There are several storage drivers like dm-multipath, iscsi, tcmu-runner,
amd nbd that have userspace components that can run in the IO path. For
example, iscsi and nbd's userspace deamons may need to recreate a socket
and/or send IO on it, and dm-multipath's daemon multipathd may need to
send SG IO or read/write IO to figure out the state of paths and re-set
them up.
In the kernel these drivers have access to GFP_NOIO/GFP_NOFS and the
memalloc_*_save/restore functions to control the allocation behavior,
but for userspace we would end up hitting an allocation that ended up
writing data back to the same device we are trying to allocate for.
The device is then in a state of deadlock, because to execute IO the
device needs to allocate memory, but to allocate memory the memory
layers want execute IO to the device.
Here is an example with nbd using a local userspace daemon that performs
network IO to a remote server. We are using XFS on top of the nbd device,
but it can happen with any FS or other modules layered on top of the nbd
device that can write out data to free memory. Here a nbd daemon helper
thread, msgr-worker-1, is performing a write/sendmsg on a socket to execute
a request. This kicks off a reclaim operation which results in a WRITE to
the nbd device and the nbd thread calling back into the mm layer.
[ 1626.609191] msgr-worker-1 D 0 1026 1 0x00004000
[ 1626.609193] Call Trace:
[ 1626.609195] ? __schedule+0x29b/0x630
[ 1626.609197] ? wait_for_completion+0xe0/0x170
[ 1626.609198] schedule+0x30/0xb0
[ 1626.609200] schedule_timeout+0x1f6/0x2f0
[ 1626.609202] ? blk_finish_plug+0x21/0x2e
[ 1626.609204] ? _xfs_buf_ioapply+0x2e6/0x410
[ 1626.609206] ? wait_for_completion+0xe0/0x170
[ 1626.609208] wait_for_completion+0x108/0x170
[ 1626.609210] ? wake_up_q+0x70/0x70
[ 1626.609212] ? __xfs_buf_submit+0x12e/0x250
[ 1626.609214] ? xfs_bwrite+0x25/0x60
[ 1626.609215] xfs_buf_iowait+0x22/0xf0
[ 1626.609218] __xfs_buf_submit+0x12e/0x250
[ 1626.609220] xfs_bwrite+0x25/0x60
[ 1626.609222] xfs_reclaim_inode+0x2e8/0x310
[ 1626.609224] xfs_reclaim_inodes_ag+0x1b6/0x300
[ 1626.609227] xfs_reclaim_inodes_nr+0x31/0x40
[ 1626.609228] super_cache_scan+0x152/0x1a0
[ 1626.609231] do_shrink_slab+0x12c/0x2d0
[ 1626.609233] shrink_slab+0x9c/0x2a0
[ 1626.609235] shrink_node+0xd7/0x470
[ 1626.609237] do_try_to_free_pages+0xbf/0x380
[ 1626.609240] try_to_free_pages+0xd9/0x1f0
[ 1626.609245] __alloc_pages_slowpath+0x3a4/0xd30
[ 1626.609251] ? ___slab_alloc+0x238/0x560
[ 1626.609254] __alloc_pages_nodemask+0x30c/0x350
[ 1626.609259] skb_page_frag_refill+0x97/0xd0
[ 1626.609274] sk_page_frag_refill+0x1d/0x80
[ 1626.609279] tcp_sendmsg_locked+0x2bb/0xdd0
[ 1626.609304] tcp_sendmsg+0x27/0x40
[ 1626.609307] sock_sendmsg+0x54/0x60
[ 1626.609308] ___sys_sendmsg+0x29f/0x320
[ 1626.609313] ? sock_poll+0x66/0xb0
[ 1626.609318] ? ep_item_poll.isra.15+0x40/0xc0
[ 1626.609320] ? ep_send_events_proc+0xe6/0x230
[ 1626.609322] ? hrtimer_try_to_cancel+0x54/0xf0
[ 1626.609324] ? ep_read_events_proc+0xc0/0xc0
[ 1626.609326] ? _raw_write_unlock_irq+0xa/0x20
[ 1626.609327] ? ep_scan_ready_list.constprop.19+0x218/0x230
[ 1626.609329] ? __hrtimer_init+0xb0/0xb0
[ 1626.609331] ? _raw_spin_unlock_irq+0xa/0x20
[ 1626.609334] ? ep_poll+0x26c/0x4a0
[ 1626.609337] ? tcp_tsq_write.part.54+0xa0/0xa0
[ 1626.609339] ? release_sock+0x43/0x90
[ 1626.609341] ? _raw_spin_unlock_bh+0xa/0x20
[ 1626.609342] __sys_sendmsg+0x47/0x80
[ 1626.609347] do_syscall_64+0x5f/0x1c0
[ 1626.609349] ? prepare_exit_to_usermode+0x75/0xa0
[ 1626.609351] entry_SYSCALL_64_after_hwframe+0x44/0xa9
This patch adds a new prctl command that daemons can use after they have
done their initial setup, and before they start to do allocations that
are in the IO path. It sets the PF_MEMALLOC_NOIO and PF_LESS_THROTTLE
flags so both userspace block and FS threads can use it to avoid the
allocation recursion and try to prevent from being throttled while
writing out data to free up memory.
Signed-off-by: Mike Christie <[email protected]>
Acked-by: Michal Hocko <[email protected]>
Tested-by: Masato Suzuki <[email protected]>
Reviewed-by: Damien Le Moal <[email protected]>
Reviewed-by: Bart Van Assche <[email protected]>
Reviewed-by: Dave Chinner <[email protected]>
Reviewed-by: Darrick J. Wong <[email protected]>
Link: https://lore.kernel.org/r/[email protected]
Signed-off-by: Christian Brauner <[email protected]>
This fixes a regression added with:
commit e9e006f5fc
Author: Mike Christie <[email protected]>
Date: Sun Aug 4 14:10:06 2019 -0500
nbd: fix max number of supported devs
where we can deadlock during device shutdown. The problem occurs if
the recv_work's nbd_config_put occurs after nbd_start_device_ioctl has
returned and the userspace app has droppped its reference via closing
the device and running nbd_release. The recv_work nbd_config_put call
would then drop the refcount to zero and try to destroy the config which
would try to do destroy_workqueue from the recv work.
This patch just has nbd_start_device_ioctl do a flush_workqueue when it
wakes so we know after the ioctl returns running works have exited. This
also fixes a possible race where we could try to reuse the device while
old recv_works are still running.
Cc: [email protected]
Fixes: e9e006f5fc ("nbd: fix max number of supported devs")
Signed-off-by: Mike Christie <[email protected]>
Signed-off-by: Jens Axboe <[email protected]>
nbd requires socket families to support the shutdown method so the nbd
recv workqueue can be woken up from its sock_recvmsg call. If the socket
does not support the callout we will leave recv works running or get hangs
later when the device or module is removed.
This adds a check during socket connection/reconnection to make sure the
socket being passed in supports the needed callout.
Reported-by: [email protected]
Fixes: e9e006f5fc ("nbd: fix max number of supported devs")
Tested-by: Richard W.M. Jones <[email protected]>
Signed-off-by: Mike Christie <[email protected]>
Signed-off-by: Jens Axboe <[email protected]>
This fixes a bug added in 4.10 with commit:
commit 9561a7ade0
Author: Josef Bacik <[email protected]>
Date: Tue Nov 22 14:04:40 2016 -0500
nbd: add multi-connection support
that limited the number of devices to 256. Before the patch we could
create 1000s of devices, but the patch switched us from using our
own thread to using a work queue which has a default limit of 256
active works.
The problem is that our recv_work function sits in a loop until
disconnection but only handles IO for one connection. The work is
started when the connection is started/restarted, but if we end up
creating 257 or more connections, the queue_work call just queues
connection257+'s recv_work and that waits for connection 1 - 256's
recv_work to be disconnected and that work instance completing.
Instead of reverting back to kthreads, this has us allocate a
workqueue_struct per device, so we can block in the work.
Cc: [email protected]
Reviewed-by: Josef Bacik <[email protected]>
Signed-off-by: Mike Christie <[email protected]>
Signed-off-by: Jens Axboe <[email protected]>
This fixes a regression added in 4.9 with commit:
commit 0eadf37afc
Author: Josef Bacik <[email protected]>
Date: Thu Sep 8 12:33:40 2016 -0700
nbd: allow block mq to deal with timeouts
where before the patch userspace would set the timeout to 0 to disable
it. With the above patch, a zero timeout tells the block layer to use
the default value of 30 seconds. For setups where commands can take a
long time or experience transient issues like network disruptions this
then results in IO errors being sent to the application.
To fix this, the patch still uses the common block layer timeout
framework, but if zero is set, nbd just logs a message and then resets
the timer when it expires.
Reviewed-by: Josef Bacik <[email protected]>
Signed-off-by: Mike Christie <[email protected]>
Signed-off-by: Jens Axboe <[email protected]>
Fix bug added with the patch:
commit 8f3ea35929
Author: Josef Bacik <[email protected]>
Date: Mon Jul 16 12:11:35 2018 -0400
nbd: handle unexpected replies better
where if the timeout handler runs when the completion path is and we fail
to grab the mutex in the timeout handler we will leave a config reference
and cannot free the config later.
Reviewed-by: Josef Bacik <[email protected]>
Signed-off-by: Mike Christie <[email protected]>
Signed-off-by: Jens Axboe <[email protected]>
This adds a helper function to convert a block req op to a nbd cmd type.
It will be used in the last patch to log the type in the timeout
handler.
Reviewed-by: Josef Bacik <[email protected]>
Signed-off-by: Mike Christie <[email protected]>
Signed-off-by: Jens Axboe <[email protected]>
Add a helper to set the cmd timeout. It does not really do a lot now,
but will be more useful in the next patches.
Reviewed-by: Josef Bacik <[email protected]>
Signed-off-by: Mike Christie <[email protected]>
Signed-off-by: Jens Axboe <[email protected]>
If the device is setup with ioctl we can resize the device after the
initial setup, but if the device is setup with netlink we cannot use the
resize related ioctls and there is no netlink reconfigure size ATTR
handling code.
This patch adds netlink reconfigure resize support to match the ioctl
interface.
Reviewed-by: Josef Bacik <[email protected]>
Signed-off-by: Mike Christie <[email protected]>
Signed-off-by: Jens Axboe <[email protected]>
If iscsi_login_init_conn fails it can free conn_ops.
__iscsi_target_login_thread will then call iscsi_target_login_sess_out
which will also free it.
This fixes the problem by organizing conn allocation/setup into parts that
are needed through the life of the conn and parts that are only needed for
the login. The free functions then release what was allocated in the alloc
functions.
With this patch we have:
iscsit_alloc_conn/iscsit_free_conn - allocs/frees the conn we need for the
entire life of the conn.
iscsi_login_init_conn/iscsi_target_nego_release - allocs/frees the parts
of the conn that are only needed during login.
Signed-off-by: Mike Christie <[email protected]>
Signed-off-by: Martin K. Petersen <[email protected]>
The problem is that iscsi_login_zero_tsih_s1 sets conn->sess early in
iscsi_login_set_conn_values. If the function fails later like when we
alloc the idr it does kfree(sess) and leaves the conn->sess pointer set.
iscsi_login_zero_tsih_s1 then returns -Exyz and we then call
iscsi_target_login_sess_out and access the freed memory.
This patch has iscsi_login_zero_tsih_s1 either completely setup the
session or completely tear it down, so later in
iscsi_target_login_sess_out we can just check for it being set to the
connection.
Cc: [email protected]
Fixes: 0957627a99 ("iscsi-target: Fix sess allocation leak in...")
Signed-off-by: Mike Christie <[email protected]>
Acked-by: Martin K. Petersen <[email protected]>
Signed-off-by: Matthew Wilcox <[email protected]>
This converts drivers that were only calling transport_deregister_session
to use target_remove_session. The calling of
transport_deregister_session_configfs via target_remove_session for these
types of drivers is ok, because they were not exporting info from fields
like sess_acl_list, sess->se_tpg and sess->fabric_sess_ptr from configfs
accessible functions, so they will see no difference.
Signed-off-by: Mike Christie <[email protected]>
Reviewed-by: Bart Van Assche <[email protected]>
Reviewed-by: Christoph Hellwig <[email protected]>
Cc: Felipe Balbi <[email protected]>
Cc: Sebastian Andrzej Siewior <[email protected]>
Cc: Andrzej Pietrasiewicz <[email protected]>
Cc: Michael S. Tsirkin <[email protected]>
Cc: Juergen Gross <[email protected]>
Signed-off-by: Martin K. Petersen <[email protected]>
This converts tcm_fc to use target_remove_session
tcm_fc was calling transport_deregister_session_configfs then calling
transport_deregister_session when commands have completed. It should be ok
for it to call transport_deregister_session_configfs later via
target_remove_session because transport_deregister_session_configfs only
prevents access from configfs via tpg removal and its call to the
close_session callback for that driver, and this is already protected by
the ft_lport_lock and its port lookup handling.
Signed-off-by: Mike Christie <[email protected]>
Reviewed-by: Bart Van Assche <[email protected]>
Reviewed-by: Christoph Hellwig <[email protected]>
Cc: Johannes Thumshirn <[email protected]>
Signed-off-by: Martin K. Petersen <[email protected]>
Rename target_alloc_session to target_setup_session to avoid confusion with
the other transport session allocation function that only allocates the
session and because the target_alloc_session does so much more. It
allocates the session, sets up the nacl and registers the session.
The next patch will then add a remove function to match the setup in this
one, so it should make sense for all drivers, except iscsi, to just call
those 2 functions to setup and remove a session.
iscsi will continue to be the odd driver.
Signed-off-by: Mike Christie <[email protected]>
Reviewed-by: Bart Van Assche <[email protected]>
Reviewed-by: Christoph Hellwig <[email protected]>
Cc: Chris Boot <[email protected]>
Cc: Bryant G. Ly <[email protected]>
Cc: Michael Cyr <[email protected]>
Cc: <[email protected]>
Cc: Johannes Thumshirn <[email protected]>
Cc: Felipe Balbi <[email protected]>
Cc: Sebastian Andrzej Siewior <[email protected]>
Cc: Andrzej Pietrasiewicz <[email protected]>
Cc: Michael S. Tsirkin <[email protected]>
Cc: Juergen Gross <[email protected]>
Signed-off-by: Martin K. Petersen <[email protected]>
transport_init_session_tags is only called from target_core_transport.c so
make it static.
Signed-off-by: Mike Christie <[email protected]>
Reviewed-by: Bart Van Assche <[email protected]>
Reviewed-by: Christoph Hellwig <[email protected]>
Signed-off-by: Martin K. Petersen <[email protected]>
Just have iscsit_start_nopin_timer grab the lock and call
__iscsit_start_nopin_timer.
Signed-off-by: Mike Christie <[email protected]>
Reviewed-by: Christoph Hellwig <[email protected]>
Signed-off-by: Martin K. Petersen <[email protected]>
When __transport_register_session is called from transport_register_session
irqs will already have been disabled, so we do not want the unlock irq call
to enable them until the higher level has done the final
spin_unlock_irqrestore/ spin_unlock_irq.
This has __transport_register_session use the save/restore call.
Signed-off-by: Mike Christie <[email protected]>
Reviewed-by: Bart Van Assche <[email protected]>
Reviewed-by: Christoph Hellwig <[email protected]>
Signed-off-by: Martin K. Petersen <[email protected]>
We use unsigned long, size_t and u64 for dev_size. This has us standardize
on u64.
Signed-off-by: Mike Christie <[email protected]>
Reviewed-by: Xiubo Li <[email protected]>
Signed-off-by: Martin K. Petersen <[email protected]>
Instead of doing strdup and kstrto* just use match_int for dev params.
It will be ok to use int instead of unsigned long in tcmu_set_dev_attrib
because that is only being used for max sectors and block size and the
supported values for them are well under the max possible integer value.
Signed-off-by: Mike Christie <[email protected]>
Reviewed-by: Xiubo Li <[email protected]>
Signed-off-by: Martin K. Petersen <[email protected]>
This patch prevents a bug where data_bitmap is allocated in
tcmu_configure_device, userspace changes the max_blocks setting, the device
is mapped to a LUN, then we try to access the data_bitmap based on the new
max_blocks limit which may now be out of range.
To prevent this, we just check if data_bitmap has been setup. If it has
then we fail the max_blocks update operation.
Signed-off-by: Mike Christie <[email protected]>
Reviewed-by: Xiubo Li <[email protected]>
Signed-off-by: Martin K. Petersen <[email protected]>
The tcmu dev is added to the list of tcmu devices during configuration. At
this time the tcmu setup has completed, but lio core has not completed its
setup. The device is not yet usable so do not try to unmap blocks from it
Signed-off-by: Mike Christie <[email protected]>
Reviewed-by: Xiubo Li <[email protected]>
Signed-off-by: Martin K. Petersen <[email protected]>
Do not allow userspace to block or reset the ring until the device has been
configured. This will prevent the bug where userspace can write to those
files and access mb_addr before it has been setup.
Signed-off-by: Mike Christie <[email protected]>
Reviewed-by: Xiubo Li <[email protected]>
Signed-off-by: Martin K. Petersen <[email protected]>
Use the lio core helper to check if the device is configured.
Signed-off-by: Mike Christie <[email protected]>
Reviewed-by: Xiubo Li <[email protected]>
Signed-off-by: Martin K. Petersen <[email protected]>
This just adds a helper function to check if a device is configured and it
converts the target users to use it. The next patch will add a backend
module user so those types of modules do not have to know the lio core
details.
Signed-off-by: Mike Christie <[email protected]>
Reviewed-by: Xiubo Li <[email protected]>
Signed-off-by: Martin K. Petersen <[email protected]>
The caller of queue_cmd_ring grabs and releases the lock, so the
tcmu_setup_cmd_timer failure handling inside queue_cmd_ring should not call
mutex_unlock.
Signed-off-by: Mike Christie <[email protected]>
Reviewed-by: Xiubo Li <[email protected]>
Signed-off-by: Martin K. Petersen <[email protected]>
This patch fixes a bug where configfs_register_group had added
a group in a tree, and userspace has done a rmdir on a dir somewhere
above that group and we hit a kernel crash. The problem is configfs_rmdir
will detach everything under it and unlink groups on the default_groups
list. It will not unlink groups added with configfs_register_group so when
configfs_unregister_group is called to drop its references to the group/items
we crash when we try to access the freed dentrys.
The patch just adds a check for if a rmdir has been done above
us and if so just does the unlink part of unregistration.
Sorry if you are getting this multiple times. I thouhgt I sent
this to some of you and lkml, but I do not see it.
Signed-off-by: Mike Christie <[email protected]>
Cc: Christoph Hellwig <[email protected]>
Cc: Joel Becker <[email protected]>
Signed-off-by: Christoph Hellwig <[email protected]>
Fix warning:
smatch warnings:
drivers/target/target_core_user.c:301 tcmu_genl_cmd_done() warn: KERN_*
level not at start of string
Signed-off-by: Mike Christie <[email protected]>
Signed-off-by: Martin K. Petersen <[email protected]>
This patch based on Xiubo's patches adds 2 tcmu attr to block and reset the
netlink interface. It's used during userspace daemon reinitialization after
the daemon has crashed while there is outstanding nl requests. The daemon
can block the nl interface, kill outstanding requests in the kernel and
then reopen the netlink socket and unblock it to allow new requests.
[mkp: typo]
Signed-off-by: Mike Christie <[email protected]>
Tested-by: Xiubo Li <[email protected]>
Signed-off-by: Martin K. Petersen <[email protected]>
Some misc cleanup of the nl rework patches.
1. Fix space instead of tabs use and extra newline
2. Drop initializing variables to 0 when not needed
3. Just pass the skb_buff and msg_header pointers to
tcmu_netlink_event_send.
Signed-off-by: Mike Christie <[email protected]>
Tested-by: Xiubo Li <[email protected]>
Signed-off-by: Martin K. Petersen <[email protected]>
Just return EBUSY if a nl request comes in while processing one. The upper
layers do not support sending multiple create/remove requests at the same
time (you cannot have a create and remove at the same time or do multiple
creates or removes at the same time) and doing a reconfig while a
create/remove is still executing does not make sense.
Signed-off-by: Mike Christie <[email protected]>
Tested-by: Xiubo Li <[email protected]>
Signed-off-by: Martin K. Petersen <[email protected]>
The next patch is going to fix the hung nl command issue so this adds a
list of outstanding nl commands that we can later abort when the daemon is
restarted.
Signed-off-by: Mike Christie <[email protected]>
Signed-off-by: Martin K. Petersen <[email protected]>
If we are failing the command due to a qfull timeout we are
also freeing the tcmu command, so we cannot access it later
to get the se_cmd.
Note: The clearing of cmd->se_cmd is not needed. We do not check
it later for something like determining if the command was failed
due to a timeout. As a result I am dropping it.
Signed-off-by: Mike Christie <[email protected]>
Signed-off-by: Nicholas Bellinger <[email protected]>
This patch adds 2 tcmu attrs to block/unblock a device and
reset the ring buffer. They are used when the userspace
daemon has crashed or forced to shutdown while IO is executing.
On restart, the daemon can block the device so new IO is not
sent to userspace while it puts the ring in a clean state.
Notes: The reset ring opreation is specific to tcmu, but the
block one could be generic. I kept it tcmu specific, because
it requires some extra locking/state checks in the main IO
path and since other backend modules did not need this
functionality I thought only tcmu should take the perf hit.
Signed-off-by: Mike Christie <[email protected]>
Signed-off-by: Nicholas Bellinger <[email protected]>
This patch adds a new group of files that are to be used to
have the kernel module execution some action. The next patch
will have target_core_user use the group/files to be able to block
a device and to reset its memory buffer used to pass commands
between user/kernel space.
This type of file is different from the existing device attributes
in that they may be write only and when written to they result in
the kernel module executing some function. These need to be
separate from the normal device attributes which get/set device
values so userspace can continue to loop over all the attribs and
get/set them during initialization.
Signed-off-by: Mike Christie <[email protected]>
Signed-off-by: Nicholas Bellinger <[email protected]>
Log cmd id that was not found in the tcmu_handle_completions
lookup failure path.
Signed-off-by: Mike Christie <[email protected]>
Signed-off-by: Nicholas Bellinger <[email protected]>
Add SAM_STAT_BUSY sense_reason. The next patch will have
target_core_user return this value while it is temporarily
blocked and restarting.
Signed-off-by: Mike Christie <[email protected]>
Signed-off-by: Nicholas Bellinger <[email protected]>
We will always have a page mapped for cmd data if it is
valid command. If the mapping does not exist then something
bad happened in userspace and it should not proceed. This
has us return VM_FAULT_SIGBUS when this happens instead of
returning a freshly allocated paged. The latter can cause
corruption because userspace might write the pages data
overwriting valid data or return it to the initiator.
Signed-off-by: Mike Christie <[email protected]>
Signed-off-by: Nicholas Bellinger <[email protected]>
Users might have a physical system to a target so they could
have a lot more than 2 gigs of memory they want to devote to
tcmu. OTOH, we could be running in a vm and so a 2 gig
global and 1 gig per dev limit might be too high. This patch
allows the user to specify the limits.
Signed-off-by: Mike Christie <[email protected]>
Signed-off-by: Nicholas Bellinger <[email protected]>
This adds a timer, qfull_time_out, that controls how long a
device will wait for ring buffer space to open before
failing the commands in the queue. It is useful to separate
this timer from the cmd_time_out and default 30 sec one,
because for HA setups cmd_time_out may be disbled and 30
seconds is too long to wait when some OSs like ESX will
timeout commands after as little as 8 - 15 seconds.
Signed-off-by: Mike Christie <[email protected]>
Signed-off-by: Nicholas Bellinger <[email protected]>
This patch has tcmu internally queue cmds if its ring buffer
is full. It also makes the TCMU_GLOBAL_MAX_BLOCKS limit a
hint instead of a hard limit, so we do not have to add any
new locks/atomics in the main IO path except when IO is not
running.
This fixes the following bugs:
1. We cannot sleep from the submitting context because it might be
called from a target recv context. This results in transport level
commands timing out. For example if the ring is full, we would
sleep, and a iscsi initiator would send a iscsi ping/nop which
times out because the target's recv thread is sleeping here.
2. Devices were not fairly scheduled to run when they hit the global
limit so they could time out waiting for ring space while others
got run.
Signed-off-by: Mike Christie <[email protected]>
Signed-off-by: Nicholas Bellinger <[email protected]>
We do not really save a lot by trying to increase thresh
a multiple of the existing value. This just simplifies the
code by increasing it to whatever is needed for the command
being executed.
Signed-off-by: Mike Christie <[email protected]>
Signed-off-by: Nicholas Bellinger <[email protected]>
In the next patches we will call queue_cmd_ring from the submitting
context and also the completion path. This changes the queue_cmd_ring
return code so in the next patches we can return a sense_reason_t
and also signal if a command was requeued.
Signed-off-by: Mike Christie <[email protected]>
Signed-off-by: Nicholas Bellinger <[email protected]>
The blocks_left calculation does not account for free blocks
between 0 and thresh, so we could be queueing/waiting when
there are enough blocks free.
This has us add in the blocks between 0 and thresh as well as
at the end from thresh to DATA_BLOCK_BITS.
Signed-off-by: Mike Christie <[email protected]>
Signed-off-by: Nicholas Bellinger <[email protected]>
scatter_data_area always returns 0, so stop checking
for errors.
Signed-off-by: Mike Christie <[email protected]>
Signed-off-by: Nicholas Bellinger <[email protected]>
If we cannot setup a cmd because we run out of ring space
or global pages release the blocks before sleeping. This
prevents a deadlock where dev0 has waiting_blocks set and
needs N blocks, but dev1 to devX have each allocated N / X blocks
and also hit the global block limit so they went to sleep.
find_free_blocks is not able to take the sleeping dev's
blocks becaause their waiting_blocks is set and even
if it was not the block returned by find_last_bit could equal
dbi_max. The latter will probably never happen because
DATA_BLOCK_BITS is so high but in the next patches
DATA_BLOCK_BITS and TCMU_GLOBAL_MAX_BLOCKS will be settable so
it might be lower and could happen.
Signed-off-by: Mike Christie <[email protected]>
Signed-off-by: Nicholas Bellinger <[email protected]>
No need for the commands_lock. The cmdr_lock is already held during
idr addition and deletion, so just grab it during traversal.
Note: This also fixes a issue where we should have been using at
least _bh locking in tcmu_handle_completions when taking the commands
lock to prevent the case where tcmu_handle_completions could be
interrupted by a timer softirq while the commands_lock is held.
Signed-off-by: Mike Christie <[email protected]>
Signed-off-by: Nicholas Bellinger <[email protected]>
This moves the expired command completion handling to
the unmap wq, so the next patch can use a mutex
in tcmu_check_expired_cmd.
Note:
tcmu_device_timedout's use of spin_lock_irq was not needed.
The commands_lock is used between thread context (tcmu_queue_cmd_ring
and tcmu_irqcontrol (even though this is named irqcontrol it is not
run in irq context)) and timer/bh context. In the timer/bh context
bhs are disabled, so you need to use the _bh lock calls from the
thread context callers.
Signed-off-by: Mike Christie <[email protected]>
Signed-off-by: Nicholas Bellinger <[email protected]>
If the unmap thread has already run find_free_blocks
but not yet run prepare_to_wait when a wake_up(&unmap_wait)
call is done, the unmap thread is going to miss the wake
call. Instead of adding checks for if new waiters were added
this just has us use a work queue which will run us again
in this type of case.
Signed-off-by: Mike Christie <[email protected]>
Signed-off-by: Nicholas Bellinger <[email protected]>
Separate unmap_thread_fn to make it easier to read.
Note: this patch does not fix the bug where we might
miss a wake up call. The next patch will fix that.
This patch only separates the code into functions.
Signed-off-by: Mike Christie <[email protected]>
Signed-off-by: Nicholas Bellinger <[email protected]>
If cmd_time_out != 0, then tcmu_queue_cmd_ring could end up
sleeping waiting for ring space, timing out and then returning
failure to lio, and tcmu_check_expired_cmd could also detect
the timeout and call target_complete_cmd on the cmd.
This patch just delays setting up the deadline value and adding
the cmd to the udev->commands idr until we have allocated ring
space and are about to send the cmd to userspace.
Signed-off-by: Mike Christie <[email protected]>
Signed-off-by: Nicholas Bellinger <[email protected]>
TCM_OUT_OF_RESOURCES is getting translated to
TCM_LOGICAL_UNIT_COMMUNICATION_FAILURE which seems like a heavy
error when we just cannot allocate a resource that may be
allocatable later. This has us translate TCM_OUT_OF_RESOURCES
to SAM_STAT_TASK_SET_FULL instead.
Signed-off-by: Mike Christie <[email protected]>
Signed-off-by: Nicholas Bellinger <[email protected]>
Older kernels could crash or hang if the user write/read some ALUA files
with pscsi and tcmu backends. This patch exports if LIO supports
executing PGR and ALUA scsi commands/checks for the se_device, so userspace
can easily test.
Signed-off-by: Mike Christie <[email protected]>
Signed-off-by: Nicholas Bellinger <[email protected]>
If the scsi status was not SAM_STAT_GOOD or there was no transport
sense, we would ignore the scsi status and do a generic not ready
LUN communication failure check condition failure.
The problem is that LUN COMM failure is treated as a hard error
sometimes and will cause apps to get IO errors instead of the OS's SCSI
layer retrying. For example, the tcmu daemon will return SAM_STAT_QUEUE_FULL
when memory runs low and can still make progress but wants the initiator to
reduce the work load. Windows will fail this error directly the app
instead of retrying.
This patch is based on Nick's "target/iblock: Use -EAGAIN/-ENOMEM to
propigate SAM BUSY/TASK_SET_FULL" patch here:
http://comments.gmane.org/gmane.linux.scsi.target.devel/11301
but instead of only setting SAM_STAT_GOOD, SAM_STAT_TASK_SET_FULL
and SAM_STAT_BUSY as success, it sets all non check condition status
as success so they are passed back to the initiator, so passthrough
type backends can return all SCSI status codes. Since only passthrough
uses this, I was not sure if we wanted to add checks for non-passthrough
and specific codes.
Signed-off-by: Mike Christie <[email protected]>
Signed-off-by: Nicholas Bellinger <[email protected]>
transport_complete is no longer used, so drop the code.
Signed-off-by: Mike Christie <[email protected]>
Signed-off-by: Nicholas Bellinger <[email protected]>
This patch performs the pscsi_transport_complete operations from
pscsi_req_done. It looks like the only difference the transport_complete
callout provides is that it is called under t_state_lock which seems to
only be needed for the SCF_TRANSPORT_TASK_SENSE bit handling. We can
now use transport_copy_sense_to_cmd to handle the se_cmd sense bits, and
we can then drop the code where we have to copy the request info to
the pscsi_plugin_task for transport_complete use.
Signed-off-by: Mike Christie <[email protected]>
Signed-off-by: Nicholas Bellinger <[email protected]>
We were just copying the sense to the cmd sense_buffer and
did not implement a transport_complete or set the
SCF_TRANSPORT_TASK_SENSE, so the sense was ignored.
Signed-off-by: Mike Christie <[email protected]>
Signed-off-by: Nicholas Bellinger <[email protected]>
This adds a helper to copy sense from backend module buffer to
the se_cmd's sense buffer.
Signed-off-by: Mike Christie <[email protected]>
Signed-off-by: Nicholas Bellinger <[email protected]>
tcmu needs to pass raw sense to target_complete_cmd, but a a
transport_complete callout is akward to implement for it.
This moves the check for SCF_TRANSPORT_TASK_SENSE so any backend
can pass sense.
Signed-off-by: Mike Christie <[email protected]>
Signed-off-by: Nicholas Bellinger <[email protected]>
If the uio device is open and closed multiple times, the
kref count will be off due to tcmu_release getting called
multiple times for each close. This patch integrates
Wenji Tang's patch to add a kref_get on open that now
matches the kref_put done on tcmu_release and adds
a kref_put in tcmu_destroy_device to match the kref_get
done in succesful tcmu_configure_device calls.
Signed-off-by: Mike Christie <[email protected]>
Cc: Wenji Tang <[email protected]>
Signed-off-by: Nicholas Bellinger <[email protected]>
destroy_device is only called if we have successfully run
configure_device, so drop the duplicate tcmu_dev_configured check.
Signed-off-by: Mike Christie <[email protected]>
Signed-off-by: Nicholas Bellinger <[email protected]>
g_device_list is no longer needed because we now use the idr code
for lookups and seaches.
Signed-off-by: Mike Christie <[email protected]>
Reviewed-by: Bart Van Assche <[email protected]>
Signed-off-by: Nicholas Bellinger <[email protected]>
This converts the xcopy code to use the idr helper. The next patch
will drop the g_device_list and make g_device_mutex local to the
target_core_device.c file.
Signed-off-by: Mike Christie <[email protected]>
Reviewed-by: Bart Van Assche <[email protected]>
Signed-off-by: Nicholas Bellinger <[email protected]>
This adds a wrapper around idr_for_each so the xcopy code can loop over
the devices in the next patch.
Signed-off-by: Mike Christie <[email protected]>
Reviewed-by: Bart Van Assche <[email protected]>
Signed-off-by: Nicholas Bellinger <[email protected]>
This makes the device add, del reconfig operations sync. It fixes
the issue where for add and reconfig, we do not know if userspace
successfully completely the operation, so we leave invalid kernel
structs or report incorrect status for the config/reconfig operations.
Signed-off-by: Mike Christie <[email protected]>
Signed-off-by: Nicholas Bellinger <[email protected]>
This adds a helper to find a se_device by dev_index. It will
be used in the next patches so tcmu's netlink interface can
execute commands on specific devices.
Signed-off-by: Mike Christie <[email protected]>
Reviewed-by: Bart Van Assche <[email protected]>
Signed-off-by: Nicholas Bellinger <[email protected]>
In the next patches we will add tcmu netlink support that allows
userspace to send commands to target_core_user. To execute operations
on a se_device/tcmu_dev we need to be able to look up a dev by any old
id. This patch replaces the se_device->dev_index with a idr created
id.
The next patches will also remove the g_device_list and replace it with
the idr.
Signed-off-by: Mike Christie <[email protected]>
Signed-off-by: Nicholas Bellinger <[email protected]>
With this patch free_device is now used to free what is allocated in the
alloc_device callback and destroy_device tears down the resources that are
setup in the configure_device callback.
This patch will be needed in the next patch where tcmu needs
to be able to look up the device in the destroy callback.
Signed-off-by: Mike Christie <[email protected]>
Reviewed-by: Bart Van Assche <[email protected]>
Signed-off-by: Nicholas Bellinger <[email protected]>
1. TCMU_ATTR_TYPE is too generic when it describes only the
reconfiguration type, so rename to TCMU_ATTR_RECONFIG_TYPE.
2. Only return the reconfig type when it is a
TCMU_CMD_RECONFIG_DEVICE command.
3. CONFIG_* type is not needed. We can pass the value along with an
ATTR to userspace, so it does not need to read sysfs/configfs.
4. Fix leak in tcmu_dev_path_store and rename to dev_config to
reflect it is more than just a path that can be changed.
6. Don't update kernel struct value if netlink sending fails.
Signed-off-by: Mike Christie <[email protected]>
Reviewed-by: "Bryant G. Ly" <[email protected]>
Signed-off-by: Nicholas Bellinger <[email protected]>
We currently do
tcmu_free_device ->tcmu_netlink_event(TCMU_CMD_REMOVED_DEVICE) ->
uio_unregister_device -> kfree(tcmu_dev).
The problem is that the kernel does not wait for userspace to
do the close() on the uio device before freeing the tcmu_dev.
We can then hit a race where the kernel frees the tcmu_dev before
userspace does close() and so when close() -> release -> tcmu_release
is done, we try to access a freed tcmu_dev.
This patch made over the target-pending master branch moves the freeing
of the tcmu_dev to when the last reference has been dropped.
This also fixes a leak where if tcmu_configure_device was not called on a
device we did not free udev->name which was allocated at tcmu_alloc_device time.
Signed-off-by: Mike Christie <[email protected]>
Signed-off-by: Nicholas Bellinger <[email protected]>
We need to do a kthread_should_stop to check when kthread_stop has been
called.
This was a regression added in
b6df4b79a5
tcmu: Add global data block pool support
so not sure if you wanted to merge it in with that patch or what.
Signed-off-by: Mike Christie <[email protected]>
Signed-off-by: Nicholas Bellinger <[email protected]>
Multiple threads could be writing to alua_access_state at
the same time, or there could be multiple STPGs in flight
(different initiators sending them or one initiator sending
them to different ports), or a combo of both and the
core_alua_do_transition_tg_pt calls will race with each other.
Because from the last patches we no longer delay running
core_alua_do_transition_tg_pt_work, there does not seem to be
any point in running that in a workqueue. And, we always
wait for it to complete one way or another, so we can sleep
in this code path. So, this patch made over target-pending just adds a
mutex and does the work core_alua_do_transition_tg_pt_work was doing in
core_alua_do_transition_tg_pt.
There is also no need to use an atomic for the
tg_pt_gp_alua_access_state. In core_alua_do_transition_tg_pt we will
test and set it under the transition mutex. And, it is a int/32 bits
so in the other places where it is read, we will never see it partially
updated.
Signed-off-by: Mike Christie <[email protected]>
Signed-off-by: Nicholas Bellinger <[email protected]>
A single daemon could implement multiple types of devices
using multuple types of real devices that may not support
restarting from crashes and/or handling tcmu timeouts. This
makes the cmd timeout configurable, so handlers that do not
support it can turn if off for now.
Signed-off-by: Mike Christie <[email protected]>
Signed-off-by: Nicholas Bellinger <[email protected]>
This adds a helper to check if the dev was configured. It
will be used in the next patch to prevent updates to some
config settings after the device has been setup.
Signed-off-by: Mike Christie <[email protected]>
Signed-off-by: Nicholas Bellinger <[email protected]>
This fixes the following races:
1. core_alua_do_transition_tg_pt could have read
tg_pt_gp_alua_access_state and gone into this if chunk:
if (!explicit &&
atomic_read(&tg_pt_gp->tg_pt_gp_alua_access_state) ==
ALUA_ACCESS_STATE_TRANSITION) {
and then core_alua_do_transition_tg_pt_work could update the
state. core_alua_do_transition_tg_pt would then only set
tg_pt_gp_alua_pending_state and the tg_pt_gp_alua_access_state would
not get updated with the second calls state.
2. core_alua_do_transition_tg_pt could be setting
tg_pt_gp_transition_complete while the tg_pt_gp_transition_work
is already completing. core_alua_do_transition_tg_pt then waits on the
completion that will never be called.
To handle these issues, we just call flush_work which will return when
core_alua_do_transition_tg_pt_work has completed so there is no need
to do the complete/wait. And, if core_alua_do_transition_tg_pt_work
was running, instead of trying to sneak in the state change, we just
schedule up another core_alua_do_transition_tg_pt_work call.
Note that this does not handle a possible race where there are multiple
threads call core_alua_do_transition_tg_pt at the same time. I think
we need a mutex in target_tg_pt_gp_alua_access_state_store.
Signed-off-by: Mike Christie <[email protected]>
Signed-off-by: Nicholas Bellinger <[email protected]>
Userspace target_core_user handlers like tcmu-runner may want to set the
ALUA state to transitioning while it does implicit transitions. This
patch allows that state when set from configfs.
Signed-off-by: Mike Christie <[email protected]>
Signed-off-by: Nicholas Bellinger <[email protected]>
The implicit transition time tells initiators the min time
to wait before timing out a transition. We currently schedule
the transition to occur in tg_pt_gp_implicit_trans_secs
seconds so there is no room for delays. If
core_alua_do_transition_tg_pt_work->core_alua_update_tpg_primary_metadata
needs to write out info to a remote file, then the initiator can
easily time out the operation.
Signed-off-by: Mike Christie <[email protected]>
Signed-off-by: Nicholas Bellinger <[email protected]>
If tcmu-runner is processing a STPG and needs to change the kernel's
ALUA state then we cannot use the same work queue for task management
requests and ALUA transitions, because we could deadlock. The problem
occurs when a STPG times out before tcmu-runner is able to
call into target_tg_pt_gp_alua_access_state_store->
core_alua_do_port_transition -> core_alua_do_transition_tg_pt ->
queue_work. In this case, the tmr is on the work queue waiting for
the STPG to complete, but the STPG transition is now queued behind
the waiting tmr.
Note:
This bug will also be fixed by this patch:
http://www.spinics.net/lists/target-devel/msg14560.html
which switches the tmr code to use the system workqueues.
For both, I am not sure if we need a dedicated workqueue since
it is not a performance path and I do not think we need WQ_MEM_RECLAIM
to make forward progress to free up memory like the block layer does.
Signed-off-by: Mike Christie <[email protected]>
Signed-off-by: Nicholas Bellinger <[email protected]>
We do not setup the LU group for pscsi devices, so if you write
a state to alua_access_state that will cause a transition you will
get a NULL pointer dereference.
This patch will fail attempts to try and transition the path
for backend devices that set the TRANSPORT_FLAG_PASSTHROUGH_ALUA
flag.
Signed-off-by: Mike Christie <[email protected]>
Signed-off-by: Nicholas Bellinger <[email protected]>
This patch allows passthrough backends to use the core/base LIO
ALUA setup and state checks, but still handle the execution of
commands.
This will allow the target_core_user module to execute STPG and RTPG
in userspace, and not have to duplicate the ALUA state checks, path
information (needed so we can check if command is executable on
specific paths) and setup (rtslib sets/updates the configfs ALUA
interface like it does for iblock or file).
For STPG, the target_core_user userspace daemon, tcmu-runner will
still execute the STPG, and to update the core/base LIO state it
will use the existing configfs interface. For RTPG, tcmu-runner
will loop over configfs and/or cache the state.
Signed-off-by: Mike Christie <[email protected]>
Signed-off-by: Nicholas Bellinger <[email protected]>
We only were returing failure if the last opt to be parsed failed.
This has a return failure when we first detect a failure.
Signed-off-by: Mike Christie <[email protected]>
Signed-off-by: Nicholas Bellinger <[email protected]>
tcmu hard codes the hw_max_sectors to 128 which is a litle small.
Userspace uses the max_sectors to report the optimal IO size and
some initiators perform better with larger IOs (open-iscsi seems
to do better with 256 to 512 depending on the test).
(Fix do not display hw max sectors twice - MNC)
Signed-off-by: Mike Christie <[email protected]>
Signed-off-by: Nicholas Bellinger <[email protected]>
I think the transport statistics device file was supposed
to show scsiTransportProtocolType. It instead shows the
fabric name which is normally closer to the driver name.
I was thinking I cannot change from fabric name to protocol
type name incase people are expecting the driver name, so
this patch adds another file proto_id that exports the SCSI
protocol identifier ID.
Signed-off-by: Mike Christie <[email protected]>
Signed-off-by: Nicholas Bellinger <[email protected]>
This adds a force close option, so we can force the unmapping
of a rbd device that is open. If a path/device is blacklisted, apps
like multipathd can map a new device and then unmap the old one.
The unmapping cleanup would then be handled by the generic hotunplug
code paths in multipahd like is done for iSCSI, FC/FCOE, SAS, etc.
Signed-off-by: Mike Christie <[email protected]>
Signed-off-by: Ilya Dryomov <[email protected]>
Export the info used to setup the rbd image, so it can be used to remap
the image.
Signed-off-by: Mike Christie <[email protected]>
[[email protected]: do_rbd_add() EH]
Signed-off-by: Ilya Dryomov <[email protected]>
Export snap id in sysfs, so tools like multipathd can use it in a uuid.
Signed-off-by: Mike Christie <[email protected]>
Signed-off-by: Ilya Dryomov <[email protected]>
Export the cluster fsid, so tools like udev and multipath-tools can use
it for part of the uuid.
Signed-off-by: Mike Christie <[email protected]>
Signed-off-by: Ilya Dryomov <[email protected]>
The rw_page users were not converted to use bio/req ops. As a result
bdev_write_page is not passing down REQ_OP_WRITE and the IOs will
be sent down as reads.
Signed-off-by: Mike Christie <[email protected]>
Fixes: 4e1b2d52a8 ("block, fs, drivers: remove REQ_OP compat defs and related code")
Modified by me to:
1) Drop op_flags passing into ->rw_page(), as we don't use it.
2) Make op_is_write() and friends safe to use for !CONFIG_BLOCK
Signed-off-by: Jens Axboe <[email protected]>
max_discard_sectors only 32bits, and some non scsi backend
devices will set this to the max 0xffffffff, so we can end up
overflowing during the max_unmap_lba_count calculation.
This fixes a regression caused by my patch:
commit 8a9ebe717a
Author: Mike Christie <[email protected]>
Date: Mon Jan 18 14:09:27 2016 -0600
target: Fix WRITE_SAME/DISCARD conversion to linux 512b sectors
which can result in extra discards being sent to due the overflow
causing max_unmap_lba_count to be smaller than what the backing
device can actually support.
Signed-off-by: Mike Christie <[email protected]>
Reviewed-by: Bart Van Assche <[email protected]>
Cc: [email protected]
Signed-off-by: Nicholas Bellinger <[email protected]>
Because we define WRITE/READ as REQ_OPs, we cannot do
switch (rq_data_dir(request))
case READ
....
case WRITE
...
without getting warnings about handling other REQ_OPs.
This just has mq_disk do a if/else like it does in other
places.
Signed-off-by: Mike Christie <[email protected]>
Signed-off-by: Jens Axboe <[email protected]>
This was missed from my last patchset.
This patch has ext4 crypto code use the bio op helper
to set the operation. The operation (discard, write, writesame,
etc) is now defined seperately from the other REQ bits. They
still share the bi_rw field to save space, so we use these
helpers so modules do not have to worry about setting/overwriting
info.
Jens, I am not sure how you handle patches on top of patches
in the next branches. If you merge patches that fix issues
in previous patches in next, then this patch could be part
of
commit 95fe6c1a20
Author: Mike Christie <[email protected]>
Date: Sun Jun 5 14:31:48 2016 -0500
block, fs, mm, drivers: use bio set/get op accessors
Signed-off-by: Mike Christie <[email protected]>
Signed-off-by: Jens Axboe <[email protected]>
To avoid confusion between REQ_OP_FLUSH, which is handled by
request_fn drivers, and upper layers requesting the block layer
perform a flush sequence along with possibly a WRITE, this patch
renames REQ_FLUSH to REQ_PREFLUSH.
Signed-off-by: Mike Christie <[email protected]>
Reviewed-by: Christoph Hellwig <[email protected]>
Reviewed-by: Hannes Reinecke <[email protected]>
Signed-off-by: Jens Axboe <[email protected]>
The last patch added a REQ_OP_FLUSH for request_fn drivers
and the next patch renames REQ_FLUSH to REQ_PREFLUSH which
will be used by file systems and make_request_fn drivers so
they can send a write/flush combo.
This patch drops xen's use of REQ_FLUSH to track if it supports
REQ_OP_FLUSH requests, so REQ_FLUSH can be deleted.
Signed-off-by: Mike Christie <[email protected]>
Reviewed-by: Hannes Reinecke <[email protected]>
Signed-off-by: Juergen Gross <[email protected]>
Signed-off-by: Jens Axboe <[email protected]>
This adds a REQ_OP_FLUSH operation that is sent to request_fn
based drivers by the block layer's flush code, instead of
sending requests with the request->cmd_flags REQ_FLUSH bit set.
Signed-off-by: Mike Christie <[email protected]>
Reviewed-by: Christoph Hellwig <[email protected]>
Reviewed-by: Hannes Reinecke <[email protected]>
Signed-off-by: Jens Axboe <[email protected]>
This patch drops the compat definition of req_op where it matches
the rq_flag_bits definitions, and drops the related old and compat
code that allowed users to set either the op or flags for the operation.
We also then store the operation in the bi_rw/cmd_flags field similar
to how we used to store the bio ioprio where it sat in the upper bits
of the field.
Signed-off-by: Mike Christie <[email protected]>
Signed-off-by: Jens Axboe <[email protected]>
Mike Christie