On 1/26/23 02:06, Mike Snitzer wrote:
On Wed, Jan 25 2023 at 7:37P -0500,
Nikos Tsironis <ntsironis@xxxxxxxxxxx> wrote:
On 1/23/23 19:34, Mike Snitzer wrote:
On Thu, Jan 19 2023 at 4:36P -0500,
Nikos Tsironis <ntsironis@xxxxxxxxxxx> wrote:
On 1/18/23 18:28, Mike Snitzer wrote:
On Wed, Jan 18 2023 at 7:29P -0500,
Nikos Tsironis <ntsironis@xxxxxxxxxxx> wrote:
Under certain conditions, swapping a table, that includes a dm-era
target, with a new table, causes a deadlock.
This happens when a status (STATUSTYPE_INFO) or message IOCTL is blocked
in the suspended dm-era target.
dm-era executes all metadata operations in a worker thread, which stops
processing requests when the target is suspended, and resumes again when
the target is resumed.
So, running 'dmsetup status' or 'dmsetup message' for a suspended dm-era
device blocks, until the device is resumed.
If we then load a new table to the device, while the aforementioned
dmsetup command is blocked in dm-era, and resume the device, we
deadlock.
The problem is that the 'dmsetup status' and 'dmsetup message' commands
hold a reference to the live table, i.e., they hold an SRCU read lock on
md->io_barrier, while they are blocked.
When the device is resumed, the old table is replaced with the new one
by dm_swap_table(), which ends up calling synchronize_srcu() on
md->io_barrier.
Since the blocked dmsetup command is holding the SRCU read lock, and the
old table is never resumed, 'dmsetup resume' blocks too, and we have a
deadlock.
The only way to recover is by rebooting.
Steps to reproduce:
1. Create device with dm-era target
# dmsetup create eradev --table "0 1048576 era /dev/datavg/erameta /dev/datavg/eradata 8192"
2. Suspend the device
# dmsetup suspend eradev
3. Load new table to device, e.g., to resize the device. Note, that, we
must load the new table _after_ suspending the device to ensure the
constructor of the new target instance reads up-to-date metadata, as
committed by the post-suspend hook of dm-era.
# dmsetup load eradev --table "0 2097152 era /dev/datavg/erameta /dev/datavg/eradata 8192"
4. Device now has LIVE and INACTIVE tables
# dmsetup info eradev
Name: eradev
State: SUSPENDED
Read Ahead: 16384
Tables present: LIVE & INACTIVE
Open count: 0
Event number: 0
Major, minor: 252, 2
Number of targets: 1
5. Retrieve the status of the device. This blocks because the device is
suspended. Equivalently, any 'dmsetup message' operation would block
too. This command holds the SRCU read lock on md->io_barrier.
# dmsetup status eradev
I'll have a look at this flow, it seems to me we shouldn't stack up
'dmsetup status' and 'dmsetup message' commands if the table is
suspended.
I think it is unique to dm-era that you don't allow to _read_ metadata
operations while a device is suspended. But messages really shouldn't
be sent when the device is suspended. As-is DM is pretty silently
cutthroat about that constraint.
Resulting in deadlock is obviously cutthroat...
Hi Mike,
Thanks for the quick reply.
I couldn't find this constraint documented anywhere and since the
various DM targets seem to allow message operations while the device is
suspended I drew the wrong conclusion.
Thanks for clarifying this.
6. Resume the device. The resume operation tries to swap the old table
with the new one and deadlocks, because it synchronizes SRCU for the
old table, while the blocked 'dmsetup status' holds the SRCU read
lock. And the old table is never resumed again at this point.
# dmsetup resume eradev
7. The relevant dmesg logs are:
[ 7093.345486] dm-2: detected capacity change from 1048576 to 2097152
[ 7250.875665] INFO: task dmsetup:1986 blocked for more than 120 seconds.
[ 7250.875722] Not tainted 5.16.0-rc2-release+ #16
[ 7250.875756] "echo 0 > /proc/sys/kernel/hung_task_timeout_secs" disables this message.
[ 7250.875803] task:dmsetup state:D stack: 0 pid: 1986 ppid: 1313 flags:0x00000000
[ 7250.875809] Call Trace:
[ 7250.875812] <TASK>
[ 7250.875816] __schedule+0x330/0x8b0
[ 7250.875827] schedule+0x4e/0xc0
[ 7250.875831] schedule_timeout+0x20f/0x2e0
[ 7250.875836] ? do_set_pte+0xb8/0x120
[ 7250.875843] ? prep_new_page+0x91/0xa0
[ 7250.875847] wait_for_completion+0x8c/0xf0
[ 7250.875854] perform_rpc+0x95/0xb0 [dm_era]
[ 7250.875862] in_worker1.constprop.20+0x48/0x70 [dm_era]
[ 7250.875867] ? era_iterate_devices+0x30/0x30 [dm_era]
[ 7250.875872] ? era_status+0x64/0x1e0 [dm_era]
[ 7250.875877] era_status+0x64/0x1e0 [dm_era]
[ 7250.875882] ? dm_get_live_or_inactive_table.isra.11+0x20/0x20 [dm_mod]
[ 7250.875900] ? __mod_node_page_state+0x82/0xc0
[ 7250.875909] retrieve_status+0xbc/0x1e0 [dm_mod]
[ 7250.875921] ? dm_get_live_or_inactive_table.isra.11+0x20/0x20 [dm_mod]
[ 7250.875932] table_status+0x61/0xa0 [dm_mod]
[ 7250.875942] ctl_ioctl+0x1b5/0x4f0 [dm_mod]
[ 7250.875956] dm_ctl_ioctl+0xa/0x10 [dm_mod]
[ 7250.875966] __x64_sys_ioctl+0x8e/0xd0
[ 7250.875970] do_syscall_64+0x3a/0xd0
[ 7250.875974] entry_SYSCALL_64_after_hwframe+0x44/0xae
[ 7250.875980] RIP: 0033:0x7f20b7cd4017
[ 7250.875984] RSP: 002b:00007ffd443874b8 EFLAGS: 00000246 ORIG_RAX: 0000000000000010
[ 7250.875988] RAX: ffffffffffffffda RBX: 000055d69d6bd0e0 RCX: 00007f20b7cd4017
[ 7250.875991] RDX: 000055d69d6bd0e0 RSI: 00000000c138fd0c RDI: 0000000000000003
[ 7250.875993] RBP: 000000000000001e R08: 00007f20b81df648 R09: 00007ffd44387320
[ 7250.875996] R10: 00007f20b81deb53 R11: 0000000000000246 R12: 000055d69d6bd110
[ 7250.875998] R13: 00007f20b81deb53 R14: 000055d69d6bd000 R15: 0000000000000000
[ 7250.876002] </TASK>
[ 7250.876004] INFO: task dmsetup:1987 blocked for more than 120 seconds.
[ 7250.876046] Not tainted 5.16.0-rc2-release+ #16
[ 7250.876083] "echo 0 > /proc/sys/kernel/hung_task_timeout_secs" disables this message.
[ 7250.876129] task:dmsetup state:D stack: 0 pid: 1987 ppid: 1385 flags:0x00000000
[ 7250.876134] Call Trace:
[ 7250.876136] <TASK>
[ 7250.876138] __schedule+0x330/0x8b0
[ 7250.876142] schedule+0x4e/0xc0
[ 7250.876145] schedule_timeout+0x20f/0x2e0
[ 7250.876149] ? __queue_work+0x226/0x420
[ 7250.876156] wait_for_completion+0x8c/0xf0
[ 7250.876160] __synchronize_srcu.part.19+0x92/0xc0
[ 7250.876167] ? __bpf_trace_rcu_stall_warning+0x10/0x10
[ 7250.876173] ? dm_swap_table+0x2f4/0x310 [dm_mod]
[ 7250.876185] dm_swap_table+0x2f4/0x310 [dm_mod]
[ 7250.876198] ? table_load+0x360/0x360 [dm_mod]
[ 7250.876207] dev_suspend+0x95/0x250 [dm_mod]
[ 7250.876217] ctl_ioctl+0x1b5/0x4f0 [dm_mod]
[ 7250.876231] dm_ctl_ioctl+0xa/0x10 [dm_mod]
[ 7250.876240] __x64_sys_ioctl+0x8e/0xd0
[ 7250.876244] do_syscall_64+0x3a/0xd0
[ 7250.876247] entry_SYSCALL_64_after_hwframe+0x44/0xae
[ 7250.876252] RIP: 0033:0x7f15e9254017
[ 7250.876254] RSP: 002b:00007ffffdc59458 EFLAGS: 00000246 ORIG_RAX: 0000000000000010
[ 7250.876257] RAX: ffffffffffffffda RBX: 000055d4d99560e0 RCX: 00007f15e9254017
[ 7250.876260] RDX: 000055d4d99560e0 RSI: 00000000c138fd06 RDI: 0000000000000003
[ 7250.876261] RBP: 000000000000000f R08: 00007f15e975f648 R09: 00007ffffdc592c0
[ 7250.876263] R10: 00007f15e975eb53 R11: 0000000000000246 R12: 000055d4d9956110
[ 7250.876265] R13: 00007f15e975eb53 R14: 000055d4d9956000 R15: 0000000000000001
[ 7250.876269] </TASK>
Fix this by allowing metadata operations triggered by user space to run
in the corresponding user space thread, instead of queueing them for
execution by the dm-era worker thread.
Allowing them to run while the device is suspended is _not_ the
correct way to work-around this deadlock situation. I think it'd be
useful to understand why your userspace is tools are allowing these
messages and status to a device that is suspended.
Ack.
The sequence of operations I provided is just a way to easily reproduce
the deadlock. The exact issue I am facing is the following:
1. Create device with dm-era target
# dmsetup create eradev --table "0 1048576 era /dev/datavg/erameta /dev/datavg/eradata 8192"
2. Load new table to device, e.g., to resize the device or snapshot it.
# dmsetup load eradev --table "0 2097152 era /dev/datavg/erameta /dev/datavg/eradata 8192"
3. Suspend the device
# dmsetup suspend eradev
4. Someone else, e.g., a user or a monitoring daemon, runs an LVM
command at this point, e.g. 'vgs'.
5. 'vgs' tries to retrieve the status of the dm-era device using the
DM_TABLE_STATUS ioctl, and blocks, because the command is queued for
execution by the worker thread, which is suspended while the device
is suspended.
Note, that, LVM uses the DM_NOFLUSH_FLAG, but this doesn't make a
difference for dm-era, since the "problem" is not that it writes to
the metadata, but that it queues the metadata read operation to the
worker thread.
6. Resume the device: This deadlocks.
# dmsetup resume eradev
So, I am not the one retrieving the status of the suspended device. LVM
is. LVM, when running commands like 'lvs' and 'vgs', retrieves the
status of the devices on the system using the DM_TABLE_STATUS ioctl.
The deadlock is a race that happens when someone runs an LVM command at
the "wrong" time.
I think dm-era shouldn't be disallowing work items while suspended,
e.g.:
diff --git a/drivers/md/dm-era-target.c b/drivers/md/dm-era-target.c
index e92c1afc3677..33ea2c2374c7 100644
--- a/drivers/md/dm-era-target.c
+++ b/drivers/md/dm-era-target.c
@@ -1175,7 +1175,6 @@ struct era {
struct list_head rpc_calls;
struct digest digest;
- atomic_t suspended;
};
struct rpc {
@@ -1219,8 +1218,7 @@ static void remap_to_origin(struct era *era, struct bio *bio)
*--------------------------------------------------------------*/
static void wake_worker(struct era *era)
{
- if (!atomic_read(&era->suspended))
- queue_work(era->wq, &era->worker);
+ queue_work(era->wq, &era->worker);
}
static void process_old_eras(struct era *era)
@@ -1392,17 +1390,6 @@ static int in_worker1(struct era *era,
return perform_rpc(era, &rpc);
}
-static void start_worker(struct era *era)
-{
- atomic_set(&era->suspended, 0);
-}
-
-static void stop_worker(struct era *era)
-{
- atomic_set(&era->suspended, 1);
- drain_workqueue(era->wq);
-}
-
/*----------------------------------------------------------------
* Target methods
*--------------------------------------------------------------*/
@@ -1569,7 +1556,7 @@ static void era_postsuspend(struct dm_target *ti)
/* FIXME: fail mode */
}
- stop_worker(era);
+ drain_workqueue(era->wq);
r = metadata_commit(era->md);
if (r) {
@@ -1600,8 +1587,6 @@ static int era_preresume(struct dm_target *ti)
era->nr_blocks = new_size;
}
- start_worker(era);
-
r = in_worker0(era, metadata_era_rollover);
if (r) {
DMERR("%s: metadata_era_rollover failed", __func__);
During suspend it should certainly flush all works; but I fail to see
the value in disallowing the targets main way to do work (even while
suspended). Maybe Joe has insight on why dm-era was written this way.
But as an example dm-cache and dm-thin don't have such a strong
restriction.
The worker thread does the following:
1. Process old eras, i.e., digest the archived writesets in to the main
era array.
This doesn't involve a metadata commit, but writes to the metadata
2. Process deferred bios. This might trigger a metadata commit in
general, but when the device is suspended no bios should be reaching
the target, so it should be a no op.
3. Process RPC calls. This involves 'status' and 'message' operations.
process_rpc_calls() does commit the metadata, after running all RPC
calls. Checking if the device is suspended here is a way to prevent
the metadata commit.
But, some of the 'message' operations also commit the metadata, e.g.,
the metadata_take_snap message. I understand that no one should be
sending messages to a suspended device, but it's not enforced so it
could happen.
I believe the value of suspending the worker when the device is
suspended is to prevent anyone from writing to the metadata.
Writing to the metadata, while the device is suspended, could cause
problems. More on that in the following comments.
Using an appropriate LVM 'global_filter' is a way to prevent this, but:
1. This is just a workaround, not a proper solution.
2. This is not always an option. Imagine someone running an LVM command
in a container, for example. Or, we may not be allowed to change the
LVM configuration of the host at all.
FYI, status shouldn't trigger write IOs if the device is suspended.
Both dm-cache and dm-thin have this in status as a guard around its
work to ensure updated accounting (as a side-effect of committing
metadata), e.g.:
/* Commit to ensure statistics aren't out-of-date */
if (!(status_flags & DM_STATUS_NOFLUSH_FLAG) && !dm_suspended(ti))
(void) commit(cache, false);
Ack. The current behavior of dm-era wrt 'status' command is:
1. Queue the 'metadata_get_stats()' function for execution by the worker
thread.
2. The worker runs the function and retrieves the statistics
3. The worker commits the metadata _after_ retrieving the statistics
Shall I just change 'era_status()' to read the metadata directly and
skip the metadata commit, instead of queuing the operation to the worker
thread, when the device is suspended?
dm-era should still respect that the device is suspended (disallow
metadata commits, etc).
Probably should just do like the dm-cache code I quoted above (dm-thin
does the same): do not commit if suspended.
Seems to me that if simply retrieving stats in era_status() results in
metadata commit as a side-effect of its work that is pretty bogus --
process_rpc_calls() must be what does it just due to generic code?
If so, yes absolutely avoid that metadata commit (might we just update
process_rpc_calls() to check dm_suspended()?)
Yes, the metadata commit is due to the generic code in
process_rpc_calls(), and could be avoided if we check the device
suspended state here.
But, avoiding the metadata commit when the device is suspended is not
enough. No one should be writing to the metadata at all.
Else, we risk hitting the bug fixed by commit 9ae6e8b1c9bbf6 ("dm era:
commit metadata in postsuspend after worker stops").
In short, writing to the metadata of a suspended dm-era device results
in the committed metadata getting out-of-sync with the in-core metadata.
This could result in the corruption of the on-disk metadata if we load a
new dm-era table to the device, and destroy the old one (the destructor
flushes any cached writes to the disk).
For an exact sequence of events that can lead to this issue, please see
the commit mentioned above (9ae6e8b1c9bbf6).
Sorry for such fundamental concerns, I really should've caught this
stuff back when dm-era was introduced!
Really does feel like a read-write lock for metadata might be
warranted.
I think we would still need a lock to synchronize access to the
metadata, since, if I am not mistaken, the pre-resume hook that starts
the worker runs before clearing the DMF_SUSPENDED flag, and can run
concurrently with the 'status' IOCTL. So, the worker might run
concurrently with a 'status' operation that sees the device as
suspended.
I was initially thinking removing dm-era's suspended state (like above
patch) should take care of the pitfalls of disallowing works due to
that state.
But the dm-era could should be audited to make sure the big hammer of
disallowing works while suspended wasn't important for its
implementation (and the need to disallow device changes while
suspended).
I think a simple solution, without modifying dm-era too much, is:
1. Have status run in the user space thread, instead of queueing it to
the worker.
OK.
2. Add a mutex to avoid status running concurrently with the worker
thread.
Ideally you'd already have adequate protection.
You want to be sure you're protecting data and not just user vs worker.
I agree. Ideally we would use a read-write lock to protect access to the
metadata, like the rest of the DM targets do.
I proposed this locking scheme because it's really simple, and it
doesn't require many changes. I agree it's not ideal, but it could be a
first step, before introducing a more fine grained read-write lock.
3. Have status check if the device is suspended. If it is not, commit
the metadata before retrieving the status. If it is, skip the
metadata commit. This is in line with what the thin and cache targets
do.
Doing this avoids the deadlock in case someone, e.g., an LVM command,
runs a status operation against a suspended dm-era device.
User space could trigger the deadlock if it sends a message to a
suspended dm-era device, but, since this should never happen, it
shouldn't affect existing workflows.
Why not have message run from user thread too?
That's what this patch does, but I thought it was not the correct way to
solve the issue, that's why I didn't propose it again:
Allowing them to run while the device is suspended is _not_ the
correct way to work-around this deadlock situation.
But, probably, I misunderstood something.
Allowing message operations to run in the user thread, while the device
is suspended, would result in writing to the metadata of the suspended
device.
I understand that this should not happen.
We could fail message operations if the device is suspended (since all
of them write to the metadata), but isn't this considered a change to
the user facing API?
What if someone depends on the existing behavior of message operations
blocking if the dm-era device is suspended? Is it ok to change this
behavior?
Still, it might be a good idea to enforce this restriction, i.e.,
disallow running message operations on suspended devices, but, I guess
this requires more discussion.
It's all target specific. DM core shouldn't be the place to impose
this constraint.
Ack.
Nikos
--
dm-devel mailing list
dm-devel@xxxxxxxxxx
https://listman.redhat.com/mailman/listinfo/dm-devel
