zfs-patches/0004-Fix-zpl_mount-deadlock.patch

   1 From 0000000000000000000000000000000000000000 Mon Sep 17 00:00:00 2001
   2 From: Brian Behlendorf <behlendorf1@llnl.gov>
   3 Date: Wed, 11 Jul 2018 15:49:10 -0700
   4 Subject: [PATCH] Fix zpl_mount() deadlock
   5
   6 Commit 93b43af10 inadvertently introduced the following scenario which
   7 can result in a deadlock.  This issue was most easily reproduced by
   8 LXD containers using a ZFS storage backend but should be reproducible
   9 under any workload which is frequently mounting and unmounting.
  10
  11 -- THREAD A --
  12 spa_sync()
  13   spa_sync_upgrades()
  14     rrw_enter(&dp->dp_config_rwlock, RW_WRITER, FTAG); <- Waiting on B
  15
  16 -- THREAD B --
  17 mount_fs()
  18   zpl_mount()
  19     zpl_mount_impl()
  20       dmu_objset_hold()
  21         dmu_objset_hold_flags()
  22           dsl_pool_hold()
  23             dsl_pool_config_enter()
  24               rrw_enter(&dp->dp_config_rwlock, RW_READER, tag);
  25     sget()
  26       sget_userns()
  27         grab_super()
  28           down_write(&s->s_umount); <- Waiting on C
  29
  30 -- THREAD C --
  31 cleanup_mnt()
  32   deactivate_super()
  33     down_write(&s->s_umount);
  34     deactivate_locked_super()
  35       zpl_kill_sb()
  36         kill_anon_super()
  37           generic_shutdown_super()
  38             sync_filesystem()
  39               zpl_sync_fs()
  40                 zfs_sync()
  41                   zil_commit()
  42                     txg_wait_synced() <- Waiting on A
  43
  44 Reviewed by: Alek Pinchuk <apinchuk@datto.com>
  45 Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
  46 Closes #7598
  47 Closes #7659
  48 Closes #7691
  49 Closes #7693
  50
  51 (Cherry-picked from ac09630d8b0bf6c92084a30fdaefd03fd0adbdc1)
  52 Signed-off-by: Stoiko Ivanov <s.ivanov@proxmox.com>
  53 ---
  54  include/sys/zfs_vfsops.h |  1 +
  55  module/zfs/zpl_super.c   | 11 ++++++++++-
  56  2 files changed, 11 insertions(+), 1 deletion(-)
  57
  58 diff --git a/include/sys/zfs_vfsops.h b/include/sys/zfs_vfsops.h
  59 index 2326da42..927153b2 100644
  60 --- a/include/sys/zfs_vfsops.h
  61 +++ b/include/sys/zfs_vfsops.h
  62 @@ -32,6 +32,7 @@
  63  #include <sys/zil.h>
  64  #include <sys/sa.h>
  65  #include <sys/rrwlock.h>
  66 +#include <sys/dsl_dataset.h>
  67  #include <sys/zfs_ioctl.h>
  68
  69  #ifdef __cplusplus
  70 diff --git a/module/zfs/zpl_super.c b/module/zfs/zpl_super.c
  71 index fc10271b..5c426b0a 100644
  72 --- a/module/zfs/zpl_super.c
  73 +++ b/module/zfs/zpl_super.c
  74 @@ -271,8 +271,17 @@ zpl_mount_impl(struct file_system_type *fs_type, int flags, zfs_mnt_t *zm)
  75         if (err)
  76                 return (ERR_PTR(-err));
  77
  78 +       /*
  79 +        * The dsl pool lock must be released prior to calling sget().
  80 +        * It is possible sget() may block on the lock in grab_super()
  81 +        * while deactivate_super() holds that same lock and waits for
  82 +        * a txg sync.  If the dsl_pool lock is held over over sget()
  83 +        * this can prevent the pool sync and cause a deadlock.
  84 +        */
  85 +       dsl_pool_rele(dmu_objset_pool(os), FTAG);
  86         s = zpl_sget(fs_type, zpl_test_super, set_anon_super, flags, os);
  87 -       dmu_objset_rele(os, FTAG);
  88 +       dsl_dataset_rele(dmu_objset_ds(os), FTAG);
  89 +
  90         if (IS_ERR(s))
  91                 return (ERR_CAST(s));
  92