Thread (1 message) flat view 1 message, 1 author, 2025-06-12

Re: [PATCH] man/man2/memfd_secret.2: Correct the flags

From: Alejandro Colomar <alx@kernel.org>
Date: 2025-06-12 09:45:42
Also in: linux-fsdevel, linux-man, linux-mm, lkml

[CC += people related to memfd_{create,secret}(2) in the kernel]

Hi Zhengyi,

On Thu, Jun 12, 2025 at 02:17:05PM +0800, Zhengyi Fu wrote:
memfd_secret returns EINVAL when called with FD_CLOEXEC.  The
correct flag should be O_CLOEXEC.
Thanks for the report!  It seems like a bug in the kernel.  The
documentation was written (relatively) consistent with memfd_create(2),
but the implementation was made different.  I say the documentation was
relatively consistent, because memfd_create(2) uses MFD_CLOEXEC, and
memfd_secret(2) documents FD_CLOEXEC, which could be confused, and since
they have the same value, it could be considered just a typo.  However,
O_CLOEXEC is an entirely different flag, which doesn't seem to make
sense here.

	$ grepc -tfld memfd_create . | grep -A4 -e '^[{}.]' -e CLOEXEC;
	./mm/memfd.c:SYSCALL_DEFINE2(memfd_create,
			const char __user *, uname,
			unsigned int, flags)
	{
		struct file *file;
		int fd, error;
		char *name;

	--
		fd = get_unused_fd_flags((flags & MFD_CLOEXEC) ? O_CLOEXEC : 0);
		if (fd < 0) {
			error = fd;
			goto err_name;
		}
	--
	}

	$ grepc -tfld memfd_secret . | grep -A3 -e '^[{}.]' -e CLOEXEC;
	./mm/secretmem.c:SYSCALL_DEFINE1(memfd_secret, unsigned int, flags)
	{
		struct file *file;
		int fd, err;

	--
		BUILD_BUG_ON(SECRETMEM_FLAGS_MASK & O_CLOEXEC);

		if (!secretmem_enable || !can_set_direct_map())
			return -ENOSYS;
	--
		if (flags & ~(SECRETMEM_FLAGS_MASK | O_CLOEXEC))
			return -EINVAL;
		if (atomic_read(&secretmem_users) < 0)
			return -ENFILE;
	--
		fd = get_unused_fd_flags(flags & O_CLOEXEC);
		if (fd < 0)
			return fd;

	--
	}

Let's see who added memfd_create(2):

	alx@devuan:~/src/linux/linux/master$ git blame -- ./mm/memfd.c | grep _CLOEXEC
	105ff5339f498 (Jeff Xu                 2022-12-15 00:12:03 +0000 306) #define MFD_ALL_FLAGS (MFD_CLOEXEC | MFD_ALLOW_SEALING | MFD_HUGETLB | MFD_NOEXEC_SEAL | MFD_EXEC)
	f5dbcd90dacd3 (Isaac J. Manjarres      2025-01-10 08:58:59 -0800 475) 	fd = get_unused_fd_flags((flags & MFD_CLOEXEC) ? O_CLOEXEC : 0);
	alx@devuan:~/src/linux/linux/master$ git show f5dbcd90dacd3 | grep -e _CLOEXEC -e ^diff | grep -B1 -v ^d
	diff --git a/mm/memfd.c b/mm/memfd.c
	-	fd = get_unused_fd_flags((flags & MFD_CLOEXEC) ? O_CLOEXEC : 0);
	+	fd = get_unused_fd_flags((flags & MFD_CLOEXEC) ? O_CLOEXEC : 0);
	alx@devuan:~/src/linux/linux/master$ git blame f5dbcd90dacd3^ -- mm/memfd.c | grep _CLOEXEC
	105ff5339f498 (Jeff Xu                 2022-12-15 00:12:03 +0000 305) #define MFD_ALL_FLAGS (MFD_CLOEXEC | MFD_ALLOW_SEALING | MFD_HUGETLB | MFD_NOEXEC_SEAL | MFD_EXEC)
	5d752600a8c37 (Mike Kravetz            2018-06-07 17:06:01 -0700 423) 	fd = get_unused_fd_flags((flags & MFD_CLOEXEC) ? O_CLOEXEC : 0);
	alx@devuan:~/src/linux/linux/master$ git show 5d752600a8c37 | grep -e _CLOEXEC -e ^diff | grep -B1 -v ^d
	diff --git a/mm/memfd.c b/mm/memfd.c
	+#define MFD_ALL_FLAGS (MFD_CLOEXEC | MFD_ALLOW_SEALING | MFD_HUGETLB)
	+	fd = get_unused_fd_flags((flags & MFD_CLOEXEC) ? O_CLOEXEC : 0);
	diff --git a/mm/shmem.c b/mm/shmem.c
	-#define MFD_ALL_FLAGS (MFD_CLOEXEC | MFD_ALLOW_SEALING | MFD_HUGETLB)
	-	fd = get_unused_fd_flags((flags & MFD_CLOEXEC) ? O_CLOEXEC : 0);
	alx@devuan:~/src/linux/linux/master$ git blame 5d752600a8c37^ -- mm/shmem.c | grep _CLOEXEC
	749df87bd7bee (Mike Kravetz            2017-09-06 16:24:16 -0700 3684) #define MFD_ALL_FLAGS (MFD_CLOEXEC | MFD_ALLOW_SEALING | MFD_HUGETLB)
	9183df25fe7b1 (David Rheinsberg        2014-08-08 14:25:29 -0700 3729) 	fd = get_unused_fd_flags((flags & MFD_CLOEXEC) ? O_CLOEXEC : 0);
	alx@devuan:~/src/linux/linux/master$ git show 9183df25fe7b1 | grep -e _CLOEXEC -e ^diff | grep -B1 -v ^d
	diff --git a/include/uapi/linux/memfd.h b/include/uapi/linux/memfd.h
	+#define MFD_CLOEXEC		0x0001U
	--
	diff --git a/mm/shmem.c b/mm/shmem.c
	+#define MFD_ALL_FLAGS (MFD_CLOEXEC | MFD_ALLOW_SEALING)
	+	fd = get_unused_fd_flags((flags & MFD_CLOEXEC) ? O_CLOEXEC : 0);

	alx@devuan:~/src/linux/linux/master$ git show 9183df25fe7b1 | head -n5
	commit 9183df25fe7b194563db3fec6dc3202a5855839c
	Author: David Rheinsberg [off-list ref]
	Date:   Fri Aug 8 14:25:29 2014 -0700

	    shm: add memfd_create() syscall

	alx@devuan:~/src/linux/linux/master$ git log -1 9183df25fe7b1 | grep @
	Author: David Rheinsberg [off-list ref]
	    Signed-off-by: David Herrmann [off-list ref]
	    Acked-by: Hugh Dickins [off-list ref]
	    Cc: Michael Kerrisk [off-list ref]
	    Cc: Ryan Lortie [off-list ref]
	    Cc: Lennart Poettering [off-list ref]
	    Cc: Daniel Mack [off-list ref]
	    Cc: Andy Lutomirski [off-list ref]
	    Signed-off-by: Andrew Morton [off-list ref]
	    Signed-off-by: Linus Torvalds [off-list ref]

And memfd_secret(2):

	alx@devuan:~/src/linux/linux/master$ git blame -- ./mm/secretmem.c | grep _CLOEXEC
	1507f51255c9f (Mike Rapoport           2021-07-07 18:08:03 -0700 238) 	BUILD_BUG_ON(SECRETMEM_FLAGS_MASK & O_CLOEXEC);
	1507f51255c9f (Mike Rapoport           2021-07-07 18:08:03 -0700 243) 	if (flags & ~(SECRETMEM_FLAGS_MASK | O_CLOEXEC))
	1507f51255c9f (Mike Rapoport           2021-07-07 18:08:03 -0700 248) 	fd = get_unused_fd_flags(flags & O_CLOEXEC);
	alx@devuan:~/src/linux/linux/master$ git show 1507f51255c9f | grep -e _CLOEXEC -e ^diff | grep -B1 -v ^d
	diff --git a/mm/secretmem.c b/mm/secretmem.c
	+	BUILD_BUG_ON(SECRETMEM_FLAGS_MASK & O_CLOEXEC);
	+	if (flags & ~(SECRETMEM_FLAGS_MASK | O_CLOEXEC))
	+	fd = get_unused_fd_flags(flags & O_CLOEXEC);
	alx@devuan:~/src/linux/linux/master$ git show 1507f51255c9f | head -n5
	commit 1507f51255c9ff07d75909a84e7c0d7f3c4b2f49
	Author: Mike Rapoport [off-list ref]
	Date:   Wed Jul 7 18:08:03 2021 -0700

	    mm: introduce memfd_secret system call to create "secret" memory areas
	alx@devuan:~/src/linux/linux/master$ git log -1 1507f51255c9f | grep @
	Author: Mike Rapoport [off-list ref]
	    [1] https://lore.kernel.org/linux-mm/213b4567-46ce-f116-9cdf-bbd0c884eb3c@linux.intel.com/ (local)
	    [akpm@linux-foundation.org: suppress Kconfig whine]
	    Link: https://lkml.kernel.org/r/20210518072034.31572-5-rppt@kernel.org
	    Signed-off-by: Mike Rapoport [off-list ref]
	    Acked-by: Hagen Paul Pfeifer [off-list ref]
	    Acked-by: James Bottomley [off-list ref]
	    Cc: Alexander Viro [off-list ref]
	    Cc: Andy Lutomirski [off-list ref]
	    Cc: Arnd Bergmann [off-list ref]
	    Cc: Borislav Petkov [off-list ref]
	    Cc: Catalin Marinas [off-list ref]
	    Cc: Christopher Lameter [off-list ref]
	    Cc: Dan Williams [off-list ref]
	    Cc: Dave Hansen [off-list ref]
	    Cc: Elena Reshetova [off-list ref]
	    Cc: "H. Peter Anvin" [off-list ref]
	    Cc: Ingo Molnar [off-list ref]
	    Cc: James Bottomley [off-list ref]
	    Cc: "Kirill A. Shutemov" [off-list ref]
	    Cc: Matthew Wilcox [off-list ref]
	    Cc: Mark Rutland [off-list ref]
	    Cc: Michael Kerrisk [off-list ref]
	    Cc: Palmer Dabbelt [off-list ref]
	    Cc: Palmer Dabbelt [off-list ref]
	    Cc: Paul Walmsley [off-list ref]
	    Cc: Peter Zijlstra [off-list ref]
	    Cc: Rick Edgecombe [off-list ref]
	    Cc: Roman Gushchin [off-list ref]
	    Cc: Shakeel Butt [off-list ref]
	    Cc: Shuah Khan [off-list ref]
	    Cc: Thomas Gleixner [off-list ref]
	    Cc: Tycho Andersen [off-list ref]
	    Cc: Will Deacon [off-list ref]
	    Cc: David Hildenbrand [off-list ref]
	    Cc: kernel test robot [off-list ref]
	    Signed-off-by: Andrew Morton [off-list ref]
	    Signed-off-by: Linus Torvalds [off-list ref]

I've added to CC everyone who had something different than Cc, and
everyone who had Cc in both.

Now about the situation: it seems there is only one user of CLOEXEC
with memfd_secret(2) in Debian: systemtap.
<https://codesearch.debian.net/search?q=memfd_secret.*CLOEXEC&literal=0>

Do we want to fix the bug, or do we want to document it?  This is for
kernel people to respond.

Also, was O_CLOEXEC used on purpose, or was it by accident?  I expect
that either MFD_CLOEXEC should have been used, by imitating
memfd_create(2), or a new MFDS_CLOEXEC could have been invented, but
O_CLOEXEC doesn't make much sense, IMO.


Have a lovely day!
Alex
quoted hunk
Signed-off-by: Zhengyi Fu <redacted>
---
 man/man2/memfd_secret.2 | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
diff --git a/man/man2/memfd_secret.2 b/man/man2/memfd_secret.2
index 5ba7813c1..c6abd2f5f 100644
--- a/man/man2/memfd_secret.2
+++ b/man/man2/memfd_secret.2
@@ -51,7 +51,7 @@ The following values may be bitwise ORed in
 to control the behavior of
 .BR memfd_secret ():
 .TP
-.B FD_CLOEXEC
+.B O_CLOEXEC
 Set the close-on-exec flag on the new file descriptor,
 which causes the region to be removed from the process on
 .BR execve (2).
-- 
<https://www.alejandro-colomar.es/>

Attachments

Keyboard shortcuts
hback out one level
jnext message in thread
kprevious message in thread
ldrill in
Escclose help / fold thread tree
?toggle this help