Xiang Mei (Microsoft) [off-list ref] wrote:

xfrm_hash_rebuild()'s first loop preallocates the bins/chains the reinsert
loop needs, so the reinsert (after hlist_del_rcu()) cannot allocate or
fail. But its guard is inverted: it skips policies with prefixlen <
threshold and preallocates for the rest.

prefixlen < threshold is exactly when policy_hash_bysel() returns NULL and
the reinsert takes the allocating xfrm_policy_inexact_insert() path. So the
loop preallocates for the exact policies (which never allocate) and skips
the inexact ones, whose bin/node is then allocated GFP_ATOMIC during
reinsert. On failure the error path only WARN_ONCE()s and continues,
leaving a poisoned bydst node; the next rebuild's hlist_del_rcu()
dereferences LIST_POISON2 and takes a GPF. Reachable under memory pressure,
deterministic via failslab.

Invert the guard so preallocation covers exactly the reinserted policies;
the reinsert then allocates nothing and cannot fail.

Crash:
  Oops: general protection fault, probably for non-canonical address
  0xfbd59c0000000024: 0000 [#1] SMP KASAN NOPTI
  KASAN: maybe wild-memory-access in range [0xdead...]
  ...
  Workqueue: events xfrm_hash_rebuild
  RIP: 0010:xfrm_hash_rebuild+0x5b3/0x1190
  RAX: dead000000000122   (LIST_POISON2 + offset)
  ...
  Call Trace:
   hlist_del_rcu (include/linux/rculist.h:599)
   xfrm_hash_rebuild (net/xfrm/xfrm_policy.c:1365)
   process_one_work (kernel/workqueue.c:3322)
   worker_thread (kernel/workqueue.c:3486)
   kthread (kernel/kthread.c:436)
   ret_from_fork (arch/x86/kernel/process.c:158)
   ret_from_fork_asm (arch/x86/entry/entry_64.S:245)
   ...
  Kernel panic - not syncing: Fatal exception in interrupt

Fixes: 24969facd704 ("xfrm: policy: store inexact policies in an rhashtable")
Reported-by: AutonomousCodeSecurity@microsoft.com
Signed-off-by: Xiang Mei (Microsoft) <redacted>

Reviewed-by: Florian Westphal <fw@strlen.de>