keyctl_pkey_query(), keyctl_pkey_e_d_s() and keyctl_pkey_verify() check that key->type has the ->asym_*() method they need and call it later, without holding key->sem. A reference keeps the key alive, but does not pin the module that provides its type. When that module is unloaded, unregister_key_type() waits for the key GC to process every remaining key of the type. For each such key, the GC takes key->sem for writing, switches key->type to key_type_dead, whose ->asym_*() methods are NULL, and destroys the payload. If this happens between the check and the call, the keyctl calls a NULL method: KEYCTL_PKEY_SIGN task Module unload and key GC ------------------------ ------------------------ keyctl_pkey_e_d_s() if (!key->type->asym_eds_op) unregister_key_type() GC: down_write(&key->sem) GC: retype to key_type_dead GC: up_write(&key->sem) free_module() in = memdup_user(...) key->type->asym_eds_op(...) -> NULL call If it happens while the method is running, the method is left running in freed module text. In-tree, turris-signing-key provides ->asym_query() and ->asym_eds_op() and can be built as a module. Take key->sem for reading around each ->asym_*() call, and check the method again under it, right before the call. While key->sem is held, the GC cannot switch the key to key_type_dead, so unregister_key_type(), and with it the module unload, cannot complete: the type, its module and the payload stay in place until the method returns. If the type is already gone when key->sem is taken, the check finds key_type_dead and the keyctl fails with -EOPNOTSUPP, as it does for any key type without the method. Otherwise, behaviour is unchanged: - The existing checks in keyctl_pkey_params_get(), keyctl_pkey_e_d_s() and keyctl_pkey_verify() are kept, so key types without the method still fail at the same point with the same error. No call relies on them any more. - memdup_user() and copy_to_user() stay outside key->sem, as in keyctl_read_key(), so no page fault is taken with key->sem held. - key->sem is only taken for reading, so pkey operations and KEYCTL_READ on the same key still run concurrently. Writers such as KEYCTL_UPDATE, KEYCTL_REVOKE and KEYCTL_SETPERM now wait for a pkey operation in progress, as they already wait for ->read(). The in-tree ->asym_*() methods do not take key->sem themselves. Nothing else in the kernel calls these three methods, so the four call sites below are the full coverage: - keyctl_pkey_query(): ->asym_query() for KEYCTL_PKEY_QUERY - keyctl_pkey_params_get_2(): ->asym_query() to check the buffer sizes for KEYCTL_PKEY_ENCRYPT, _DECRYPT, _SIGN and _VERIFY - keyctl_pkey_e_d_s(): ->asym_eds_op() for KEYCTL_PKEY_ENCRYPT, _DECRYPT and _SIGN - keyctl_pkey_verify(): ->asym_verify_signature() for KEYCTL_PKEY_VERIFY This problem was recently fixed for keyctl_read_key() by 925c84e3cc21 ("keys: Protect the read method check against type removal") and for key_update() by aef3a8a4f988 ("KEYS: Protect the update method check against type removal"). LLM assistants, OpenAI Codex (gpt-6.1-sol) and Anthropic Claude (claude-opus-5-5), found this while looking for other cases of the KEYCTL_READ problem, and drafted this patch and this description. Fixes: 00d60fd3b932 ("KEYS: Provide keyctls to drive the new key type ops for asymmetric keys [ver #2]") Cc: stable@vger.kernel.org Signed-off-by: Yuansheng Yang <0xbeac@gmail.com> --- Changes in v5: - Keep the asym_query check in keyctl_pkey_params_get() (Jarkko), and for the same reason keep the asym_eds_op and asym_verify_signature checks that keyctl_pkey_e_d_s() and keyctl_pkey_verify() do before copying the input. All four call sites now get the same change and nothing else moves. This also restores the error for KEYCTL_PKEY_VERIFY with a bad input pointer on a key type without ->asym_verify_signature, such as turris-signing-key: v4 returned -EFAULT there instead of -EOPNOTSUPP. - Rewrite the commit message around that change: what it does, why it is safe, and why the four call sites are all of them. - Say in the changelog that LLM assistants found this bug and drafted the patch and its description. They also wrote the tests below. Tested on x86_64 under QEMU/KVM, with KASAN and lockdep, on 7.4-rc1 with and without this patch. A test-only key type module (not for merging) is unloaded while KEYCTL_PKEY_SIGN or KEYCTL_PKEY_VERIFY waits in memdup_user() on a userfaultfd page, after the early check: without the patch: NULL call from keyctl_pkey_e_d_s() and keyctl_pkey_verify() with the patch: both return -EOPNOTSUPP When the module's ->asym_eds_op() is sleeping as the unload starts, the unpatched kernel faults on the freed module text; with the patch, delete_module() returns only after the method has returned. KEYCTL_PKEY_QUERY, _ENCRYPT, _DECRYPT, _SIGN and _VERIFY on an RSA-2048 asymmetric key give the same results with and without the patch. In the two forced races lockdep also reports "bad unlock balance" for key->sem. Its lock class is the type's lock_class, which lockdep drops when the module is freed, so any key->sem user that took its key reference before the unload can hit this, e.g. keyctl_read_key(). Not tested on Turris hardware. v4: https://lore.kernel.org/all/20261105193632.34363-1-0xbeac@gmail.com/ v3: https://lore.kernel.org/all/9d47e202bd11857b3555e6b95694b3c3a4f2460b.1793852399.git.0xbeac@gmail.com/ v2: https://lore.kernel.org/all/20261104021414.50696-1-0xbeac@gmail.com/ v1: https://lore.kernel.org/all/20261103032940.51826-1-0xbeac@gmail.com/ security/keys/keyctl_pkey.c | 49 ++++++++++++++++++++++++++++++++----- 1 file changed, 43 insertions(+), 6 deletions(-) diff --git a/security/keys/keyctl_pkey.c b/security/keys/keyctl_pkey.c index 15b6bf6399..0ea6f2790e 100644 --- a/security/keys/keyctl_pkey.c +++ b/security/keys/keyctl_pkey.c @@ -129,7 +129,16 @@ static int keyctl_pkey_params_get_2(const struct keyctl_pkey_params __user *_par if (ret < 0) return ret; + down_read(¶ms->key->sem); + if (!params->key->type->asym_query) { + ret = -EOPNOTSUPP; + goto out_up_key_sem; + } + ret = params->key->type->asym_query(params, &info); + +out_up_key_sem: + up_read(¶ms->key->sem); if (ret < 0) return ret; @@ -185,7 +194,16 @@ long keyctl_pkey_query(key_serial_t id, if (ret < 0) goto error; + down_read(¶ms.key->sem); + if (!params.key->type->asym_query) { + ret = -EOPNOTSUPP; + goto out_up_key_sem; + } + ret = params.key->type->asym_query(¶ms, &res); + +out_up_key_sem: + up_read(¶ms.key->sem); if (ret < 0) goto error; @@ -260,7 +278,16 @@ long keyctl_pkey_e_d_s(int op, if (!out) goto error_in; + down_read(¶ms.key->sem); + if (!params.key->type->asym_eds_op) { + ret = -EOPNOTSUPP; + goto out_up_key_sem; + } + ret = params.key->type->asym_eds_op(¶ms, in, out); + +out_up_key_sem: + up_read(¶ms.key->sem); if (ret < 0) goto error_out; @@ -305,31 +332,41 @@ long keyctl_pkey_verify(const struct keyctl_pkey_params __user *_params, ret = keyctl_pkey_params_get_2(_params, _info, KEYCTL_PKEY_VERIFY, ¶ms); if (ret < 0) - goto error_params; + goto out_free_params; ret = -EOPNOTSUPP; if (!params.key->type->asym_verify_signature) - goto error_params; + goto out_free_params; in = memdup_user(_in, params.in_len); if (IS_ERR(in)) { ret = PTR_ERR(in); - goto error_params; + goto out_free_params; } in2 = memdup_user(_in2, params.in2_len); if (IS_ERR(in2)) { ret = PTR_ERR(in2); - goto error_in; + goto out_free_in; } params.op = kernel_pkey_verify; + down_read(¶ms.key->sem); + if (!params.key->type->asym_verify_signature) { + ret = -EOPNOTSUPP; + goto out_up_key_sem; + } + ret = params.key->type->asym_verify_signature(¶ms, in, in2); +out_up_key_sem: + up_read(¶ms.key->sem); kfree(in2); -error_in: + +out_free_in: kfree(in); -error_params: + +out_free_params: keyctl_pkey_params_free(¶ms); return ret; } -- 2.43.0