When requeuing work to a draining workqueue the last work instance may
not be idle, so sas_queue_work() must not touch work->entry. Introduce
sas_work with a drain_node list_head to have a private list for
collecting work deferred due to drain collision.
Fixes reports like:
BUG: unable to handle kernel NULL pointer dereference at (null)
IP: [<ffffffff810410d4>] process_one_work+0x2e/0x338
Signed-off-by: Dan Williams <dan.j.williams@intel.com>
Signed-off-by: James Bottomley <JBottomley@Parallels.com>
libata issues follow up srsts when the controller has a hard time
recording the signature-fis after a reset, or if the link supports port
multipliers. libsas does not support port multipliers and no current
libsas lldds appear to need help retrieving the signature fis. Revert
it for now to remove confusion.
Signed-off-by: Dan Williams <dan.j.williams@intel.com>
Signed-off-by: James Bottomley <JBottomley@Parallels.com>
libsas ata error handling is already async but this does not help the
scan case. Move initial link recovery out from under host->scan_mutex,
and delay synchronization with eh until after all port probe/recovery
work has been queued.
Device ordering is maintained with scan order by still calling
sas_rphy_add() in order of domain discovery.
Since we now scan the domain list when invoking libata-eh we need to be
careful to check for fully initialized ata ports.
Acked-by: Jack Wang <jack_wang@usish.com>
Acked-by: Jeff Garzik <jgarzik@redhat.com>
Signed-off-by: Dan Williams <dan.j.williams@intel.com>
Signed-off-by: James Bottomley <JBottomley@Parallels.com>
If the top level expander is hot removed, mark all child devices as gone
before unregistration to short circuit futile recovery.
Signed-off-by: Dan Williams <dan.j.williams@intel.com>
Signed-off-by: James Bottomley <JBottomley@Parallels.com>
In the direct-attached case this routine returns the phy on which this
device was first discovered. Which is broken if we want to support
wide-targets, as this phy reference can become stale even though the
port is still active.
In the expander-attached case this routine tries to lookup the phy by
scanning the attached sas addresses of the parent expander, and BUG_ONs
if it can't find it. However since eh and the libsas workqueue run
independently we can still be attempting device recovery via eh after
libsas has recorded the device as detached. This is even easier to hit
now that eh is blocked while device domain rediscovery takes place, and
that libata is fed more timed out commands increasing the chances that
it will try to recover the ata device.
Arrange for dev->phy to always point to a last known good phy, it may be
stale after the port is torn down, but it will catch up for wide port
reconfigurations, and never be NULL.
Signed-off-by: Dan Williams <dan.j.williams@intel.com>
Signed-off-by: James Bottomley <JBottomley@Parallels.com>
Use ata_wait_after_reset() to poll for link recovery after a reset.
This combined with sas_ha->eh_mutex prevents expander rediscovery from
probing phys in an intermediate state. Local discovery does not have a
mechanism to filter link status changes during this timeout, so it
remains the responsibility of lldds to prevent premature port teardown.
Although once all lldd's support ->lldd_ata_check_ready() that could be
used as a gate to local port teardown.
The signature fis is re-transmitted when the link comes back so we
should be revalidating the ata device class, but that is left to a future
patch.
Signed-off-by: Dan Williams <dan.j.williams@intel.com>
Signed-off-by: James Bottomley <JBottomley@Parallels.com>
SAS does not tag SMP requests, and at least one lldd (isci) does not permit
more than one in-flight request at a time.
[jejb: fix sas_init_dev tab issues while we're at it]
Signed-off-by: Jeff Skirvin <jeffrey.d.skirvin@intel.com>
Signed-off-by: Dan Williams <dan.j.williams@intel.com>
Signed-off-by: James Bottomley <JBottomley@Parallels.com>
Execute the link-reset triggered by sas_phy_enable via
transport_sas_phy_reset so that it can be managed by libata.
Signed-off-by: Dan Williams <dan.j.williams@intel.com>
Signed-off-by: James Bottomley <JBottomley@Parallels.com>
libsas-eh if it successfully aborts an ata command will hide the timeout
condition (AC_ERR_TIMEOUT) from libata. The command likely completes
with the all-zero task->task_status it started with. Instead, interpret
a TMF_RESP_FUNC_COMPLETE as the end of the sas_task but keep the scmd
around for libata-eh to handle.
Tested-by: Andrzej Jakowski <andrzej.jakowski@intel.com>
Signed-off-by: Dan Williams <dan.j.williams@intel.com>
Signed-off-by: James Bottomley <JBottomley@Parallels.com>
Until we have told the lldd to forget a task a timed out operation can
return from the hardware at any time. Since completion frees the task
we need to make sure that no tasks run their normal completion handler
once eh has decided to manage the task. Similar to
ata_scsi_cmd_error_handler() freeze completions to let eh judge the
outcome of the race.
Task collector mode is problematic because it presents a situation where
a task can be timed out and aborted before the lldd has even seen it.
For this case we need to guarantee that a task that an lldd has been
told to forget does not get queued after the lldd says "never seen it".
With sas_scsi_timed_out we achieve this with the ->task_queue_flush
mutex, rather than adding more time.
Signed-off-by: Dan Williams <dan.j.williams@intel.com>
Signed-off-by: James Bottomley <JBottomley@Parallels.com>
We invoke task->task_done() to free the task in the eh case, but at this
point we are prepared for scsi_eh_flush_done_q() to finish off the scmd.
Introduce sas_end_task() to capture the final response status from the
lldd and free the task.
Also take the opportunity to kill this warning.
drivers/scsi/libsas/sas_scsi_host.c: In function ‘sas_end_task’:
drivers/scsi/libsas/sas_scsi_host.c:102:3: warning: case value ‘2’ not in enumerated type ‘enum exec_status’ [-Wswitch]
Signed-off-by: Dan Williams <dan.j.williams@intel.com>
Signed-off-by: James Bottomley <JBottomley@Parallels.com>
Since sas_ata does not implement ->freeze(), completions for scmds and
internal commands can still arrive concurrent with
ata_scsi_cmd_error_handler() and sas_ata_post_internal() respectively.
By the time either of those is called libata has committed to completing
the qc, and the ATA_PFLAG_FROZEN flag tells sas_ata_task_done() it has
lost the race.
In the sas_ata_post_internal() case we take on the additional
responsibility of freeing the sas_task to close the race with
sas_ata_task_done() freeing the the task while sas_ata_post_internal()
is in the process of invoking ->lldd_abort_task().
Signed-off-by: Dan Williams <dan.j.williams@intel.com>
Signed-off-by: James Bottomley <JBottomley@Parallels.com>
sas_discover_sata() notifies lldds of sata devices twice. Once to allow
the 'identify' to be sent, and a second time to allow aic94xx (the only
libsas driver that cares about sata_dev.identify) to setup NCQ
parameters before the device becomes known to the midlayer. Replace
this double notification and intervening 'identify' with an explicit
->lldd_ata_set_dmamode notification. With this change all ata internal
commands are issued by libata, so we no longer need sas_issue_ata_cmd().
The data from the identify command only needs to be cached in one
location so ata_device.id replaces domain_device.sata_dev.identify.
Signed-off-by: Dan Williams <dan.j.williams@intel.com>
Signed-off-by: James Bottomley <JBottomley@Parallels.com>
libata error handling provides for a timeout for link recovery. libsas
must not rescan for previously known devices in this interval otherwise
it may remove a device that is simply waiting for its link to recover.
Let libata-eh make the determination of when the link is stable and
prevent libsas (host workqueue) from taking action while this
determination is pending.
Using a mutex (ha->disco_mutex) to flush and disable revalidation while
eh is running requires any discovery action that may block on eh be
moved to its own context outside the lock. Probing ATA devices
explicitly waits on ata-eh and the cache-flush-io issued during device
removal may also pend awaiting eh completion. Essentially any rphy
add/remove activity needs to run outside the lock.
This adds two new cleanup states for sas_unregister_domain_devices()
'allocated-but-not-probed', and 'flagged-for-destruction'. In the
'allocated-but-not-probed' state dev->rphy points to a rphy that is
known to have not been through a sas_rphy_add() event. At domain
teardown check if this device is still pending probe and cleanup
accordingly. Similarly if a device has already been queued for removal
then sas_unregister_domain_devices has nothing to do.
Signed-off-by: Dan Williams <dan.j.williams@intel.com>
Signed-off-by: James Bottomley <JBottomley@Parallels.com>
When an lldd invokes ->notify_port_event() it can trigger a chain of libsas
events to:
1/ form the port and find the direct attached device
2/ if the attached device is an expander perform domain discovery
A call to flush_workqueue() will only flush the initial port formation work.
Currently libsas users need to call scsi_flush_work() up to the max depth of
chain (which will grow from 2 to 3 when ata discovery is moved to its own
discovery event). Instead of open coding multiple calls switch to use
drain_workqueue() to flush sas work.
drain_workqueue() does not handle new work submitted during the drain so
libsas needs a bit of infrastructure to hold off unchained work submissions
while a drain is in flight. A lldd ->notify() event is considered 'unchained'
while a sas_discover_event() is 'chained'. As Tejun notes:
"For now, I think it would be best to add private wrapper in libsas to
support deferring unchained work items while draining."
Signed-off-by: Dan Williams <dan.j.williams@intel.com>
Signed-off-by: James Bottomley <JBottomley@Parallels.com>
In preparation for adding new states (SAS_HA_DRAINING, SAS_HA_FROZEN),
convert ha->state into a set of flags.
Signed-off-by: Dan Williams <dan.j.williams@intel.com>
Signed-off-by: James Bottomley <JBottomley@Parallels.com>
The locks only served to make sure the pending event bitmask was updated
consistently.
Signed-off-by: Dan Williams <dan.j.williams@intel.com>
Signed-off-by: James Bottomley <JBottomley@Parallels.com>
Arrange for the deallocation of a struct domain_device object when it no
longer has:
1/ any children
2/ references by any scsi_targets
3/ references by a lldd
The comment about domain_device lifetime in
Documentation/scsi/libsas.txt is stale as it appears mainline never had
a version of a struct domain_device that was registered as a kobject.
We now manage domain_device reference counts on behalf of external
agents.
Reviewed-by: Jack Wang <jack_wang@usish.com>
Signed-off-by: Dan Williams <dan.j.williams@intel.com>
Signed-off-by: James Bottomley <JBottomley@Parallels.com>
Per commit 3e4ec344 "libata: kill ATA_FLAG_DISABLED" needing to set
ATA_DEV_NONE is a holdover from before libsas converted to the
"new-style" ata-eh.
Signed-off-by: Dan Williams <dan.j.williams@intel.com>
Signed-off-by: James Bottomley <JBottomley@Parallels.com>
Commit 1e34c838 "[SCSI] libsas: remove spurious sata control register
read/write" removed the routines to fake the presence of the sata
control registers, now remove the unused data structure fields to kill
any remaining confusion.
Acked-by: Jack Wang <jack_wang@usish.com>
Signed-off-by: Dan Williams <dan.j.williams@intel.com>
Signed-off-by: James Bottomley <JBottomley@Parallels.com>
port->dev_list maintains a list of devices attached to a given sas root port.
It needs to be mutated under a lock as contexts outside of the
single-threaded-libsas-workqueue access the list via sas_find_dev_by_rphy().
Fixup locations where the list was being mutated without a lock.
This is a follow-up to commit 5911e963 "[SCSI] libsas: remove expander
from dev list on error", where Luben noted [1]:
> 2/ We have unlocked list manipulations in sas_ex_discover_end_dev(),
> sas_unregister_common_dev(), and sas_ex_discover_end_dev()
Yes, I can see that and that is very unfortunate.
[1]: http://marc.info/?l=linux-scsi&m=131480962006471&w=2
Signed-off-by: Dan Williams <dan.j.williams@intel.com>
Signed-off-by: James Bottomley <JBottomley@Parallels.com>
Based on original implementation from Jiangbi Liu and Maciej Trela.
ATAPI transfers happen in two-to-three stages. The two stage atapi
commands are those that include a dma data transfer. The data transfer
portion of these operations is handled by the hardware packet-dma
acceleration. The three-stage commands do not have a data transfer and
are handled without hardware assistance in raw frame mode.
stage1: transmit host-to-device fis to notify the device of an incoming
atapi cdb. Upon reception of the pio-setup-fis repost the task_context
to perform the dma transfer of the cdb+data (go to stage3), or repost
the task_context to transmit the cdb as a raw frame (go to stage 2).
stage2: wait for hardware notification of the cdb transmission and then
go to stage 3.
stage3: wait for the arrival of the terminating device-to-host fis and
terminate the command.
To keep the implementation simple we only support ATAPI packet-dma
protocol (for commands with data) to avoid needing to handle the data
transfer manually (like we do for SATA-PIO). This may affect
compatibility for a small number of devices (see
ATA_HORKAGE_ATAPI_MOD16_DMA).
If the data-transfer underruns, or encounters an error the
device-to-host fis is expected to arrive in the unsolicited frame queue
to pass to libata for disposition. However, in the DONE_UNEXP_FIS (data
underrun) case it appears we need to craft a response. In the
DONE_REG_ERR case we do receive the UF and propagate it to libsas.
Signed-off-by: Maciej Trela <maciej.trela@intel.com>
Signed-off-by: Dan Williams <dan.j.williams@intel.com>
Signed-off-by: James Bottomley <JBottomley@Parallels.com>
Allow expander table-to-table attachments for
expanders that support it.
Signed-off-by: Luben Tuikov <ltuikov@yahoo.com>
Signed-off-by: James Bottomley <JBottomley@Parallels.com>
