723000149.969416-7-tyreld@linux.ibm.com
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
Prepare the driver for protocol-specific target management by moving
the target list and target count out of struct ibmvfc_host and into
struct ibmvfc_channels.
Today the driver only maintains a single SCSI target list, but NVMe/FC
support will require separate target tracking for each protocol-specific
channel group. Update the existing target iteration, allocation, and
discovery paths to use the SCSI channel group's target list instead of a
host-wide list.
This is a preparatory refactoring only. No functional change is intended
for existing SCSI operation.
Signed-off-by: Tyrel Datwyler <tyreld@linux.ibm.com>
Link: https://patch.msgid.link/20260723000149.969416-6-tyreld@linux.ibm.com
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
The ibmvfc_target *tgt structures for discovery are allocated using
mempool_alloc but in turn free'd via kfree in ibmvfc_release_target
breaking pool guarentees.
Fixup ibmvfc_release_target to call mempool_free instead so that memory
is properly returned to the mempool instead of the general allocator.
Signed-off-by: Tyrel Datwyler <tyreld@linux.ibm.com>
Link: https://patch.msgid.link/20260723000149.969416-5-tyreld@linux.ibm.com
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
Unlike ibmvfc_tgt_implicit_logout the ibmvfc_tgt_implicit_logout_and_del
path fails to check for a NULL evt returned from
ibmvfc_get_reserved_event and passes that directly to ibmvfc_send_event
which will unconditionally derefence the evt.
Check for a NULL evt and perform same escalation as
ibmvfc_implicit_logout.
Signed-off-by: Tyrel Datwyler <tyreld@linux.ibm.com>
Link: https://patch.msgid.link/20260723000149.969416-4-tyreld@linux.ibm.com
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
INIT_WORK is called on rport_add_work_q during driver probe. When the
adapter is removed, ibmvfc_remove() stops the work thread and proceeds
to tear down the FC host, but any work item queued via schedule_work()
that raced the teardown could still be running or pending in the
system workqueue.
Add a flush_work() call after kthread_stop() and before fc_remove_host()
to drain any outstanding rport add work before the FC host and its rports
are torn down.
Signed-off-by: Tyrel Datwyler <tyreld@linux.ibm.com>
Link: https://patch.msgid.link/20260723000149.969416-3-tyreld@linux.ibm.com
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
MADs are processed with the requirement that the host_lock is held for
MAD sends and completions. When a MAD job step is called the host lock
is already held and there is an unlikely scenario where h_send_crq fails
with something other than H_CLOSED. If this happens there is a special
done handler for MADs that tries to take the host lock.
In this failure scenario fall back to the saved evt->_done unlocked
completion handler.
Signed-off-by: Tyrel Datwyler <tyreld@linux.ibm.com>
Link: https://patch.msgid.link/20260723000149.969416-2-tyreld@linux.ibm.com
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
The variable 'fact' was used for debug but this was removed in commit
62b434b0db ("scsi: dc395x: Remove DEBUG conditional compilation"). The
variable is now redundant and can be removed.
Cleans up clang scan build warning:
drivers/scsi/dc395x.c: In function ‘msgin_set_sync’:
drivers/scsi/dc395x.c:2185:13: warning: variable ‘fact’ set but not used [-Wunused-but-set-variable]
2185 | int fact;
Signed-off-by: Colin Ian King <colin.i.king@gmail.com>
Link: https://patch.msgid.link/20260714073510.43289-1-colin.i.king@gmail.com
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
Louis Sautier <sautier.louis@gmail.com> says:
Expose the IOC and board temperature sensors of LSI / Broadcom / Avago
SAS HBAs that bind to mpt3sas through the hwmon interface. The data
lives in MPI IO Unit Page 7.
The same fields are exposed by Broadcom's userspace tooling through
the /dev/mpt[23]ctl ioctl path (typically root-only): IOCTemperature
and BoardTemperature in lsiutil; ROC and Controller in storcli.
With this driver, sensors(1) shows them unprivileged:
$ sensors mpt3sas-pci-0200
mpt3sas-pci-0200
Adapter: PCI adapter
IOC: +42.0°C
Testing
-------
Build-tested all four SCSI_MPT3SAS x HWMON combinations (=y and =m
each), including the SCSI_MPT3SAS=y with HWMON=m case that requires
IS_REACHABLE().
Validated across three Broadcom SAS chip generations. None of the
cards had a board sensor present, so the testing only covers the
IOC channel:
* LSI 9500-8i / SAS3816, SAS-3:
- hwmon device registers as "mpt3sas" with only temp1 (IOC) exposed
- IOC reading matches `storcli /c0 show temperature` and
`lsiutil -p1 -a 25,2,0,0`
- rmmod / modprobe cycle goes through the explicit
unregister/register paths cleanly
* LSI 9305-24i / SAS3224, SAS-3: same behaviour.
* LSI 9211-4i / SAS2004, SAS-2: firmware reports both
*TemperatureUnits = NOT_PRESENT, no hwmon device registered
(graceful-skip path).
Not verified (no available hardware):
* Path with both IOC and board sensors present.
* Fahrenheit-units conversion.
* Sub-zero readings (signed-cast path).
Link: https://patch.msgid.link/20260630224922.2543096-1-sautier.louis@gmail.com
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
The PCI_INTERRUPT_PIN reports if the device supports the INTx. However,
when the device is assigned to a guest via vfio, the PCI_INTERRUPT_PIN
is set to 0 (i.e none) if the line is not connected and/or the platform
cannot route the interrupt.
In such cases, the guest PCI_INTERRUPT_PIN is 0 and the port number
becomes -1 (255, uint8_t underflow) for qla[25|27|28]xx and qla2031
devices. The flt_region_nvram is never set, and subsequently the LUN
detection fails. Below warnings show the NVRAM configuration failure:
[]-0073:1: Inconsistent NVRAM checksum=0xffffffc0 id=HCAM version=0x100.
[]-0074:1: Falling back to functioning (yet invalid -- WWPN) defaults.
[]-0076:1: NVRAM configuration failed.
Handle this case and set the port_no to devfn like its done everywhere
else.
Reference: commit 2bd42b03ab ("vfio/pci: Virtualize zero INTx PIN if no pdev->irq")
Signed-off-by: Shivaprasad G Bhat <sbhat@linux.ibm.com>
Reviewed-by: Kyle Mahlkuch <kmahlkuc@linux.ibm.com>
Link: https://patch.msgid.link/177885270578.1573.14283751510936407585.stgit@linux.ibm.com
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
Once the UFS WriteBooster fails to allocate memory, the situation will
stay until fstrim or equivalent is ran.
Mark is as a warning since it impacts the performance but only print it
once for the lifetime of the kernel since it's not fatal.
Otherwise it will be printed each time the device is resumed:
[ 31.666880] ufshcd-qcom 1d84000.ufshc: dCurWBBuf: 0 WB disabled until free-space is available
[ 52.655594] ufshcd-qcom 1d84000.ufshc: dCurWBBuf: 0 WB disabled until free-space is available
[ 62.890469] ufshcd-qcom 1d84000.ufshc: dCurWBBuf: 0 WB disabled until free-space is available
...
Signed-off-by: Neil Armstrong <neil.armstrong@linaro.org>
Reviewed-by: Bart Van Assche <bvanassche@acm.org>
Link: https://patch.msgid.link/20260618-topic-ufs-wb-empty-warn-v1-1-ec744a153e0e@linaro.org
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
ed.tsai@mediatek.com says:
The first patch adds the get_hba_nortt() callback to the UFS core
layer, allowing vendor drivers to provide dynamic, platform-specific
RTT capability handling.
The second patch implements this callback in the MediaTek UFS driver,
distinguishing between legacy platforms (which require the RTT to be
limited to 2) and newer MT6995 B0+ platforms (which can use the value
from the capability register directly).
The third patch removes the max_num_rtt field from ufs_hba_variant_ops
as it is now replaced by the get_hba_nortt() callback.
Link: https://patch.msgid.link/20260615055802.105479-1-ed.tsai@mediatek.com
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
Can Guo <can.guo@oss.qualcomm.com> says:
TX Equalization training currently has a few error-path gaps that can
make the flow brittle and can leave variant/device cleanup incomplete.
This series hardens TX EQTR in three places:
1. ufs-qcom: route SW FOM setup failures through the shared cleanup
path so temporary device TX EQ settings are restored and link
recovery is always attempted before exit.
2. core: treat RX_FOM DME read failures as best effort so TX EQTR can
continue, and force failed lanes to deterministic 0 FOM.
3. core: always run tx_eqtr POST_CHANGE notify once PRE_CHANGE
succeeds, even when TX EQTR fails, so variant cleanup is not
skipped.
Together these changes improve TX EQTR robustness without changing the
normal success path.
Link: https://patch.msgid.link/20260625121306.1655467-1-can.guo@oss.qualcomm.com
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
ufshcd_tx_eqtr() skips POST_CHANGE notify when __ufshcd_tx_eqtr()
fails. That can leave variant cleanup incomplete when PRE_CHANGE saved
temporary state that POST_CHANGE is expected to restore.
Always call POST_CHANGE once PRE_CHANGE has succeeded. Keep the TX EQTR
result as the primary return value, and only propagate POST_CHANGE failure
when TX EQTR itself succeeded.
Log PRE_CHANGE and POST_CHANGE notify failures to make variant callback
failures visible in TX EQTR error paths.
Reviewed-by: Manivannan Sadhasivam <mani@kernel.org>
Reviewed-by: Peter Wang <peter.wang@mediatek.com>
Signed-off-by: Can Guo <can.guo@oss.qualcomm.com>
Reviewed-by: Bean Huo <beanhuo@micron.com>
Reviewed-by: Bart Van Assche <bvanassche@acm.org>
Reviewed-by: Ziqi Chen <ziqi.chen@oss.qualcomm.com>
Link: https://patch.msgid.link/20260625121306.1655467-4-can.guo@oss.qualcomm.com
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
ufshcd_get_rx_fom() aborted TX EQTR when a per-lane RX_FOM DME read failed.
That makes the whole training flow fragile even though these reads can be
treated as best effort.
Keep TX EQTR running by logging RX_FOM read failures and continuing. Make
failed lanes deterministic by initializing each lane FOM to 0 before
reading and only updating it when the DME read succeeds. This avoids
propagating stale or uninitialized values into EQTR evaluation.
Also update the kerneldoc return description to match behavior: RX_FOM DME
read failures are handled as warnings, while get_rx_fom() vops failures are
still propagated to the caller.
Signed-off-by: Can Guo <can.guo@oss.qualcomm.com>
Reviewed-by: Peter Wang <peter.wang@mediatek.com>
Reviewed-by: Bean Huo <beanhuo@micron.com>
Reviewed-by: Ziqi Chen <ziqi.chen@oss.qualcomm.com>
Link: https://patch.msgid.link/20260625121306.1655467-3-can.guo@oss.qualcomm.com
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
ufs_qcom_get_rx_fom() applies temporary device TX Equalization values
before forcing HS mode and running the EOM-based SW FOM scan.
When one of these steps fails, the function can bypass the shared cleanup
path and leave temporary TX Equalization settings programmed.
Route those failures through the cleanup label so the original TX EQ
settings are restored and link recovery runs before exit.
This path also reuses ret for cleanup, so it may overwrite the original
error. Keep that on purpose: if cleanup succeeds, the caller can proceed
with the FOM result for the current iteration.
Signed-off-by: Can Guo <can.guo@oss.qualcomm.com>
Reviewed-by: Bean Huo <beanhuo@micron.com>
Reviewed-by: Ziqi Chen <ziqi.chen@oss.qualcomm.com>
Link: https://patch.msgid.link/20260625121306.1655467-2-can.guo@oss.qualcomm.com
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
TX EQTR may run while devfreq gear scaling has quiesced the UFS
tagset. In that context, functions ufshcd_tx_eqtr(), __ufshcd_tx_eqtr()
and ufs_qcom_get_rx_fom() allocate memory with GFP_KERNEL. If direct
reclaim is triggered, reclaim/writeback can depend on I/O to UFS
device. Because the queue is quiesced, this can cause deadlock.
Use memalloc_noio_save/restore() in ufshcd_tx_eqtr() to cover all
allocations in the TX EQTR call tree, including:
- params->eqtr_record in ufshcd_tx_eqtr()
- eqtr_data in __ufshcd_tx_eqtr()
- params in ufs_qcom_get_rx_fom()
This is preferred over tagging individual call sites with GFP_NOIO, as it
automatically covers any future allocations added anywhere in the call tree
without requiring each caller to be aware of this constraint.
[mkp: fix label as suggested by Bart]
Fixes: 03e5d38e2f ("scsi: ufs: core: Add support for TX Equalization")
Closes: https://sashiko.dev/#/patchset/20260615132834.2985346-1-can.guo@oss.qualcomm.com?part=2
Signed-off-by: Can Guo <can.guo@oss.qualcomm.com>
Reviewed-by: Ziqi Chen <ziqi.chen@oss.qualcomm.com>
Reviewed-by: Manivannan Sadhasivam <mani@kernel.org>
Link: https://patch.msgid.link/20260618140941.902000-1-can.guo@oss.qualcomm.com
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
Can Guo <can.guo@oss.qualcomm.com> says:
Hi,
This series adds support for board-specific static TX Equalization settings
provided through Device Tree.
This series is based on the earlier TX Equalization enablement work and
persistent storage/retrieval of optimal TX Equalization settings work:
https://lore.kernel.org/all/20260325152154.1604082-1-can.guo@oss.qualcomm.comhttps://lore.kernel.org/all/20260424151420.111675-1-can.guo@oss.qualcomm.com
Background
==========
UFS v5.0/UFSHCI v5.0 adds HS-G6 support (46.6 Gbps/lane) via UniPro v3.0
and M-PHY v6.0. These specs define TX Equalization for all High-Speed
Gears (not only HS-G6) to compensate channel loss and improve signal
integrity at high speed.
For HS-G6, M-PHY uses PAM4 1b1b line coding. Pre-Coding may also be
required depending on channel characteristics.
This series adds vendor-neutral DT properties:
- patternProperties: txeq-preshoot-g[1-6], txeq-deemphasis-g[1-6]
- fixed property: tx-precode-enable-g6
All properties use per-lane Host/Device tuples and accept 2 or 4 values
for x1/x2 lane configurations:
- txeq-preshoot-g[1-6]: values 0..7
- txeq-deemphasis-g[1-6]: values 0..7
- tx-precode-enable-g6: values 0/1
These properties carry board-level SI characterization data used as static
TX Equalization settings for each High-Speed Gear.
Example DTS snippet
===================
The following x2-lane example shows the expected DT encoding:
ufs@1d84000 {
lanes-per-direction = <2>;
txeq-preshoot-g6 = <1 2>, <3 4>;
txeq-deemphasis-g6 = <0 1>, <2 3>;
tx-precode-enable-g6 = <1 0>, <0 1>;
};
Relationship with Adaptive TX Equalization
==========================================
Adaptive TX Equalization remains the primary path when enabled.
Static TX Equalization settings from DT are board-specific baseline values,
but when adaptive TX Equalization is used, static settings are not final:
- If valid settings are retrieved from qTxEQGnSettings/wTxEQGnSettingsExt,
those retrieved settings override static DT settings.
- If retrieval is not available/valid, TX EQTR runs and trained settings
override static DT settings.
So static DT settings are a fallback and are intended for cases where
adaptive TX Equalization is not enabled/used.
No behavior changes for platforms that do not provide these properties.
What this series adds
=====================
1. dt-bindings:
- Document txeq-preshoot-g[1-6], txeq-deemphasis-g[1-6], and
tx-precode-enable-g6 in ufs-common.yaml.
- Define tuple encoding for host/device values per lane.
- Add per-property value validation ranges in schema.
2. UFS core/platform integration:
- Parse and validate per-gear DT TX EQ settings during platform init.
- Store parsed values into per-gear TX EQ params and track DT origin using
the from_dt flag.
- Integrate static-state handling in TX EQ flow so DT-provided entries are
fed through the adaptive TX Equalization path and then converted to
normal runtime params.
Link: https://patch.msgid.link/20260616113348.1168248-1-can.guo@oss.qualcomm.com
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
Expose the IOC and board temperature sensors of LSI / Broadcom SAS HBAs
through hwmon. Readings come from MPI IO Unit Page 7 via the accessor added
in the preceding patch.
The same fields are exposed by Broadcom's userspace tooling through the
/dev/mpt[23]ctl ioctl path (typically root-only): IOCTemperature and
BoardTemperature in lsiutil; ROC and Controller in storcli. With this
driver, sensors(1) shows them unprivileged:
$ sensors mpt3sas-pci-0200
mpt3sas-pci-0200
Adapter: PCI adapter
IOC: +42.0°C
Each channel is gated independently by its *TemperatureUnits field through
is_visible(); cards that populate only one sensor expose only one input
file, and cards that populate neither do not register an hwmon device.
The hwmon code is gated directly on CONFIG_HWMON. IS_REACHABLE() is used
rather than IS_ENABLED() so that SCSI_MPT3SAS=y with HWMON=m still builds;
in that configuration, the sensors are not exposed (same pattern as i915
and xe).
Assisted-by: Claude:claude-opus-4-7
Signed-off-by: Louis Sautier <sautier.louis@gmail.com>
Reviewed-by: Damien Le Moal <dlemoal@kernel.org>
Link: https://patch.msgid.link/20260630224922.2543096-3-sautier.louis@gmail.com
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
