MDS diagnostics fail because of frame count mismatch.

Unavailability of SGL is the trigger for this issue. If ELS SGL is not
available to process MDS frame, IOCB is put in FCP txq but not attempted to
post afterwards. So, driver stops processing incoming frames as it runs out
of IOCB.  lpfc_drain_txq attempts to submit IOCBS that are queued in ELS
txq but MDS frames are posted to FCP WQ.

Attempt to submit IOCBs that are present in FCP txq when MDS loopback is
running.
Signed-off-by: NDick Kennedy <dick.kennedy@broadcom.com>
Signed-off-by: NJames Smart <james.smart@broadcom.com>
Reviewed-by: NHannes Reinecke <hare@suse.com>
Signed-off-by: NMartin K. Petersen <martin.petersen@oracle.com>

When issuing a nexus reset for directly attached device, we want to ignore
the PHY down events so libsas will not deform and reform the port.

In the case that the attached SAS changes for the reset, libsas will deform
and form a port.

For scenario that the PHY does not come up after a timeout period, then
report the PHY down to libsas.
Signed-off-by: NXiang Chen <chenxiang66@hisilicon.com>
Signed-off-by: NJohn Garry <john.garry@huawei.com>
Signed-off-by: NMartin K. Petersen <martin.petersen@oracle.com>

It is an step of executing task to get free slot. If the step fails, we
will cleanup LLDD resources and should return failure to upper layer or
internal caller to abort task execution of this time.

But in the current code, the caller of get_free_slot() doesn't return
failure when get_free_slot() failed. This patch is to fix it.
Signed-off-by: NXiaofei Tan <tanxiaofei@huawei.com>
Signed-off-by: NJohn Garry <john.garry@huawei.com>
Signed-off-by: NMartin K. Petersen <martin.petersen@oracle.com>

For v2 hw, STP link from target is rejected after host reset because of a
SoC bug. The STP reject will be terminated after we have sent IO from each
PHY of a port.

This is not an problem before, as we don't need to setup STP link from
target immediately after host reset. But now, it is.  Because we want to
send soft-reset immediately after host reset.

In order to terminate STP reject quickly, this patch send ATA reset command
through each PHY of a port. Notes: ATA reset command don't need target's
response.

Besides, we do abort dev for each device before terminating STP reject.
This is a quirk of v2 hw.
Signed-off-by: NXiaofei Tan <tanxiaofei@huawei.com>
Signed-off-by: NJohn Garry <john.garry@huawei.com>
Signed-off-by: NMartin K. Petersen <martin.petersen@oracle.com>

This patch adds a force PHY function for internal ATA command for v2 hw.

Because there is an SoC bug in v2 hw, and need send an IO through each PHY
of a port to work around a bug which occurs after a controller reset.

This force PHY function will be used in the later patch.
Signed-off-by: NXiaofei Tan <tanxiaofei@huawei.com>
Signed-off-by: NJohn Garry <john.garry@huawei.com>
Signed-off-by: NMartin K. Petersen <martin.petersen@oracle.com>

In future scenarios we will want to use the TMF struct for more task types
than SSP.

As such, we can add struct hisi_sas_tmf_task directly into struct
hisi_sas_slot, and this will mean we can remove the TMF parameters from the
task prep functions.
Signed-off-by: NXiaofei Tan <tanxiaofei@huawei.com>
Signed-off-by: NJohn Garry <john.garry@huawei.com>
Signed-off-by: NMartin K. Petersen <martin.petersen@oracle.com>

We may reset the controller in many scenarios, such as SCSI EH and HW
errors. There should be no IO which returns from target when SCSI EH is
active. But for other scenarios, there may be.  It is not necessary to make
such IOs fail.

This patch adds an function of trying to wait for any commands, or IO, to
complete before host reset. If no more CQ returned from host controller in
100ms, we assume no more IO can return, and then stop waiting. We wait 5s
at most.

The HW has a register CQE_SEND_CNT to indicate the total number of CQs that
has been reported to driver. We can use this register and it is reliable to
resd this register in such scenarios that require host reset.
Signed-off-by: NXiaofei Tan <tanxiaofei@huawei.com>
Signed-off-by: NJohn Garry <john.garry@huawei.com>
Signed-off-by: NMartin K. Petersen <martin.petersen@oracle.com>

After the controller is reset, it is possible that the disks attached still
have outstanding IO to complete.

Thus, when the PHYs come back up after controller reset, it is possible
that these IOs complete at some unknown point later.

We want to ensure that all IOs are complete after the controller reset so
that all associated IPTT and other resources can be recycled safely.

To achieve this, re-init the disks by TMF or softreset (in case of ATA
devices).

If the init fails - maybe because the device was removed or link has not
come up - then do not release the device resources, but rather rely on SCSI
EH to handle the timeout for these resources later on.

This patch also does some cleanup to hisi_sas_init_disk(), including
removing superfluous cases in the switch statement.
Signed-off-by: NXiaofei Tan <tanxiaofei@huawei.com>
Signed-off-by: NJohn Garry <john.garry@huawei.com>
Signed-off-by: NMartin K. Petersen <martin.petersen@oracle.com>

When a SCSI host is registered, the SCSI mid-layer takes a reference to a
module in Scsi_host.hostt.module. In doing this, we are prevented from
removing the driver module for the host in dangerous scenario, like when a
disk is mounted.

Currently there is only one scsi_host_template (sht) for all HW versions,
and this is the main.c module. So this means that we can possibly remove
the HW module in this dangerous scenario, as SCSI mid-layer is only
referencing the main.c module.

To fix this, create a sht per module, referencing that same module to
create the Scsi host.
Signed-off-by: NXiang Chen <chenxiang66@hisilicon.com>
Signed-off-by: NJohn Garry <john.garry@huawei.com>
Signed-off-by: NMartin K. Petersen <martin.petersen@oracle.com>

When a disk is discovered, it may be in an error state, or there may be
residual commands remaining in the disk.

To ensure any disk is in good state after discovery, reset via TMF (for SAS
disk) or softreset (for a SATA disk).
Signed-off-by: NXiang Chen <chenxiang66@hisilicon.com>
Signed-off-by: NXiaofei Tan <tanxiaofei@huawei.com>
Signed-off-by: NJohn Garry <john.garry@huawei.com>
Signed-off-by: NMartin K. Petersen <martin.petersen@oracle.com>

This patch implements LED feature of directly attached disk for v3 hw.

In fact, this hw has created an SGPIO component for LED feature, and we can
control LEDs just by internal registers.
Signed-off-by: NXiaofei Tan <tanxiaofei@huawei.com>
Signed-off-by: NJohn Garry <john.garry@huawei.com>
Signed-off-by: NMartin K. Petersen <martin.petersen@oracle.com>

To reduce possibility of hitting unknown SoC bugs and aid debugging and
test, change allocation mode of device id from last used device id instead
of lowest available index.
Signed-off-by: NXiang Chen <chenxiang66@hisilicon.com>
Signed-off-by: NJohn Garry <john.garry@huawei.com>
Signed-off-by: NMartin K. Petersen <martin.petersen@oracle.com>

Currently we find the lowest available empty bit in the IPTT bitmap to
allocate the IPTT for a command.

To reduce possibility of hitting unknown SoC bugs and also aid in the
debugging of those same bugs, change the allocation mode.

The next allocation method is to use the next free slot adjacent to the
most recently allocated slot, in a round-robin fashion.
Signed-off-by: NXiang Chen <chenxiang66@hisilicon.com>
Signed-off-by: NJohn Garry <john.garry@huawei.com>
Signed-off-by: NMartin K. Petersen <martin.petersen@oracle.com>

There is much common code and functionality between the HW versions to set
the PHY linkrate.

As such, this patch factors out the common code into a generic function
hisi_sas_phy_set_linkrate().
Signed-off-by: NJohn Garry <john.garry@huawei.com>
Signed-off-by: NMartin K. Petersen <martin.petersen@oracle.com>

Fix a typo in hisi_sas_task_prep().

Fixes: 7eee4b92 ("scsi: hisi_sas: relocate smp sg map")
Signed-off-by: NWei Yongjun <weiyongjun1@huawei.com>
Signed-off-by: NMartin K. Petersen <martin.petersen@oracle.com>

Trivial fix to spelling mistake in pm8001_printk message text; also I
believe NOT_UNSUPPORTED should probably be NOT_SUPPORTED. Also fix the
indent of the pm8001_printk statement.
Signed-off-by: NColin Ian King <colin.king@canonical.com>
Acked-by: NJack Wang <jinpu.wang@profitbricks.com>
Signed-off-by: NMartin K. Petersen <martin.petersen@oracle.com>

Commit 505aa4b6 ("scsi: sd: Defer spinning up drive while SANITIZE is
in progress") may not be sufficient, especially if the SCSI SANITIZE
command is sent via the bsg or sg pass-throughs, since they don't use the
sd driver.

Add "Sanitize in progress" plus some other recent "... in progress"
additional sense codes into the scsi mid-level so they are treated in a
similar fashion to "Format in progress".

[mkp: checkpatch]
Signed-off-by: NDouglas Gilbert <dgilbert@interlog.com>
Signed-off-by: NMartin K. Petersen <martin.petersen@oracle.com>

Trivial fix to spelling mistakes/typos:
"SNIC_IOREQ_ABTS_COMPELTE" -> "SNIC_IOREQ_ABTS_COMPLETE"
"SNIC_IOREQ_LR_COMPELTE"   -> "SNIC_IOREQ_LR_COMPLETE"
"SNIC_IOREQ_CMD_COMPELTE"  -> "SNIC_IOREQ_CMD_COMPLETE"
Signed-off-by: NColin Ian King <colin.king@canonical.com>
Signed-off-by: NMartin K. Petersen <martin.petersen@oracle.com>

The 'free_irq()' call is not at the right place in the error handling
path.  The changed order has been introduced in commit 3d4253d9
("[SCSI] qlogicpti: Convert to new SBUS device framework.")

Fixes: 3d4253d9 ("[SCSI] qlogicpti: Convert to new SBUS device framework.")
Signed-off-by: NChristophe JAILLET <christophe.jaillet@wanadoo.fr>
Reviewed-by: NDan Carpenter <dan.carpenter@oracle.com>
Signed-off-by: NMartin K. Petersen <martin.petersen@oracle.com>

UFS driver can receive a request during memory reclaim by kswapd.  So
when ufs driver puts the ungate work in queue, and if there are no idle
workers, kthreadd is invoked to create a new kworker. Since kswapd task
holds a mutex which kthreadd also needs, this can cause a deadlock
situation. So ungate work must be done in a separate work queue with
WQ_MEM_RECLAIM flag enabled.  Such a workqueue will have a rescue thread
which will be called when the above deadlock condition is possible.
Signed-off-by: NVijay Viswanath <vviswana@codeaurora.org>
Signed-off-by: NCan Guo <cang@codeaurora.org>
Signed-off-by: NAsutosh Das <asutoshd@codeaurora.org>
Reviewed-by: NSubhash Jadavani <subhashj@codeaurora.org>
Signed-off-by: NMartin K. Petersen <martin.petersen@oracle.com>

As multiple requests are submitted to the ufs host controller in
parallel there could be instances where the command completion interrupt
arrives later for a request that is already processed earlier as the
corresponding doorbell was cleared when handling the previous
interrupt. Read the interrupt status in a loop after processing the
received interrupt to catch such interrupts and handle it.
Signed-off-by: NVenkat Gopalakrishnan <venkatg@codeaurora.org>
Signed-off-by: NAsutosh Das <asutoshd@codeaurora.org>
Reviewed-by: NSubhash Jadavani <subhashj@codeaurora.org>
Signed-off-by: NMartin K. Petersen <martin.petersen@oracle.com>

UFSHCD_QUIRK_BROKEN_UFS_HCI_VERSION is only applicable for QCOM UFS host
controller version 2.x.y and this has been fixed from version 3.x.y
onwards, hence this change removes this quirk for version 3.x.y onwards.

[mkp: applied by hand]
Signed-off-by: NSubhash Jadavani <subhashj@codeaurora.org>
Signed-off-by: NAsutosh Das <asutoshd@codeaurora.org>
Signed-off-by: NMartin K. Petersen <martin.petersen@oracle.com>

Currently we call the scsi_block_requests()/scsi_unblock_requests()
whenever we want to block/unblock scsi requests but as there is no
reference counting, nesting of these calls could leave us in undesired
state sometime. Consider following call flow sequence:

1. func1() calls scsi_block_requests() but calls func2() before
   calling scsi_unblock_requests()
2. func2() calls scsi_block_requests()
3. func2() calls scsi_unblock_requests()
4. func1() calls scsi_unblock_requests()

As there is no reference counting, we will have scsi requests unblocked
after #3 instead of it to be unblocked only after #4. Though we may not
have failures seen with this, we might run into some failures in future.
Better solution would be to fix this by adding reference counting.
Signed-off-by: NSubhash Jadavani <subhashj@codeaurora.org>
Signed-off-by: NCan Guo <cang@codeaurora.org>
Signed-off-by: NAsutosh Das <asutoshd@codeaurora.org>
Signed-off-by: NMartin K. Petersen <martin.petersen@oracle.com>

Vendor specific setup_clocks ops may depend on clocks managed by ufshcd
driver so if the vendor specific setup_clocks callback is called when
the required clocks are turned off, it results into unclocked register
access.

This change make sure that required clocks are enabled before vendor
specific setup_clocks callback is called.
Signed-off-by: NSubhash Jadavani <subhashj@codeaurora.org>
Signed-off-by: NVenkat Gopalakrishnan <venkatg@codeaurora.org>
Signed-off-by: NCan Guo <cang@codeaurora.org>
Signed-off-by: NAsutosh Das <asutoshd@codeaurora.org>
Signed-off-by: NMartin K. Petersen <martin.petersen@oracle.com>

The device can set the exception event bit in one of the response UPIU,
for example to notify the need for urgent BKOPs operation.  In such a
case, the host driver calls ufshcd_exception_event_handler to handle
this notification.  When trying to check the exception event status (for
finding the cause for the exception event), the device may be busy with
additional SCSI commands handling and may not respond within the 100ms
timeout.

To prevent that, we need to block SCSI commands during handling of
exception events and allow retransmissions of the query requests, in
case of timeout.
Signed-off-by: NSubhash Jadavani <subhashj@codeaurora.org>
Signed-off-by: NMaya Erez <merez@codeaurora.org>
Signed-off-by: NCan Guo <cang@codeaurora.org>
Signed-off-by: NAsutosh Das <asutoshd@codeaurora.org>
Reviewed-by: NSubhash Jadavani <subhashj@codeaurora.org>
Signed-off-by: NMartin K. Petersen <martin.petersen@oracle.com>

On the quest to remove all VLAs from the kernel[1] this moves the
sg_list variable off the stack, as already done for other allocated
buffers in adpt_i2o_passthru(). Additionally consolidates the error path
for kfree().

[1] https://lkml.kernel.org/r/CA+55aFzCG-zNmZwX4A2FQpadafLfEzK6CC=qPXydAacU1RqZWA@mail.gmail.comSigned-off-by: NKees Cook <keescook@chromium.org>
Signed-off-by: NMartin K. Petersen <martin.petersen@oracle.com>

devfreq requires that the client operates on actual frequencies, not only 0
and UMAX_INT and as such UFS brok with the introduction of f1d981ea ("PM
/ devfreq: Use the available min/max frequency").

This patch registers the frequencies of the first clock with devfreq and use
these to determine if we're trying to step up or down.

Reviewed-by: Chanwoo Choi <cw00.choi@samsung.com> [for devfreq & OPP part]
Reviewed-by: NSubhash Jadavani <subhashj@codeaurora.org>
Signed-off-by: NBjorn Andersson <bjorn.andersson@linaro.org>
Signed-off-by: NMartin K. Petersen <martin.petersen@oracle.com>

Failing to register with devfreq leaves hba->devfreq assigned, which causes
the error path to dereference the ERR_PTR(). Rather than bolting on more
conditionals, move the call of devm_devfreq_add_device() into it's own
function and only update hba->devfreq once it's successfully registered.

The subsequent patch builds upon this to make UFS actually work again, as
it's been broken since f1d981ea ("PM / devfreq: Use the available
min/max frequency")

Also switch to use DEVFREQ_GOV_SIMPLE_ONDEMAND constant.
Reviewed-by: NChanwoo Choi <cw00.choi@samsung.com>
Signed-off-by: NBjorn Andersson <bjorn.andersson@linaro.org>
Signed-off-by: NMartin K. Petersen <martin.petersen@oracle.com>

Current code allocates 240 Kbytes (in typical configs) for each synthetic
SCSI controller to use as temp cpumask variables.  Recode to avoid needing
the temp cpumask variables and remove the memory allocation.
Signed-off-by: NMichael Kelley <mikelley@microsoft.com>
Acked-by: NStephen Hemminger <sthemmin@microsoft.com>
Signed-off-by: NMartin K. Petersen <martin.petersen@oracle.com>

Depending on the underlying transport, cxlflash has a dependency on either
the CXL or OCXL drivers, which are enabled via their Kconfig option.
Instead of having a module wide dependency on these config options, it is
better to isolate the object modules that are dependent on the CXL and OCXL
drivers and adjust the module dependencies accordingly.

This commit isolates the object files that are dependent on CXL and/or
OCXL. The cxl/ocxl fops used in the core driver are tucked under an ifdef to
avoid compilation errors.
Signed-off-by: NUma Krishnan <ukrishn@linux.vnet.ibm.com>
Acked-by: NMatthew R. Ochs <mrochs@linux.vnet.ibm.com>
Signed-off-by: NMartin K. Petersen <martin.petersen@oracle.com>

As a staging cleanup to support transport specific builds of the cxlflash
module, relocate device dependent assignments to header files. This will
avoid littering the core driver with conditional compilation logic.
Signed-off-by: NUma Krishnan <ukrishn@linux.vnet.ibm.com>
Acked-by: NMatthew R. Ochs <mrochs@linux.vnet.ibm.com>
Signed-off-by: NMartin K. Petersen <martin.petersen@oracle.com>

The new header file, backend.h, that was recently added is missing the
include guards. This commit adds the guards.
Signed-off-by: NUma Krishnan <ukrishn@linux.vnet.ibm.com>
Acked-by: NMatthew R. Ochs <mrochs@linux.vnet.ibm.com>
Signed-off-by: NMartin K. Petersen <martin.petersen@oracle.com>

AFUs can only process a single AFU command at a time. This is enforced with
a global mutex situated within the AFU send routine. As this mutex has a
global scope, it has the potential to unnecessarily block commands destined
for other AFUs.

Instead of using a global mutex, transition the mutex to be per-AFU. This
will allow commands to only be blocked by siblings of the same AFU.
Signed-off-by: NMatthew R. Ochs <mrochs@linux.vnet.ibm.com>
Signed-off-by: NUma Krishnan <ukrishn@linux.vnet.ibm.com>
Signed-off-by: NMartin K. Petersen <martin.petersen@oracle.com>

When a superpipe process that makes use of virtual LUNs is terminated or
killed abruptly, there is a possibility that the cxlflash driver could hang
and deprive other operations on the adapter.

The release fop registered to be invoked on a context close, detaches every
LUN associated with the context. The underlying service to detach the LUN
assumes it has been called with the read semaphore held, and releases the
semaphore before any operation that could be time consuming.

When invoked without holding the read semaphore, an opportunity is created
for the semaphore's count to become negative when it is temporarily released
during one of these potential lengthy operations. This negative count
results in subsequent acquisition attempts taking forever, leading to the
hang.

To support the current design point of holding the semaphore on the ioctl()
paths, the release fop should acquire it before invoking any ioctl services.
Signed-off-by: NUma Krishnan <ukrishn@linux.vnet.ibm.com>
Acked-by: NMatthew R. Ochs <mrochs@linux.vnet.ibm.com>
Signed-off-by: NMartin K. Petersen <martin.petersen@oracle.com>

The kernel log can get filled with debug messages from send_cmd_ioarrin()
when dynamic debug is enabled for the cxlflash module and there is a lot of
legacy I/O traffic.

While these messages are necessary to debug issues that involve command
tracking, the abundance of data can overwrite other useful data in the
log. The best option available is to limit the messages that should serve
most of the common use cases.
Signed-off-by: NUma Krishnan <ukrishn@linux.vnet.ibm.com>
Acked-by: NMatthew R. Ochs <mrochs@linux.vnet.ibm.com>
Signed-off-by: NMartin K. Petersen <martin.petersen@oracle.com>

The following Oops may be encountered if the device is reset, i.e. EEH
recovery, while there is heavy I/O traffic:

59:mon> t
[c000200db64bb680] c008000009264c40 cxlflash_queuecommand+0x3b8/0x500
					[cxlflash]
[c000200db64bb770] c00000000090d3b0 scsi_dispatch_cmd+0x130/0x2f0
[c000200db64bb7f0] c00000000090fdd8 scsi_request_fn+0x3c8/0x8d0
[c000200db64bb900] c00000000067f528 __blk_run_queue+0x68/0xb0
[c000200db64bb930] c00000000067ab80 __elv_add_request+0x140/0x3c0
[c000200db64bb9b0] c00000000068daac blk_execute_rq_nowait+0xec/0x1a0
[c000200db64bba00] c00000000068dbb0 blk_execute_rq+0x50/0xe0
[c000200db64bba50] c0000000006b2040 sg_io+0x1f0/0x520
[c000200db64bbaf0] c0000000006b2e94 scsi_cmd_ioctl+0x534/0x610
[c000200db64bbc20] c000000000926208 sd_ioctl+0x118/0x280
[c000200db64bbcc0] c00000000069f7ac blkdev_ioctl+0x7fc/0xe30
[c000200db64bbd20] c000000000439204 block_ioctl+0x84/0xa0
[c000200db64bbd40] c0000000003f8514 do_vfs_ioctl+0xd4/0xa00
[c000200db64bbde0] c0000000003f8f04 SyS_ioctl+0xc4/0x130
[c000200db64bbe30] c00000000000b184 system_call+0x58/0x6c

When there is no room to send the I/O request, the cached room is refreshed
by reading the memory mapped command room value from the AFU. The AFU
register mapping is refreshed during a reset, creating a race condition that
can lead to the Oops above.

During a device reset, the AFU should not be unmapped until all the active
send threads quiesce. An atomic counter, cmds_active, is currently used to
track internal AFU commands and quiesce during reset. This same counter can
also be used for the active send threads.
Signed-off-by: NUma Krishnan <ukrishn@linux.vnet.ibm.com>
Acked-by: NMatthew R. Ochs <mrochs@linux.vnet.ibm.com>
Signed-off-by: NMartin K. Petersen <martin.petersen@oracle.com>

Currently we don't check that device is not gone before dereferencing
its elements in the function hisi_sas_task_exec() (specifically, the DQ
pointer).

This patch fixes this issue by filling in the DQ pointer in
hisi_sas_task_prep() after we check that the device pointer is still
safe to reference.

[mkp: typo]
Signed-off-by: NXiaofei Tan <tanxiaofei@huawei.com>
Signed-off-by: NJohn Garry <john.garry@huawei.com>
Signed-off-by: NMartin K. Petersen <martin.petersen@oracle.com>

The IPTT of a slot is unique, and we currently use hisi_hba lock to
protect it.

Now slot is managed on hisi_sas_device.list, so use DQ lock to protect
for allocating and freeing the slot.
Signed-off-by: NXiang Chen <chenxiang66@hisilicon.com>
Signed-off-by: NJohn Garry <john.garry@huawei.com>
Signed-off-by: NMartin K. Petersen <martin.petersen@oracle.com>

Currently we lock the DQ to protect whole delivery process.  So this
stops us building slots for the same queue in parallel, and can affect
performance.

To optimise it, only lock the DQ during special periods, specifically
when allocating a slot from the DQ and when delivering a slot to the HW.

This approach is now safe, thanks to the previous patches to ensure that
we always deliver a slot to the HW once allocated.
Signed-off-by: NXiang Chen <chenxiang66@hisilicon.com>
Signed-off-by: NJohn Garry <john.garry@huawei.com>
Signed-off-by: NMartin K. Petersen <martin.petersen@oracle.com>

Currently we allocate the slot's memory buffer after allocating the DQ
slot.

To aid DQ lockout reduction, and allow slots to be built in parallel,
move this step (which can fail) prior to allocating the slot.

Also a stray spin_unlock_irqrestore() is removed from internal task exec
function.
Signed-off-by: NXiang Chen <chenxiang66@hisilicon.com>
Signed-off-by: NJohn Garry <john.garry@huawei.com>
Signed-off-by: NMartin K. Petersen <martin.petersen@oracle.com>