On 24.06.25 10:07, David Hildenbrand wrote:
On 24.06.25 06:24, Li Wang wrote:
The current implementation of test_unmerge_uffd_wp() explicitly sets
`uffdio_api.features = UFFD_FEATURE_PAGEFAULT_FLAG_WP` before calling
UFFDIO_API. This can cause the ioctl() call to fail with EINVAL on kernels
that do not support UFFD-WP, leading the test to fail unnecessarily:
# ------------------------------
# running ./ksm_functional_tests
# ------------------------------
# TAP version 13
# 1..9
# # [RUN] test_unmerge
# ok 1 Pages were unmerged
# # [RUN] test_unmerge_zero_pages
# ok 2 KSM zero pages were unmerged
# # [RUN] test_unmerge_discarded
# ok 3 Pages were unmerged
# # [RUN] test_unmerge_uffd_wp
# not ok 4 UFFDIO_API failed <-----
# # [RUN] test_prot_none
# ok 5 Pages were unmerged
# # [RUN] test_prctl
# ok 6 Setting/clearing PR_SET_MEMORY_MERGE works
# # [RUN] test_prctl_fork
# # No pages got merged
# # [RUN] test_prctl_fork_exec
# ok 7 PR_SET_MEMORY_MERGE value is inherited
# # [RUN] test_prctl_unmerge
# ok 8 Pages were unmerged
# Bail out! 1 out of 8 tests failed
# # Planned tests != run tests (9 != 8)
# # Totals: pass:7 fail:1 xfail:0 xpass:0 skip:0 error:0
# [FAIL]
This patch improves compatibility and robustness of the UFFD-WP test
(test_unmerge_uffd_wp) by correctly implementing the UFFDIO_API
two-step handshake as recommended by the userfaultfd(2) man page.
Key changes:
1. Use features=0 in the initial UFFDIO_API call to query supported
feature bits, rather than immediately requesting WP support.
2. Skip the test gracefully if:
- UFFDIO_API fails with EINVAL (e.g. unsupported API version), or
- UFFD_FEATURE_PAGEFAULT_FLAG_WP is not advertised by the kernel.
3. Close the initial userfaultfd and create a new one before enabling
the required feature, since UFFDIO_API can only be called once per fd.
4. Improve diagnostics by distinguishing between expected and unexpected
failures, using strerror() to report errors.
This ensures the test behaves correctly across a wider range of kernel
versions and configurations, while preserving the intended behavior on
kernels that support UFFD-WP.
Suggestted-by: David Hildenbrand <[email protected]>
Signed-off-by: Li Wang <[email protected]>
Cc: Aruna Ramakrishna <[email protected]>
Cc: Bagas Sanjaya <[email protected]>
Cc: Catalin Marinas <[email protected]>
Cc: Dave Hansen <[email protected]>
Cc: Joey Gouly <[email protected]>
Cc: Johannes Weiner <[email protected]>
Cc: Keith Lucas <[email protected]>
Cc: Ryan Roberts <[email protected]>
Cc: Shuah Khan <[email protected]>
---
Notes:
v1 --> v2:
* Close the original userfaultfd and open a new one before enabling
features
* Reworked UFFDIO_API negotiation to follow the official two-step
handshake
.../selftests/mm/ksm_functional_tests.c | 28 +++++++++++++++++--
1 file changed, 26 insertions(+), 2 deletions(-)
diff --git a/tools/testing/selftests/mm/ksm_functional_tests.c
b/tools/testing/selftests/mm/ksm_functional_tests.c
index b61803e36d1c..19e5b741893a 100644
--- a/tools/testing/selftests/mm/ksm_functional_tests.c
+++ b/tools/testing/selftests/mm/ksm_functional_tests.c
@@ -393,9 +393,13 @@ static void test_unmerge_uffd_wp(void)
/* See if UFFD-WP is around. */
uffdio_api.api = UFFD_API;
- uffdio_api.features = UFFD_FEATURE_PAGEFAULT_FLAG_WP;
+ uffdio_api.features = 0;
if (ioctl(uffd, UFFDIO_API, &uffdio_api) < 0) {
- ksft_test_result_fail("UFFDIO_API failed\n");
+ if (errno == EINVAL)
+ ksft_test_result_skip("The API version requested is not
supported\n");
+ else
+ ksft_test_result_fail("UFFDIO_API failed: %s\n",
strerror(errno));
+
Not sure if that is really required. If UFFDIO_API failed after
__NR_userfaultfd worked something unexpected is happening.
goto close_uffd;
}
if (!(uffdio_api.features & UFFD_FEATURE_PAGEFAULT_FLAG_WP)) {
@@ -403,6 +407,26 @@ static void test_unmerge_uffd_wp(void)
goto close_uffd;
}
+ /*
+ * UFFDIO_API must only be called once to enable features.
+ * So we close the old userfaultfd and create a new one to
+ * actually enable UFFD_FEATURE_PAGEFAULT_FLAG_WP.
+ */
+ close(uffd);
Is that actually required?
The man page explicitly documents:
" EINVAL A previous UFFDIO_API call already enabled one or more
features for this userfaultfd. Calling UFF‐
DIO_API twice, the first time with no features set, is
explicitly allowed as per the two-step feature
detection handshake.
"
So if that doesn't work, something might be broken.
CCing Nadav and Peter:
Could it be that
commit 22e5fe2a2a279d9a6fcbdfb4dffe73821bef1c90
Author: Nadav Amit <[email protected]>
Date: Thu Sep 2 14:58:59 2021 -0700
userfaultfd: prevent concurrent API initialization
userfaultfd assumes that the enabled features are set once and never
changed after UFFDIO_API ioctl succeeded.
However, currently, UFFDIO_API can be called concurrently from two
different threads, succeed on both threads and leave userfaultfd's
features in non-deterministic state. Theoretically, other uffd operations
(ioctl's and page-faults) can be dispatched while adversely affected by
such changes of features.
Moreover, the writes to ctx->state and ctx->features are not ordered,
which can - theoretically, again - let userfaultfd_ioctl() think that
userfaultfd API completed, while the features are still not initialized.
To avoid races, it is arguably best to get rid of ctx->state. Since there
are only 2 states, record the API initialization in ctx->features as the
uppermost bit and remove ctx->state.
Accidentally broke the documented two-step handshake in the man page where we
can avoid closing + reopening the fd?
Without testing, the following might fix it if I am right:
diff --git a/fs/userfaultfd.c b/fs/userfaultfd.c
index 22f4bf956ba1c..f03e7c980e1c5 100644
--- a/fs/userfaultfd.c
+++ b/fs/userfaultfd.c
@@ -1944,9 +1944,9 @@ static int userfaultfd_move(struct userfaultfd_ctx *ctx,
static int userfaultfd_api(struct userfaultfd_ctx *ctx,
unsigned long arg)
{
+ unsigned int new_features, old_features = 0;
struct uffdio_api uffdio_api;
void __user *buf = (void __user *)arg;
- unsigned int ctx_features;
int ret;
__u64 features;
@@ -1990,9 +1990,12 @@ static int userfaultfd_api(struct userfaultfd_ctx *ctx,
goto out;
/* only enable the requested features for this uffd context */
- ctx_features = uffd_ctx_features(features);
+ new_features = uffd_ctx_features(features);
+ /* allow two-step handshake */
+ if (userfaultfd_is_initialized(ctx))
+ old_features = UFFD_FEATURE_INITIALIZED;
ret = -EINVAL;
- if (cmpxchg(&ctx->features, 0, ctx_features) != 0)
+ if (cmpxchg(&ctx->features, old_features, new_features) != old_features)
goto err_out;
ret = 0;
--
Cheers,
David / dhildenb
