We are only holding the lun_tg_pt_gp_lock in target_alua_state_check() to
make sure tg_pt_gp is not freed from under us while we copy the state,
delay, ID values. We can instead use RCU here to access the tg_pt_gp.
With this patch IOPs can increase up to 10% for jobs like:
fio --filename=/dev/sdX --direct=1 --rw=randrw --bs=4k \
--ioengine=libaio --iodepth=64 --numjobs=N
when there are multiple sessions (running that fio command to each /dev/sdX
or using multipath and there are over 8 paths), or more than 8 queues for
the loop or vhost with multiple threads case and numjobs > 8.
Link: https://lore.kernel.org/r/20210930020422.92578-5-michael.christie@oracle.com
Signed-off-by: Mike Christie <michael.christie@oracle.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
This patch fixes the following bugs:
1. If there are multiple ordered cmds queued and multiple simple cmds
completing, target_restart_delayed_cmds() could be called on different
CPUs and each instance could start a ordered cmd. They could then run in
different orders than they were queued.
2. target_restart_delayed_cmds() and target_handle_task_attr() can race
where:
1. target_handle_task_attr() has passed the simple_cmds == 0 check.
2. transport_complete_task_attr() then decrements simple_cmds to 0.
3. transport_complete_task_attr() runs target_restart_delayed_cmds() and
it does not see any cmds on the delayed_cmd_list.
4. target_handle_task_attr() adds the cmd to the delayed_cmd_list.
The cmd will then end up timing out.
3. If we are sent > 1 ordered cmds and simple_cmds == 0, we can execute
them out of order, because target_handle_task_attr() will hit that
simple_cmds check first and return false for all ordered cmds sent.
4. We run target_restart_delayed_cmds() after every cmd completion, so if
there is more than 1 simple cmd running, we start executing ordered cmds
after that first cmd instead of waiting for all of them to complete.
5. Ordered cmds are not supposed to start until HEAD OF QUEUE and all older
cmds have completed, and not just simple.
6. It's not a bug but it doesn't make sense to take the delayed_cmd_lock
for every cmd completion when ordered cmds are almost never used. Just
replacing that lock with an atomic increases IOPs by up to 10% when
completions are spread over multiple CPUs and there are multiple
sessions/ mqs/thread accessing the same device.
This patch moves the queued delayed handling to a per device work to
serialze the cmd executions for each device and adds a new counter to track
HEAD_OF_QUEUE and SIMPLE cmds. We can then check the new counter to
determine when to run the work on the completion path.
Link: https://lore.kernel.org/r/20210930020422.92578-3-michael.christie@oracle.com
Signed-off-by: Mike Christie <michael.christie@oracle.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
It may not always be best to complete the IO on same CPU as it was
submitted on. This commit allows userspace to configure it.
This has been useful for vhost-scsi where we have a single thread for
submissions and completions. If we force the completion on the submission
CPU we may be adding conflicts with what the user has setup in the lower
levels with settings like the block layer rq_affinity or the driver's IRQ
or softirq (the network's rps_cpus value) settings.
We may also want to set it up where the vhost thread runs on CPU N and does
its submissions/completions there, and then have LIO do its completion
booking on CPU M, but can't configure the lower levels due to issues like
using dm-multipath with lots of paths (the path selector can throw commands
all over the system because it's only taking into account latency/throughput
at its level).
The new setting is in:
/sys/kernel/config/target/$fabric/$target/param/cmd_completion_affinity
Writing:
-1 -> Gives the current default behavior of completing on the
submission CPU.
-2 -> Completes the cmd on the CPU the lower layers sent it to us from.
> 0 -> Completes on the CPU userspace has specified.
Link: https://lore.kernel.org/r/20210227170006.5077-26-michael.christie@oracle.com
Reviewed-by: Himanshu Madhani <himanshu.madhani@oracle.com>
Signed-off-by: Mike Christie <michael.christie@oracle.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
target_core_iblock is plugging and unplugging on every command and this is
causing perf issues for drivers that prefer batched cmds. With recent
patches we can now take multiple cmds from a fabric driver queue and then
pass them down the backend drivers in a batch. This patch adds this support
by adding 2 callouts to the backend for plugging and unplugging the
device. Subsequent commits will add support for iblock and tcmu device
plugging.
Link: https://lore.kernel.org/r/20210227170006.5077-22-michael.christie@oracle.com
Reviewed-by: Bodo Stroesser <bostroesser@gmail.com>
Signed-off-by: Mike Christie <michael.christie@oracle.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
loop and vhost/scsi do their target cmd submission from driver
workqueues. This allows them to avoid an issue where the backend may block
waiting for resources like tags/requests, mem/locks, etc and that ends up
blocking their entire submission path and for the case of vhost-scsi both
the submission and completion path.
This patch adds a helper drivers can use to submit from a LIO workqueue.
This code will then be extended in the next patches to fix the plugging of
backend devices.
We are only converting vhost/loop initially, but the workqueue based
submission will work for other drivers and have similar benefits where the
main target loops will not end up blocking one some backend resource.
Link: https://lore.kernel.org/r/20210227170006.5077-17-michael.christie@oracle.com
Tested-by: Laurence Oberman <loberman@redhat.com>
Reviewed-by: Bodo Stroesser <bostroesser@gmail.com>
Signed-off-by: Mike Christie <michael.christie@oracle.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
tcm_loop could be used like a normal block device, so we can't use
GFP_KERNEL and should use GFP_NOIO. This adds a gfp_t arg to
target_cmd_init_cdb() and converts the users. For every driver but loop
GFP_KERNEL is kept.
This will also be useful in subsequent patches where loop needs to do
target_submit_prep() from interrupt context to get a ref to the se_device,
and so it will need to use GFP_ATOMIC.
Link: https://lore.kernel.org/r/20210227170006.5077-16-michael.christie@oracle.com
Tested-by: Laurence Oberman <loberman@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Mike Christie <michael.christie@oracle.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
This breaks up target_submit_cmd_map_sgls() into 3 helpers:
- target_init_cmd(): Do the basic general setup and get a refcount to the
session to make sure the caller can execute the cmd.
- target_submit_prep(): Do the mapping, cdb processing and get a ref to
the LUN.
- target_submit(): Pass the cmd to LIO core for execution.
The above functions must be used by drivers that either:
1. Rely on LIO for session shutdown synchronization by calling
target_stop_session().
2. Need to map sgls.
When the next patches are applied then simple drivers that do not need the
extra functionality above can use target_submit_cmd() and not worry about
failures being returned and how to handle them, since many drivers were
getting this wrong and would have hit refcount bugs.
Also, by breaking target_submit_cmd_map_sgls() up into these 3 helper
functions, we can allow the later patches to do the init/prep from
interrupt context and then do the submission from a workqueue.
Link: https://lore.kernel.org/r/20210227170006.5077-5-michael.christie@oracle.com
Cc: Bart Van Assche <bvanassche@acm.org>
Cc: Juergen Gross <jgross@suse.com>
Cc: Hannes Reinecke <hare@suse.de>
Cc: Nilesh Javali <njavali@marvell.com>
Cc: Michael Cyr <mikecyr@linux.ibm.com>
Cc: Chris Boot <bootc@bootc.net>
Cc: Felipe Balbi <balbi@kernel.org>
Cc: "Michael S. Tsirkin" <mst@redhat.com>
Cc: Stefan Hajnoczi <stefanha@redhat.com>
Tested-by: Laurence Oberman <loberman@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Himanshu Madhani <himanshu.madhani@oracle.com>
Signed-off-by: Mike Christie <michael.christie@oracle.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
TCM doesn't properly handle underflow case for service actions. One way to
prevent it is to always complete command with
target_complete_cmd_with_length(), however it requires access to data_sg,
which is not always available.
This change introduces target_set_cmd_data_length() function which allows
to set command data length before completing it.
Link: https://lore.kernel.org/r/20210209072202.41154-2-a.miloserdov@yadro.com
Reviewed-by: Roman Bolshakov <r.bolshakov@yadro.com>
Reviewed-by: Bodo Stroesser <bostroesser@gmail.com>
Signed-off-by: Aleksandr Miloserdov <a.miloserdov@yadro.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
According to FCP-4 (9.4.2):
If the command requested that data beyond the length specified by the
FCP_DL field be transferred, then the device server shall set the
FCP_RESID_OVER bit (see 9.5.8) to one in the FCP_RSP IU and:
a) process the command normally except that data beyond the FCP_DL count
shall not be requested or transferred;
b) transfer no data and return CHECK CONDITION status with the sense key
set to ILLEGAL REQUEST and the additional sense code set to INVALID FIELD
IN COMMAND INFORMATION UNIT; or
c) may transfer data and return CHECK CONDITION status with the sense key
set to ABORTED COMMAND and the additional sense code set to INVALID FIELD
IN COMMAND INFORMATION UNIT.
TCM follows b) and transfers no data for residual writes but returns
INVALID FIELD IN CDB instead of INVALID FIELD IN COMMAND INFORMATION UNIT.
Change the ASCQ to INVALID FIELD IN COMMAND INFORMATION UNIT to meet the
standard.
Link: https://lore.kernel.org/r/20201203082035.54566-4-a.kovaleva@yadro.com
Signed-off-by: Anastasia Kovaleva <a.kovaleva@yadro.com>
Signed-off-by: Roman Bolshakov <r.bolshakov@yadro.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
cmd.bad_sector currently gets packed into the sense INFORMATION field for
TCM_LOGICAL_BLOCK_{GUARD,APP_TAG,REF_TAG}_CHECK_FAILED errors, which carry
an .add_sector_info flag in the sense_detail_table to ensure this.
In preparation for propagating a byte offset on COMPARE AND WRITE
TCM_MISCOMPARE_VERIFY error, rename cmd.bad_sector to cmd.sense_info and
sense_detail.add_sector_info to sense_detail.add_sense_info so that it
better reflects the sense INFORMATION field destination.
[ddiss: update previously overlooked ib_isert]
Link: https://lore.kernel.org/r/20201031233211.5207-3-ddiss@suse.de
Reviewed-by: Mike Christie <michael.christie@oracle.com>
Signed-off-by: David Disseldorp <ddiss@suse.de>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
iscsi_target_sk_data_ready() could be invoked indirectly by
iscsi_target_do_login_rx() from the workqueue like this:
iscsi_target_do_login_rx()
iscsi_target_do_login()
iscsi_target_do_tx_login_io()
iscsit_put_login_tx()
iscsi_login_tx_data()
tx_data()
sock_sendmsg_nosec()
tcp_sendmsg()
release_sock()
sk_backlog_rcv()
tcp_v4_do_rcv()
tcp_data_ready()
iscsi_target_sk_data_ready()
At that time LOGIN_FLAGS_READ_ACTIVE is not cleared and
iscsi_target_sk_data_ready will not read data from the socket. Some iscsi
initiators (libiscsi) will wait forever for a reply.
LOGIN_FLAGS_READ_ACTIVE should be cleared early just after doing the
receive and before writing to the socket in iscsi_target_do_login_rx.
Unfortunately, LOGIN_FLAGS_READ_ACTIVE is also used by sk_state_change to
do login cleanup if a socket was closed at login time. It is supposed to be
cleared after the login PDU is successfully processed and replied.
Introduce another flag, LOGIN_FLAGS_WRITE_ACTIVE, to cover the transmit
part.
Link: https://lore.kernel.org/r/20200716100212.4237-2-houpu@bytedance.com
Reviewed-by: Mike Christie <michael.christie@oracle.com>
Signed-off-by: Hou Pu <houpu@bytedance.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
