Btrfs: fix relocation incorrectly dropping data references

During relocation of a data block group we create a relocation tree
for each fs/subvol tree by making a snapshot of each tree using
btrfs_copy_root() and the tree's commit root, and then setting the last
snapshot field for the fs/subvol tree's root to the value of the current
transaction id minus 1. However this can lead to relocation later
dropping references that it did not create if we have qgroups enabled,
leaving the filesystem in an inconsistent state that keeps aborting
transactions.

Lets consider the following example to explain the problem, which requires
qgroups to be enabled.

We are relocating data block group Y, we have a subvolume with id 258 that
has a root at level 1, that subvolume is used to store directory entries
for snapshots and we are currently at transaction 3404.

When committing transaction 3404, we have a pending snapshot and therefore
we call btrfs_run_delayed_items() at transaction.c:create_pending_snapshot()
in order to create its dentry at subvolume 258. This results in COWing
leaf A from root 258 in order to add the dentry. Note that leaf A
also contains file extent items referring to extents from some other
block group X (we are currently relocating block group Y). Later on, still
at create_pending_snapshot() we call qgroup_account_snapshot(), which
switches the commit root for root 258 when it calls switch_commit_roots(),
so now the COWed version of leaf A, lets call it leaf A', is accessible
from the commit root of tree 258. At the end of qgroup_account_snapshot(),
we call record_root_in_trans() with 258 as its argument, which results
in btrfs_init_reloc_root() being called, which in turn calls
relocation.c:create_reloc_root() in order to create a relocation tree
associated to root 258, which results in assigning the value of 3403
(which is the current transaction id minus 1 = 3404 - 1) to the
last_snapshot field of root 258. When creating the relocation tree root
at ctree.c:btrfs_copy_root() we add a shared reference for leaf A',
corresponding to the relocation tree's root, when we call btrfs_inc_ref()
against the COWed root (a copy of the commit root from tree 258), which
is at level 1. So at this point leaf A' has 2 references, one normal
reference corresponding to root 258 and one shared reference corresponding
to the root of the relocation tree.

Transaction 3404 finishes its commit and transaction 3405 is started by
relocation when calling merge_reloc_root() for the relocation tree
associated to root 258. In the meanwhile leaf A' is COWed again, in
response to some filesystem operation, when we are still at transaction
3405. However when we COW leaf A', at ctree.c:update_ref_for_cow(), we
call btrfs_block_can_be_shared() in order to figure out if other trees
refer to the leaf and if any such trees exists, add a full back reference
to leaf A' - but btrfs_block_can_be_shared() incorrectly returns false
because the following condition is false:

  btrfs_header_generation(buf) <= btrfs_root_last_snapshot(&root->root_item)

which evaluates to 3404 <= 3403. So after leaf A' is COWed, it stays with
only one reference, corresponding to the shared reference we created when
we called btrfs_copy_root() to create the relocation tree's root and
btrfs_inc_ref() ends up not being called for leaf A' nor we end up setting
the flag BTRFS_BLOCK_FLAG_FULL_BACKREF in leaf A'. This results in not
adding shared references for the extents from block group X that leaf A'
refers to with its file extent items.

Later, after merging the relocation root we do a call to to
btrfs_drop_snapshot() in order to delete the relocation tree. This ends
up calling do_walk_down() when path->slots[1] points to leaf A', which
results in calling btrfs_lookup_extent_info() to get the number of
references for leaf A', which is 1 at this time (only the shared reference
exists) and this value is stored at wc->refs[0]. After this walk_up_proc()
is called when wc->level is 0 and path->nodes[0] corresponds to leaf A'.
Because the current level is 0 and wc->refs[0] is 1, it does call
btrfs_dec_ref() against leaf A', which results in removing the single
references that the extents from block group X have which are associated
to root 258 - the expectation was to have each of these extents with 2
references - one reference for root 258 and one shared reference related
to the root of the relocation tree, and so we would drop only the shared
reference (because leaf A' was supposed to have the flag
BTRFS_BLOCK_FLAG_FULL_BACKREF set).

This leaves the filesystem in an inconsistent state as we now have file
extent items in a subvolume tree that point to extents from block group X
without references in the extent tree. So later on when we try to decrement
the references for these extents, for example due to a file unlink operation,
truncate operation or overwriting ranges of a file, we fail because the
expected references do not exist in the extent tree.

This leads to warnings and transaction aborts like the following:

[  588.965795] ------------[ cut here ]------------
[  588.965815] WARNING: CPU: 2 PID: 2479 at fs/btrfs/extent-tree.c:1625 lookup_inline_extent_backref+0x432/0x5b0 [btrfs]
[  588.965816] Modules linked in: af_packet iscsi_ibft iscsi_boot_sysfs xfs libcrc32c ppdev acpi_cpufreq button tpm_tis e1000 i2c_piix4 pcspkr parport_pc
parport tpm qemu_fw_cfg joydev btrfs xor raid6_pq sr_mod cdrom ata_generic virtio_scsi ata_piix virtio_pci bochs_drm virtio_ring drm_kms_helper syscopyarea
sysfillrect sysimgblt fb_sys_fops virtio ttm serio_raw drm floppy sg
[  588.965831] CPU: 2 PID: 2479 Comm: kworker/u8:7 Not tainted 4.7.3-3-default-fdm+ #1
[  588.965832] Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS rel-1.9.1-0-gb3ef39f-prebuilt.qemu-project.org 04/01/2014
[  588.965844] Workqueue: btrfs-extent-refs btrfs_extent_refs_helper [btrfs]
[  588.965845]  0000000000000000 ffff8802263bfa28 ffffffff813af542 0000000000000000
[  588.965847]  0000000000000000 ffff8802263bfa68 ffffffff81081e8b 0000065900000000
[  588.965848]  ffff8801db2af000 000000012bbe2000 0000000000000000 ffff880215703b48
[  588.965849] Call Trace:
[  588.965852]  [<ffffffff813af542>] dump_stack+0x63/0x81
[  588.965854]  [<ffffffff81081e8b>] __warn+0xcb/0xf0
[  588.965855]  [<ffffffff81081f7d>] warn_slowpath_null+0x1d/0x20
[  588.965863]  [<ffffffffa0175042>] lookup_inline_extent_backref+0x432/0x5b0 [btrfs]
[  588.965865]  [<ffffffff81143220>] ? trace_clock_local+0x10/0x30
[  588.965867]  [<ffffffff8114c5df>] ? rb_reserve_next_event+0x6f/0x460
[  588.965875]  [<ffffffffa0175215>] insert_inline_extent_backref+0x55/0xd0 [btrfs]
[  588.965882]  [<ffffffffa017531f>] __btrfs_inc_extent_ref.isra.55+0x8f/0x240 [btrfs]
[  588.965890]  [<ffffffffa017acea>] __btrfs_run_delayed_refs+0x74a/0x1260 [btrfs]
[  588.965892]  [<ffffffff810cb046>] ? cpuacct_charge+0x86/0xa0
[  588.965900]  [<ffffffffa017e74f>] btrfs_run_delayed_refs+0x9f/0x2c0 [btrfs]
[  588.965908]  [<ffffffffa017ea04>] delayed_ref_async_start+0x94/0xb0 [btrfs]
[  588.965918]  [<ffffffffa01c799a>] btrfs_scrubparity_helper+0xca/0x350 [btrfs]
[  588.965928]  [<ffffffffa01c7c5e>] btrfs_extent_refs_helper+0xe/0x10 [btrfs]
[  588.965930]  [<ffffffff8109b323>] process_one_work+0x1f3/0x4e0
[  588.965931]  [<ffffffff8109b658>] worker_thread+0x48/0x4e0
[  588.965932]  [<ffffffff8109b610>] ? process_one_work+0x4e0/0x4e0
[  588.965934]  [<ffffffff810a1659>] kthread+0xc9/0xe0
[  588.965936]  [<ffffffff816f2f1f>] ret_from_fork+0x1f/0x40
[  588.965937]  [<ffffffff810a1590>] ? kthread_worker_fn+0x170/0x170
[  588.965938] ---[ end trace 34e5232c933a1749 ]---
[  588.966187] ------------[ cut here ]------------
[  588.966196] WARNING: CPU: 2 PID: 2479 at fs/btrfs/extent-tree.c:2966 btrfs_run_delayed_refs+0x28c/0x2c0 [btrfs]
[  588.966196] BTRFS: Transaction aborted (error -5)
[  588.966197] Modules linked in: af_packet iscsi_ibft iscsi_boot_sysfs xfs libcrc32c ppdev acpi_cpufreq button tpm_tis e1000 i2c_piix4 pcspkr parport_pc
parport tpm qemu_fw_cfg joydev btrfs xor raid6_pq sr_mod cdrom ata_generic virtio_scsi ata_piix virtio_pci bochs_drm virtio_ring drm_kms_helper syscopyarea
sysfillrect sysimgblt fb_sys_fops virtio ttm serio_raw drm floppy sg
[  588.966206] CPU: 2 PID: 2479 Comm: kworker/u8:7 Tainted: G        W       4.7.3-3-default-fdm+ #1
[  588.966207] Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS rel-1.9.1-0-gb3ef39f-prebuilt.qemu-project.org 04/01/2014
[  588.966217] Workqueue: btrfs-extent-refs btrfs_extent_refs_helper [btrfs]
[  588.966217]  0000000000000000 ffff8802263bfc98 ffffffff813af542 ffff8802263bfce8
[  588.966219]  0000000000000000 ffff8802263bfcd8 ffffffff81081e8b 00000b96345ee000
[  588.966220]  ffffffffa021ae1c ffff880215703b48 00000000000005fe ffff8802345ee000
[  588.966221] Call Trace:
[  588.966223]  [<ffffffff813af542>] dump_stack+0x63/0x81
[  588.966224]  [<ffffffff81081e8b>] __warn+0xcb/0xf0
[  588.966225]  [<ffffffff81081eff>] warn_slowpath_fmt+0x4f/0x60
[  588.966233]  [<ffffffffa017e93c>] btrfs_run_delayed_refs+0x28c/0x2c0 [btrfs]
[  588.966241]  [<ffffffffa017ea04>] delayed_ref_async_start+0x94/0xb0 [btrfs]
[  588.966250]  [<ffffffffa01c799a>] btrfs_scrubparity_helper+0xca/0x350 [btrfs]
[  588.966259]  [<ffffffffa01c7c5e>] btrfs_extent_refs_helper+0xe/0x10 [btrfs]
[  588.966260]  [<ffffffff8109b323>] process_one_work+0x1f3/0x4e0
[  588.966261]  [<ffffffff8109b658>] worker_thread+0x48/0x4e0
[  588.966263]  [<ffffffff8109b610>] ? process_one_work+0x4e0/0x4e0
[  588.966264]  [<ffffffff810a1659>] kthread+0xc9/0xe0
[  588.966265]  [<ffffffff816f2f1f>] ret_from_fork+0x1f/0x40
[  588.966267]  [<ffffffff810a1590>] ? kthread_worker_fn+0x170/0x170
[  588.966268] ---[ end trace 34e5232c933a174a ]---
[  588.966269] BTRFS: error (device sda2) in btrfs_run_delayed_refs:2966: errno=-5 IO failure
[  588.966270] BTRFS info (device sda2): forced readonly

This was happening often on openSUSE and SLE systems using btrfs as the
root filesystem (with its default layout where multiple subvolumes are
used) where balance happens in the background triggered by a cron job and
snapshots are automatically created before/after package installations,
upgrades and removals. The issue could be triggered simply by running the
following loop on the first system boot post installation:

  while true; do
     zypper -n in nfs-kernel-server
     zypper -n rm nfs-kernel-server
  done

(If we were fast enough and made that loop before the cron job triggered
a balance operation and the balance finished)

So fix by setting the last_snapshot field of the root to the value of the
generation of its commit root. Like this btrfs_block_can_be_shared()
behaves correctly for the case where the relocation root is created during
a transaction commit and for the case where it's created before a
transaction commit.

Fixes: 6426c7ad (btrfs: qgroup: Fix qgroup accounting when creating snapshot)
Cc: stable@vger.kernel.org  # 4.7+
Signed-off-by: Filipe Manana <fdmanana@suse.com>
Reviewed-by: Josef Bacik <jbacik@fb.com>