of: unittest: Require BUG_ON_DATA_CORRUPTION to be disabled
PATCH 1/1 of: unittest: Require BUG_ON_DATA_CORRUPTION to be disabled
With CONFIG_OF_UNITTEST=y and CONFIG_BUG_ON_DATA_CORRUPTION=y, the machine no longer boots after commit 5b5224220747 ("refcount: Use CONFIG_BUG_ON_DATA_CORRUPTION for UAF-related errors").
of_unittest_lifecycle() deliberately calls of_node_put() after the node's reference count has reached zero. The test expects a warning and continues to check that a subsequent put does not produce another warning. With BUG_ON_DATA_CORRUPTION enabled, refcount underflow now triggers BUG() instead. This panics the kernel during boot:
kernel BUG at lib/refcount.c:37! Oops: invalid opcode: 0000 [#1] SMP NOPTI CPU: 1 UID: 0 PID: 1 Comm: swapper/0 Tainted: G N 7.3.0+ #14 PREEMPT(full) RIP: 0010:refcount_warn_saturate+0x89/0x90 Call Trace: of_unittest_lifecycle+0x158/0x310 of_unittest+0x105/0x130 do_one_initcall+0x127/0x350 do_initcall_level+0x7c/0xe0 do_initcalls+0x59/0xa0 kernel_init_freeable+0x118/0x1a0 kernel_init+0x1f/0x140 Kernel panic - not syncing: Attempted to kill init! exitcode=0x0000000b
Make OF_UNITTEST depend on !BUG_ON_DATA_CORRUPTION and explain the restriction in its help text.
---
Alternatively, we could also just remove the refcount underflow tests. It seems they are not really testing anything OF specific, but rather test the refcount implementation.
--- drivers/of/Kconfig | 5 ++++- 1 file changed, 4 insertions(+), 1 deletion(-)
drivers/of/Kconfig +4 -1
@@ -21,6 +21,7 @@ if OF config OF_UNITTEST bool "Device Tree runtime unit tests" depends on OF_EARLY_FLATTREE + depends on !BUG_ON_DATA_CORRUPTION select IRQ_DOMAIN select OF_RESOLVE help
@@ -32,7 +33,9 @@ config OF_UNITTEST will taint the kernel with TAINT_TEST. The tests will cause ERROR and WARNING messages to print on the console. The tests will cause stack traces to print on the console. It is possible that the tests will - leave the devicetree in a corrupted state. + leave the devicetree in a corrupted state. BUG_ON_DATA_CORRUPTION + must be disabled because the tests deliberately trigger refcount + underflows and expect execution to continue. The unittest output will be verbose. Copy the output to a file via capturing the console output or via the dmesg command. Process
base-commit: a287f97d4256b5f5527f5b5a54f8ada55e1482c7
Signature · 1 line
2.56.0
Rob Herring
maintainerQuoted 28 lines · Danilo Krummrich
With
CONFIG_OF_UNITTEST=yandCONFIG_BUG_ON_DATA_CORRUPTION=y, the machine no longer boots after commit5b5224220747("refcount: UseCONFIG_BUG_ON_DATA_CORRUPTIONfor UAF-related errors").
of_unittest_lifecycle()deliberately callsof_node_put()after the node's reference count has reached zero. The test expects a warning and continues to check that a subsequent put does not produce another warning. WithBUG_ON_DATA_CORRUPTIONenabled, refcount underflow now triggersBUG()instead. This panics the kernel during boot:kernel BUG at lib/refcount.c:37! Oops: invalid opcode: 0000 [#1] SMP NOPTI CPU: 1 UID: 0 PID: 1 Comm: swapper/0 Tainted: G N 7.3.0+ #14 PREEMPT(full) RIP: 0010:refcount_warn_saturate+0x89/0x90 Call Trace: of_unittest_lifecycle+0x158/0x310 of_unittest+0x105/0x130 do_one_initcall+0x127/0x350 do_initcall_level+0x7c/0xe0 do_initcalls+0x59/0xa0 kernel_init_freeable+0x118/0x1a0 kernel_init+0x1f/0x140 Kernel panic - not syncing: Attempted to kill init! exitcode=0x0000000bMake
OF_UNITTESTdepend on !BUG_ON_DATA_CORRUPTION and explain the restriction in its help text.
Quoted 6 lines
Fixes 5b5224220747 ("refcount: Use CONFIG_BUG_ON_DATA_CORRUPTION for UAF-related errors")Signed-off-by Danilo Krummrich <[email protected]>---
Alternatively, we could also just remove the refcount underflow tests. It seems they are not really testing anything OF specific, but rather test the refcount implementation.
I'm fine with that. I think we should keep just the first part with of_node_release() error check.
Rob