Mike Christie <michael.christie@oracle.com> says:
The following patches were made over Linus tree (Martin's 6.7 branch
was missing some changes to sd.c). They only contain the sshdr and
rdac retry fixes from the "Allow scsi_execute users to control
retries" patchset.
The patches in this set are reviewed and tested but the changes to how
we do retries will take a little longer and require more testing, so I
broke up the series to make them easier to review.
Link: https://lore.kernel.org/r/20231004210013.5601-1-michael.christie@oracle.com
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
If send_mode_select retries scsi_execute_cmd it will leave err set to
SCSI_DH_RETRY/SCSI_DH_IMM_RETRY. If on the retry, the command is
successful, then SCSI_DH_RETRY/SCSI_DH_IMM_RETRY will be returned to the
scsi_dh activation caller. On the retry, we will then detect the previous
MODE SELECT had worked, and so we will return success.
This patch has us return the correct return value, so we can avoid the
extra scsi_dh activation call and to avoid failures if the caller had hit
its activation retry limit and does not end up retrying.
Signed-off-by: Mike Christie <michael.christie@oracle.com>
Link: https://lore.kernel.org/r/20231004210013.5601-5-michael.christie@oracle.com
Reviewed-by: Martin Wilck <mwilck@suse.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
Mike Christie <michael.christie@oracle.com> says:
The following patches were made over Linus's tree but apply over
Martin's branches. They allow userspace to configure how fabric
drivers submit cmds to backend drivers.
Right now loop and vhost use a worker thread, and the other drivers
submit from the contexts they receive/process the cmd from. For
multiple LUN cases where the target can queue more cmds than the
backend can handle then deferring to a worker thread is safest because
the backend driver can block when doing things like waiting for a free
request/tag. Deferring also helps when the target has to handle
transport level requests from the recv context.
For cases where the backend devices can queue everything the target
sends, then there is no need to defer to a workqueue and you can see a
perf boost of up to 26% for small IO workloads. For a nvme device and
vhost-scsi I can see with 4K IOs:
fio jobs 1 2 4 8 10
--------------------------------------------------
workqueue
submit 94K 190K 394K 770K 890K
direct
submit 128K 252K 488K 950K -
Link: https://lore.kernel.org/r/1b1f7a5c-0988-45f9-b103-dfed2c0405b1@oracle.com
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
This exports the fabric driver's direct submit settings, so users know what
the driver supports. It will be helpful when they are exporting a device
through different targets and one doesn't support direct submission.
The new files allow the fabric to report what submission types they default
to and if they support direct submission:
default_submit_type:
1 - TARGET_DIRECT_SUBMIT - If the user has not requested a specific value
then the fabric requests direct submission.
2 - TARGET_QUEUE_SUBMIT - If the user has not requested a specific value
then the fabric requests queued submission.
Note that these fabric values are based on what the fabric driver currently
defaults to for compat with exiting setups.
direct_submit_supported:
0 - The fabric does not support direct submission.
1 - The fabric supports direct submission.
Signed-off-by: Mike Christie <michael.christie@oracle.com>
Link: https://lore.kernel.org/r/20230928020907.5730-8-michael.christie@oracle.com
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
This allows userspace to request the fabric drivers do direct submissions
if they support it. With the new device file, submit_type, users can
write 0 - 2 to control how commands are submitted to the backend:
0 - TARGET_FABRIC_DEFAULT_SUBMIT - LIO will use the fabric's default
submission type. This is the default for compat.
1 - TARGET_DIRECT_SUBMIT - LIO will submit the cmd to the backend from the
calling context if the fabric the cmd was received on supports it,
else it will use the fabric's default type.
2 - TARGET_QUEUE_SUBMIT - LIO will queue the cmd to the LIO submission
workqueue which will pass it to the backend.
When using an NVMe drive and vhost-scsi with direct submission we see
around a 20% improvement in 4K I/Os:
fio jobs 1 2 4 8 10
--------------------------------------------------
defer 94K 190K 394K 770K 890K
direct 128K 252K 488K 950K -
And when using the queueing mode, we now no longer see issues like where
the iSCSI tx thread is blocked in the block layer waiting on a tag so it
can't respond to a nop or perform I/Os for other LUs.
Signed-off-by: Mike Christie <michael.christie@oracle.com>
Link: https://lore.kernel.org/r/20230928020907.5730-6-michael.christie@oracle.com
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
Move the hack to clear some buffers to transport_handle_cdb_direct() so we
can eventually merge transport_handle_cdb_direct() and target_submit().
This also fixes up the comment so it's clear it was only for udev and
reflects that the referenced function does not exist and we now allow more
than 1 page for control CDBs.
Signed-off-by: Mike Christie <michael.christie@oracle.com>
Link: https://lore.kernel.org/r/20230928020907.5730-4-michael.christie@oracle.com
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
Move core_alua_check_nonop_delay() to transport_handle_cdb_direct() so the
iSCSI target driver doesn't have to call as many core functions
directly. We will eventually merge transport_handle_cdb_direct and
target_submit so iSCSI and the other drivers call a common function.
It will also be helpful as preparation for future changes which allow the
iSCSI target to defer command submission to the LIO submission workqueue,
because we will have a common submission function for that which will be
based on transport_handle_cdb_direct()/target_submit().
Signed-off-by: Mike Christie <michael.christie@oracle.com>
Link: https://lore.kernel.org/r/20230928020907.5730-3-michael.christie@oracle.com
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
In some cases, like with multiple LUN targets or where the target has to
respond to transport level requests from the receiving context it can be
better to defer cmd submission to a helper thread. If the backend driver
blocks on something like request/tag allocation it can block the entire
target submission path and other LUs and transport IO on that session.
In other cases like single LUN targets with storage that can support all
the commands that the target can queue, then it's best to submit the cmd
to the backend from the target's cmd receiving context.
Subsequent commits will allow the user to config what they prefer, but
drivers like loop can't directly submit because they can be called from a
context that can't sleep. And, drivers like vhost-scsi can support direct
submission, but need to keep their default behavior of deferring execution
to avoid possible regressions where the backend can block.
Make the drivers tell LIO core if they support direct submissions and their
current default, so we can prevent users from misconfiguring the system and
initialize devices correctly.
Signed-off-by: Mike Christie <michael.christie@oracle.com>
Link: https://lore.kernel.org/r/20230928020907.5730-2-michael.christie@oracle.com
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
Hannes Reinecke <hare@suse.de> says:
Hi all,
(taking up an old thread:) here's the first batch of patches for my EH
rework. It modifies the reset callbacks for SCSI drivers such that
the final conversion to drop the 'struct scsi_cmnd' argument and use
the entity in question (host, bus, target, device) as the argument to
the SCSI EH callbacks becomes possible. The first part covers drivers
which just requires minor tweaks.
Link: https://lore.kernel.org/r/20231002154328.43718-1-hare@suse.de
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
SCSI EH host reset is the final callback in the escalation chain; once we
reach this we need to reset the controller. As such it defeats the purpose
to skip controller reset if no I/Os are pending and the RAID device is to
be reset; especially after kexec there might be stale commands pending in
firmware for which we have no reference whatsoever. So this patch splits
off the check for pending I/O into a 'bus_reset' function, and leaves the
actual controller reset to the host reset.
Signed-off-by: Hannes Reinecke <hare@suse.de>
Link: https://lore.kernel.org/r/20231002154328.43718-19-hare@suse.de
Cc: Kashyap Desai <kashyap.desai@broadcom.com>
Cc: Sathya Prakash Veerichetty <sathya.prakash@broadcom.com>
Cc: Sumit Saxena <sumit.saxena@broadcom.com>
Cc: Sreekanth Reddy <sreekanth.reddy@broadcom.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Johannes Thumshirn <johannes.thumshirn@wdc.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
Clang warns (or errors with CONFIG_WERROR=y) several times along the
lines of:
drivers/scsi/ibmvscsi/ibmvfc.c:650:17: warning: implicit truncation from 'int' to a one-bit wide bit-field changes value from 1 to -1 [-Wsingle-bit-bitfield-constant-conversion]
650 | vhost->reinit = 1;
| ^ ~
A single-bit signed integer bitfield only has possible values of -1 and
0, not 0 and 1 like an unsigned one would. No context appears to check
the actual value of these bitfields, just whether or not it is zero.
However, it is easy enough to change the type of the fields to 'unsigned
int', which keeps the same size in memory and resolves the warning.
Fixes: 5144905884 ("scsi: ibmvfc: Use a bitfield for boolean flags")
Signed-off-by: Nathan Chancellor <nathan@kernel.org>
Link: https://lore.kernel.org/r/20231010-ibmvfc-fix-bitfields-type-v1-1-37e95b5a60e5@kernel.org
Reviewed-by: Nick Desaulniers <ndesaulniers@google.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
When no active_reqs, devfreq_monitor (thread A) will suspend clock scaling.
But it may have racing with clk_scaling.suspend_work (thread B) and
actually not suspend clock scaling (requeue after suspend). Next time
after polling_ms, devfreq_monitor read clk_scaling.window_start_t = 0 then
scale up clock abnormal.
Below is racing step:
devfreq->work (Thread A)
devfreq_monitor
update_devfreq
.....
ufshcd_devfreq_target
queue_work(hba->clk_scaling.workq,
1 &hba->clk_scaling.suspend_work)
.....
5 queue_delayed_work(devfreq_wq, &devfreq->work,
msecs_to_jiffies(devfreq->profile->polling_ms));
2 hba->clk_scaling.suspend_work (Thread B)
ufshcd_clk_scaling_suspend_work
__ufshcd_suspend_clkscaling
devfreq_suspend_device(hba->devfreq);
3 cancel_delayed_work_sync(&devfreq->work);
4 hba->clk_scaling.window_start_t = 0;
.....
Signed-off-by: Peter Wang <peter.wang@mediatek.com>
Link: https://lore.kernel.org/r/20230831130826.5592-4-peter.wang@mediatek.com
Reviewed-by: Bart Van Assche <bvanassche@acm.org>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
When ufshcd_clk_scaling_suspend_work (thread A) running and new command
coming, ufshcd_clk_scaling_start_busy (thread B) may get host_lock after
thread A first time release host_lock. Then thread A second time get
host_lock will set clk_scaling.window_start_t = 0 which scale up clock
abnormal next polling_ms time. Also inlines another
__ufshcd_suspend_clkscaling calls.
Below is racing step:
1 hba->clk_scaling.suspend_work (Thread A)
ufshcd_clk_scaling_suspend_work
2 spin_lock_irqsave(hba->host->host_lock, irq_flags);
3 hba->clk_scaling.is_suspended = true;
4 spin_unlock_irqrestore(hba->host->host_lock, irq_flags);
__ufshcd_suspend_clkscaling
7 spin_lock_irqsave(hba->host->host_lock, flags);
8 hba->clk_scaling.window_start_t = 0;
9 spin_unlock_irqrestore(hba->host->host_lock, flags);
ufshcd_send_command (Thread B)
ufshcd_clk_scaling_start_busy
5 spin_lock_irqsave(hba->host->host_lock, flags);
....
6 spin_unlock_irqrestore(hba->host->host_lock, flags);
Signed-off-by: Peter Wang <peter.wang@mediatek.com>
Link: https://lore.kernel.org/r/20230831130826.5592-3-peter.wang@mediatek.com
Reviewed-by: Bart Van Assche <bvanassche@acm.org>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
