GitLab

A rare BUG_ON triggered in assoc_array_gc:

    [3430308.818153] kernel BUG at lib/assoc_array.c:1609!

Which corresponded to the statement currently at line 1593 upstream:

    BUG_ON(assoc_array_ptr_is_meta(p));

Using the data from the core dump, I was able to generate a userspace
reproducer[1] and determine the cause of the bug.

[1]: https://github.com/brenns10/kernel_stuff/tree/master/assoc_array_gc

After running the iterator on the entire branch, an internal tree node
looked like the following:

    NODE (nr_leaves_on_branch: 3)
      SLOT [0] NODE (2 leaves)
      SLOT [1] NODE (1 leaf)
      SLOT [2..f] NODE (empty)

In the userspace reproducer, the pr_devel output when compressing this
node was:

    -- compress node 0x5607cc089380 --
    free=0, leaves=0
    [0] retain node 2/1 [nx 0]
    [1] fold node 1/1 [nx 0]
    [2] fold node 0/1 [nx 2]
    [3] fold node 0/2 [nx 2]
    [4] fold node 0/3 [nx 2]
    [5] fold node 0/4 [nx 2]
    [6] fold node 0/5 [nx 2]
    [7] fold node 0/6 [nx 2]
    [8] fold node 0/7 [nx 2]
    [9] fold node 0/8 [nx 2]
    [10] fold node 0/9 [nx 2]
    [11] fold node 0/10 [nx 2]
    [12] fold node 0/11 [nx 2]
    [13] fold node 0/12 [nx 2]
    [14] fold node 0/13 [nx 2]
    [15] fold node 0/14 [nx 2]
    after: 3

At slot 0, an internal node with 2 leaves could not be folded into the
node, because there was only one available slot (slot 0). Thus, the
internal node was retained. At slot 1, the node had one leaf, and was
able to be folded in successfully. The remaining nodes had no leaves,
and so were removed. By the end of the compression stage, there were 14
free slots, and only 3 leaf nodes. The tree was ascended and then its
parent node was compressed. When this node was seen, it could not be
folded, due to the internal node it contained.

The invariant for compression in this function is: whenever
nr_leaves_on_branch < ASSOC_ARRAY_FAN_OUT, the node should contain all
leaf nodes. The compression step currently cannot guarantee this, given
the corner case shown above.

To fix this issue, retry compression whenever we have retained a node,
and yet nr_leaves_on_branch < ASSOC_ARRAY_FAN_OUT. This second
compression will then allow the node in slot 1 to be folded in,
satisfying the invariant. Below is the output of the reproducer once the
fix is applied:

    -- compress node 0x560e9c562380 --
    free=0, leaves=0
    [0] retain node 2/1 [nx 0]
    [1] fold node 1/1 [nx 0]
    [2] fold node 0/1 [nx 2]
    [3] fold node 0/2 [nx 2]
    [4] fold node 0/3 [nx 2]
    [5] fold node 0/4 [nx 2]
    [6] fold node 0/5 [nx 2]
    [7] fold node 0/6 [nx 2]
    [8] fold node 0/7 [nx 2]
    [9] fold node 0/8 [nx 2]
    [10] fold node 0/9 [nx 2]
    [11] fold node 0/10 [nx 2]
    [12] fold node 0/11 [nx 2]
    [13] fold node 0/12 [nx 2]
    [14] fold node 0/13 [nx 2]
    [15] fold node 0/14 [nx 2]
    internal nodes remain despite enough space, retrying
    -- compress node 0x560e9c562380 --
    free=14, leaves=1
    [0] fold node 2/15 [nx 0]
    after: 3

Changes
=======
DH:
 - Use false instead of 0.
 - Reorder the inserted lines in a couple of places to put retained before
   next_slot.

ver #2)
 - Fix typo in pr_devel, correct comparison to "<="

Fixes: 3cb98950

 ("Add a generic associative array implementation.")
Cc: <stable@vger.kernel.org>
Signed-off-by: Stephen Brennan <stephen.s.brennan@oracle.com>
Signed-off-by: David Howells <dhowells@redhat.com>
cc: Andrew Morton <akpm@linux-foundation.org>
cc: keyrings@vger.kernel.org
Link: https://lore.kernel.org/r/20220511225517.407935-1-stephen.s.brennan@oracle.com/ # v1
Link: https://lore.kernel.org/r/20220512215045.489140-1-stephen.s.brennan@oracle.com/

 # v2
Reviewed-by: Jarkko Sakkinen <jarkko@kernel.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>