zdb <dataset> fails with Remote I/O error when Multihost is enabled

[ofaaland]

System information

Type	Version/Name
Distribution Name
Distribution Version
Linux Kernel
Architecture
ZFS Version	master
SPL Version	master

Describe the problem you're observing

The command
zdb <dataset>
fails with "failed to own dataset ...: Remote I/O error" if multihost is enabled.
First reported in #6464

Describe how to reproduce the problem

zpool create poolname /dev/some/device
zpool set multihost=on poolname
zdb poolname/

Include any warning/errors/backtraces from the system logs

``
[faaland1@per zfs]$ sudo cmd/zpool/zpool set multihost=on mypool
[faaland1@per zfs]$ cmd/zdb/zdb mypool/ 2>&1 | head -n4
failed to own dataset 'mypool': Remote I/O error
zdb: can't open 'mypool': Remote I/O error

ZFS_DBGMSG(zdb):

[ofaaland]

@KnightKu proposed a patch in a comment

#6464 (comment)

diff --git a/cmd/zdb/zdb.c b/cmd/zdb/zdb.c
index 142968d..97adab1 100644
--- a/cmd/zdb/zdb.c
+++ b/cmd/zdb/zdb.c
@@ -5801,6 +5801,15 @@ main(int argc, char **argv)
                                }
                        }
                } else {
+                       /*
+                        * Disable the activity check to allow examination of
+                        * active pools.
+                        */
+                       mutex_enter(&spa_namespace_lock);
+                       if ((spa = spa_lookup(target)) != NULL) {
+                               spa->spa_import_flags |= ZFS_IMPORT_SKIP_MMP;
+                       }
+                       mutex_exit(&spa_namespace_lock);
                        error = open_objset(target, DMU_OST_ANY, FTAG, &os);
                        if (error == 0)
                                spa = dmu_objset_spa(os);

[ofaaland]

@KnightKu, that would work but the current requirement that every open/import path needs the ZFS_IMPORT_SKIP_MMP flag added seems to me to be too easy to miss, like I did here.

For example, if zdb is used to open a checkpoint with the current code, it will fail in the same way:

[faaland1@per zfs]$ cmd/zdb/zdb -k mypool 2>&1 | head -n4
zdb: Tried to read config of pool "mypool" but spa_get_stats() failed with error 121

I think some other method is needed that automatically covers user-space opens/imports in zdb and zhack.

[ofaaland]

@behlendorf , what would you think of a new global variable spa_open_global_flags whose contents are OR'd into spa->spa_import_flags? zdb and zhack could set this to ZFS_IMPORT_SKIP_MMP once and then all opens there would skip MMP activity test, but a FUSE implementation would not set this global (or at least, not this flag) and so the activity test would still occur when appropriate.

[ofaaland]

@behlendorf points out that my proposal has the side-effect of making it easier to accidentally disable the MMP activity test when it's not safe to do so, for example in zhack which opens the pool R/W.