dc395x_init_one()->adapter_init() might fail. In this case, the acb is
already cleaned up by adapter_init(), no need to do that in
adapter_uninit(acb) again.
[ 1.252251] dc395x: adapter init failed
[ 1.254900] RIP: 0010:adapter_uninit+0x94/0x170 [dc395x]
[ 1.260307] Call Trace:
[ 1.260442] dc395x_init_one.cold+0x72a/0x9bb [dc395x]
Link: https://lore.kernel.org/r/20210907040702.1846409-1-ztong0001@gmail.com
Fixes: 1da177e4c3 ("Linux-2.6.12-rc2")
Reviewed-by: Finn Thain <fthain@linux-m68k.org>
Signed-off-by: Tong Zhang <ztong0001@gmail.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
I fixed a stringop-overread warning earlier this year, now a second copy of
the original code was added and the warning came back:
drivers/scsi/lpfc/lpfc_attr.c: In function 'lpfc_cmf_info_show':
drivers/scsi/lpfc/lpfc_attr.c:289:25: error: 'strnlen' specified bound 4095 exceeds source size 24 [-Werror=stringop-overread]
289 | strnlen(LPFC_INFO_MORE_STR, PAGE_SIZE - 1),
| ^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
Fix it the same way as the other copy.
Link: https://lore.kernel.org/r/20210920095628.1191676-1-arnd@kernel.org
Fixes: ada48ba70f ("scsi: lpfc: Fix gcc -Wstringop-overread warning")
Fixes: 74a7baa2a3 ("scsi: lpfc: Add cmf_info sysfs entry")
Reviewed-by: James Smart <jsmart2021@gmail.com>
Signed-off-by: Arnd Bergmann <arnd@arndb.de>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
The acornscsi driver has a config option to enable tagged queuing, but this
option gets disabled in the driver itself with the comment 'needs to be
debugged'. As this is a _really_ old driver I doubt anyone will be wanting
to invest time here, so remove the tagged queue vestiges and make our lives
easier.
[jpg: Use scsi_cmd_to_rq()]
Link: https://lore.kernel.org/r/1631696835-136198-3-git-send-email-john.garry@huawei.com
Signed-off-by: Hannes Reinecke <hare@suse.de>
Signed-off-by: John Garry <john.garry@huawei.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
Under the session level spinlock node->active_ios_lock in
efct_scsi_io_alloc() we are taking another spinlock for the port. This
leads to contention between sessions and even between I/Os in the same
session.
Reduce the locked region to active_ios list for which active_ios_lock is
intended. Spinlock CPU usage decreases from 18% down to 13%. IOPS are
increased from 220 kIOPS to 264 kIOPS for one LUN.
Link: https://lore.kernel.org/r/20210914105539.6942-4-d.bogdanov@yadro.com
Reviewed-by: Roman Bolshakov <r.bolshakov@yadro.com>
Reviewed-by: Ram Vegesna <ram.vegesna@broadcom.com>
Signed-off-by: Dmitry Bogdanov <d.bogdanov@yadro.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
This reverts commit a113eaaf86.
There are a couple of issues with the commit:
1. It causes deadlocks.
2. It causes the shost->eh_cmd_q list of failed requests not to be
processed, ever.
So revert it.
1. Deadlocks
The SCSI error handler runs with requests blocked beginning when
scsi_schedule_eh() sets SHOST_RECOVERY state, continuing through
scsi_error_handler() callback ->eh_strategy_handler() until
scsi_restart_operations() is called. By setting eh_strategy_handler to
ufshcd_err_handler, the patch changed the UFS error handler to run with
requests blocked, including PM requests, for the entire run of the error
handler.
That conflicts with UFS error handler existing synchronization with UFS
device PM operations. The UFS error handler synchronizes with runtime PM
by doing pm_runtime_get_sync() prior to blocking requests itself. It
synchronizes with system PM by use of hba->host_sem, again before blocking
requests itself. However, if requests are already blocked, then PM
operations will block. So:
the UFS error handler blocks waiting on PM
+ PM blocks waiting on SCSI PM requests to process or fail
+ PM requests are blocked waiting on error handling to finish
= deadlock
This happens both for runtime PM and system PM.
Prior to the patch, these deadlocks could not happen even if SCSI error
handling was running, because the presence of requests in shost->eh_cmd_q
would mean the queues could not be suspended, which would mean that, should
the UFS error handler run at the same time, it would not need to wait for
PM or vice versa.
Please note these scenarios are not just theoretical, they were found
during testing on a Samsung Galaxy Book S.
2. ->eh_strategy_handler() must process shost->eh_cmd_q list of failed
requests, as all other eh_strategy_handler's do except UFS error handler.
Refer for example: scsi_unjam_host(), ata_scsi_error() and
sas_scsi_recover_host().
Link: https://lore.kernel.org/r/20210917144349.14058-1-adrian.hunter@intel.com
Fixes: a113eaaf86 ("scsi: ufs: Synchronize SCSI and UFS error handling")
Reviewed-by: Bart Van Assche <bvanassche@acm.org>
Signed-off-by: Adrian Hunter <adrian.hunter@intel.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
Injecting errors on the PCI slot while the driver is handling NVMe I/O will
cause crashes and hangs.
There are several rather difficult scenarios occurring. The main issue is
that the adapter can report a PCI error before or simultaneously to the PCI
subsystem reporting the error. Both paths have different entry points and
currently there is no interlock between them. Thus multiple teardown paths
are competing and all heck breaks loose.
Complicating things is the NVMs path. To a large degree, I/O was able to be
shutdown for a full FC port on the SCSI stack. But on NVMe, there isn't a
similar call. At best, it works on a per-controller basis, but even at the
controller level, it's a controller "reset" call. All of which means I/O is
still flowing on different CPUs with reset paths expecting hw access
(mailbox commands) to execute properly.
The following modifications are made:
- A new flag is set in PCI error entrypoints so the driver can track being
called by that path.
- An interlock is added in the SLI hw error path and the PCI error path
such that only one of the paths proceeds with the teardown logic.
- RPI cleanup is patched such that RPIs are marked unregistered w/o mbx
cmds in cases of hw error.
- If entering the SLI port re-init calls, a case where SLI error teardown
was quick and beat the PCI calls now reporting error, check whether the
SLI port is still live on the PCI bus.
- In the PCI reset code to bring the adapter back, recheck the IRQ
settings. Different checks for SLI3 vs SLI4.
- In I/O completions, that may be called as part of the cleanup or
underway just before the hw error, check the state of the adapter. If
in error, shortcut handling that would expect further adapter
completions as the hw error won't be sending them.
- In routines waiting on I/O completions, which may have been in progress
prior to the hw error, detect the device is being torn down and abort
from their waits and just give up. This points to a larger issue in the
driver on ref-counting for data structures, as it doesn't have
ref-counting on q and port structures. We'll do this fix for now as it
would be a major rework to be done differently.
- Fix the NVMe cleanup to simulate NVMe I/O completions if I/O is being
failed back due to hw error.
- In I/O buf allocation, done at the start of new I/Os, check hw state and
fail if hw error.
Link: https://lore.kernel.org/r/20210910233159.115896-10-jsmart2021@gmail.com
Co-developed-by: Justin Tee <justin.tee@broadcom.com>
Signed-off-by: Justin Tee <justin.tee@broadcom.com>
Signed-off-by: James Smart <jsmart2021@gmail.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
A prior patch inadvertently caused lpfc_sli_sum_iocb() to exclude counting
of outstanding aborted I/Os and ABORT IOCBs. Thus,
lpfc_reset_flush_io_context() called from any TMF routine does not properly
wait to flush all outstanding FCP IOCBs leading to a block layer crash on
an invalid scsi_cmnd->request pointer.
kernel BUG at ../block/blk-core.c:1489!
RIP: 0010:blk_requeue_request+0xaf/0xc0
...
Call Trace:
<IRQ>
__scsi_queue_insert+0x90/0xe0 [scsi_mod]
blk_done_softirq+0x7e/0x90
__do_softirq+0xd2/0x280
irq_exit+0xd5/0xe0
do_IRQ+0x4c/0xd0
common_interrupt+0x87/0x87
</IRQ>
Fix by separating out the LPFC_IO_FCP, LPFC_IO_ON_TXCMPLQ,
LPFC_DRIVER_ABORTED, and CMD_ABORT_XRI_CN || CMD_CLOSE_XRI_CN checks into a
new lpfc_sli_validate_fcp_iocb_for_abort() routine when determining to
build an ABORT iocb.
Restore lpfc_reset_flush_io_context() functionality by including counting
of outstanding aborted IOCBs and ABORT IOCBs in lpfc_sli_sum_iocb().
Link: https://lore.kernel.org/r/20210910233159.115896-9-jsmart2021@gmail.com
Fixes: e136471135 ("scsi: lpfc: Fix illegal memory access on Abort IOCBs")
Cc: <stable@vger.kernel.org> # v5.12+
Co-developed-by: Justin Tee <justin.tee@broadcom.com>
Signed-off-by: Justin Tee <justin.tee@broadcom.com>
Signed-off-by: James Smart <jsmart2021@gmail.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
Currently, we hold off unregistering with NVMe transport layer until GID_FT
or ADISC completes upon receipt of RSCN. In the ADISC discovery routine,
for nodes not found in the GID_FT response, the nodes are unregistered from
the SCSI transport but not UNREG_RPI'd. Meaning outstanding WQEs continue
to be outstanding and were not failed back to the OS. If an NVMe device,
this mean there wasn't initial termination of the I/Os so they could be
issued on a different NVMe path.
Fix by unregistering the RPI so that I/O is cancelled.
Link: https://lore.kernel.org/r/20210910233159.115896-8-jsmart2021@gmail.com
Fixes: 0614568361 ("scsi: lpfc: Delay unregistering from transport until GIDFT or ADISC completes")
Co-developed-by: Justin Tee <justin.tee@broadcom.com>
Signed-off-by: Justin Tee <justin.tee@broadcom.com>
Signed-off-by: James Smart <jsmart2021@gmail.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
In pt-2-pt mode, the initiator does not log into the target after a PRLI
error. In pt-2-pt mode, the target responded to the PRLI by sending a
LOGO. The LOGO causes all ELS and I/Os to be aborted. This caused the PRLI
to fail. The PRLI completion path caused the discovery node to be dropped
to avoid being stick in an UNUSED (not logged in) state. As the node was
dropped there is no retry of the login and as it is pt-2-pt, there is no
RSCN to retrigger discovery. Thus the other end is not seen by the OS.
Fix by ensuring the discovery node is not dropped if connecting pt-2-pt.
This will cause PLOGI to be retried.
Link: https://lore.kernel.org/r/20210910233159.115896-7-jsmart2021@gmail.com
Co-developed-by: Justin Tee <justin.tee@broadcom.com>
Signed-off-by: Justin Tee <justin.tee@broadcom.com>
Signed-off-by: James Smart <jsmart2021@gmail.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
On link up and node discovery, a remote port is registered with the SCSI
transport and the driver sets fc4_xpt_flags to track transport
registration.
A link down event causes the driver to deregister with the SCSI transport,
starting the devloss timer, and calls a local unreg routine to clear the
login state. Part of the login state is the fc4_xpt_flags. However, with
tape devices that support sequence level error recovery, which wants to
preserve the login, the local unreg routine is skipped, thus the flags
aren't cleared.
A subsequent link up, ADISC is performed and the lpfc_nlp_reg_node()
routine is called. As the fc4_xpt_flags is not clear, it's believed the
node is already registered with the transport. Unfortunately, the
registration was already terminated. Eventually the devloss tmo timer
expires and tears down the device.
Fix by ensuring the tape device, known by the ADISC flag, is always
unregistered if the link drops.
Link: https://lore.kernel.org/r/20210910233159.115896-6-jsmart2021@gmail.com
Co-developed-by: Justin Tee <justin.tee@broadcom.com>
Signed-off-by: Justin Tee <justin.tee@broadcom.com>
Signed-off-by: James Smart <jsmart2021@gmail.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
A test scenario encountered an unload hang while an FLOGI ELS was in flight
when a link down condition occurred. The driver fails unload as it never
releases the fport node.
For most nodes, when the link drops, devloss tmo is started and the timeout
will cause the final node release. For the Fport, as it has not yet
registered with the SCSI transport, there is no devloss timer to be
started, so there is no final release. Additionally, the link down
sequence causes ABORTS to be issued for pending ELS's. The completions from
the ABORTS perform the release of node references. However, as the adapter
is being reset to be unloaded, those completions will never occur.
Fix by the following:
- In the ELS cleanup, recognize when unloading and place the ELS's on a
different list that immediately cleans up/completes the ELS's. It's
recognized that this condition primarily affects only the fport, with
other ports having normal clean up logic that handles things.
- Resolve the devloss issue by, when cleaning up nodes on after link down,
recognizing when the fabric node does not have a completed state (its
state is UNUSED) and removing a reference so the node can delete after
the ELS reference is released.
Link: https://lore.kernel.org/r/20210910233159.115896-5-jsmart2021@gmail.com
Co-developed-by: Justin Tee <justin.tee@broadcom.com>
Signed-off-by: Justin Tee <justin.tee@broadcom.com>
Signed-off-by: James Smart <jsmart2021@gmail.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
A test scenario has a target issuing a TPLS after accepting the driver's
PRLI. TPLS is not supported by the driver so it rejects the ELS. However,
the reject was only happening on the primary N_Port. If the TPLS was to a
NPIV vport, not only would it reject the ELS, but it would act on the TPLS,
starting devloss, then unregister from the SCSI transport and release the
node. When devloss expired, it would access the node again and cause a page
faul.
Fix by altering the NPIV code to recognize that a correctly registered node
can reject unsolicited ELS I/O and to not unregister with the SCSI
transport and tear the node down. Add a check of the fc4_xpt_flags so that
only a zero value allows the unreg and teardown.
Link: https://lore.kernel.org/r/20210910233159.115896-4-jsmart2021@gmail.com
Co-developed-by: Justin Tee <justin.tee@broadcom.com>
Signed-off-by: Justin Tee <justin.tee@broadcom.com>
Signed-off-by: James Smart <jsmart2021@gmail.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
In a rarely executed path, FLOGI failure, there is a refcounting error. If
FLOGI completed with an error, typically a timeout, the initial completion
handler would remove the job reference. However, the job completion isn't
the actual end of the job/exchange as the timeout usually initiates an
ABTS, and upon that ABTS completion, a final completion is sent. The driver
removes the reference again in the final completion. Thus the imbalance.
In the buggy cases, if there was a link bounce while the delayed response
is outstanding, the fport node may be referenced again but there was no
additional reference as it is already present. The delayed completion then
occurs and removes the last reference freeing the node and causing issues
in the link up processed that is using the node.
Fix this scenario by removing the snippet that removed the reference in the
initial FLOGI completion. The bad snippet was poorly trying to identify the
FLOGI as OK to do so by realizing the node was not registered with either
SCSI or NVMe transport.
Link: https://lore.kernel.org/r/20210910233159.115896-3-jsmart2021@gmail.com
Fixes: 618e2ee146 ("scsi: lpfc: Fix FLOGI failure due to accessing a freed node")
Cc: <stable@vger.kernel.org> # v5.13+
Co-developed-by: Justin Tee <justin.tee@broadcom.com>
Signed-off-by: Justin Tee <justin.tee@broadcom.com>
Signed-off-by: James Smart <jsmart2021@gmail.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
When parsing the txq list in lpfc_drain_txq(), the driver attempts to pass
the requests to the adapter. If such an attempt fails, a local "fail_msg"
string is set and a log message output. The job is then added to a
completions list for cancellation.
Processing of any further jobs from the txq list continues, but since
"fail_msg" remains set, jobs are added to the completions list regardless
of whether a wqe was passed to the adapter. If successfully added to
txcmplq, jobs are added to both lists resulting in list corruption.
Fix by clearing the fail_msg string after adding a job to the completions
list. This stops the subsequent jobs from being added to the completions
list unless they had an appropriate failure.
Link: https://lore.kernel.org/r/20210910233159.115896-2-jsmart2021@gmail.com
Co-developed-by: Justin Tee <justin.tee@broadcom.com>
Signed-off-by: Justin Tee <justin.tee@broadcom.com>
Signed-off-by: James Smart <jsmart2021@gmail.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
Authentication application may be running and in the past tried to probe
driver (app_start) but was unsuccessful. This could be due to the bsg layer
not being ready to service the request. On a successful link up, driver
will use the netlink Link Up event to notify the app to retry the app_start
call.
In another case, app does not poll for new NPIV host. This link up event
would notify app of the presence of a new SCSI host.
Link: https://lore.kernel.org/r/20210908164622.19240-6-njavali@marvell.com
Fixes: 4de067e5df ("scsi: qla2xxx: edif: Add N2N support for EDIF")
Reviewed-by: Himanshu Madhani <himanshu.madhani@oracle.com>
Signed-off-by: Quinn Tran <qutran@marvell.com>
Signed-off-by: Nilesh Javali <njavali@marvell.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
System crash was seen when I/O was run against an NVMe target and aborts
were occurring.
Crash stack is:
-- relevant crash stack --
BUG: kernel NULL pointer dereference, address: 0000000000000010
:
#6 [ffffae1f8666bdd0] page_fault at ffffffffa740122e
[exception RIP: qla_nvme_abort_work+339]
RIP: ffffffffc0f592e3 RSP: ffffae1f8666be80 RFLAGS: 00010297
RAX: 0000000000000000 RBX: ffff9b581fc8af80 RCX: ffffffffc0f83bd0
RDX: 0000000000000001 RSI: ffff9b5839c6c7c8 RDI: 0000000008000000
RBP: ffff9b6832f85000 R8: ffffffffc0f68160 R9: ffffffffc0f70652
R10: ffffae1f862ffdc8 R11: 0000000000000300 R12: 000000000000010d
R13: 0000000000000000 R14: ffff9b5839cea000 R15: 0ffff9b583fab170
ORIG_RAX: ffffffffffffffff CS: 0010 SS: 0018
#7 [ffffae1f8666be98] process_one_work at ffffffffa6aba184
#8 [ffffae1f8666bed8] worker_thread at ffffffffa6aba39d
#9 [ffffae1f8666bf10] kthread at ffffffffa6ac06ed
The crash was due to a stale SRB structure access after it was aborted.
Fix the issue by removing stale access.
Link: https://lore.kernel.org/r/20210908164622.19240-5-njavali@marvell.com
Fixes: 2cabf10dbb ("scsi: qla2xxx: Fix hang on NVMe command timeouts")
Cc: stable@vger.kernel.org
Reviewed-by: Himanshu Madhani <himanshu.madhani@oracle.com>
Signed-off-by: Arun Easi <aeasi@marvell.com>
Signed-off-by: Nilesh Javali <njavali@marvell.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
Commit 1f02beff22 ("scsi: pm80xx: Remove global lock from outbound queue
processing") introduced a lock per outbound queue. Prior to that change the
driver was using a global lock for all outbound queues.
While processing the I/O responses and events the driver takes the outbound
queue spinlock and is supposed to release it in pm8001_ccb_task_free_done()
before calling command done(). Since the older code was using a global
lock, pm8001_ccb_task_free_done() was releasing the global spin lock. The
change that split the lock per outbound queue did not consider this and
pm8001_ccb_task_free_done() was still releasing the global lock.
Link: https://lore.kernel.org/r/20210906170404.5682-3-Ajish.Koshy@microchip.com
Fixes: 1f02beff22 ("scsi: pm80xx: Remove global lock from outbound queue processing")
Acked-by: Jack Wang <jinpu.wang@ionos.com>
Signed-off-by: Ajish Koshy <Ajish.Koshy@microchip.com>
Signed-off-by: Viswas G <Viswas.G@microchip.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
During phyup event, the firmware provides the phy_id and port_id and driver
is supposed to use these during device handle registration. Previously the
driver was using the port id value from libsas during device handle
registration. Since id can be different from the one assigned by firmware,
this can lead to wrong device registration and drives not showing up.
Use firmware assigned port id during device registration.
Link: https://lore.kernel.org/r/20210906170404.5682-2-Ajish.Koshy@microchip.com
Acked-by: Jack Wang <jinpu.wang@ionos.com>
Signed-off-by: Ajish Koshy <Ajish.Koshy@microchip.com>
Signed-off-by: Viswas G <Viswas.G@microchip.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
sd_spinup_disk() is a little bit noisy after commit 848ade90ba ("scsi:
sd: Do not exit sd_spinup_disk() quietly"):
scsi 0:0:0:0: Direct-Access Multiple Card Reader 1.00 PQ: 0 ANSI: 0
sd 0:0:0:0: Attached scsi generic sg0 type 0
sd 0:0:0:0: [sda] Media removed, stopped polling
sd 0:0:0:0: [sda] Media removed, stopped polling
sd 0:0:0:0: [sda] Attached SCSI removable disk
sd 0:0:0:0: [sda] Media removed, stopped polling
There's not really a benefit in printing the same message multiple
times. Therefore print it only if media_present was previously set.
Link: https://lore.kernel.org/r/a2d0a249-6035-9697-626a-e14ec50ef6ee@gmail.com
Reviewed-by: Bart Van Assche <bvanassche@acm.org>
Signed-off-by: Heiner Kallweit <hkallweit1@gmail.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
