lkml.org 
[lkml]   [2019]   [Jun]   [12]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
Patch in this message
/
From
Subject[RFC PATCH] binfmt_elf: Protect mm_struct access with mmap_sem
Date
find_extend_vma assumes the caller holds mmap_sem as a reader (explained
in expand_downwards()). The path when we are extending the stack VMA to
accomodate argv[] pointers happens without the lock.

I was not able to cause an mm_struct corruption but
BUG_ON(!rwsem_is_locked(&mm->mmap_sem)) in find_extend_vma could be
triggered as

# <bigfile xargs echo
xargs: echo: terminated by signal 11

(bigfile needs to have more than RLIMIT_STACK / sizeof(char *) rows)

Other accesses to mm_struct in exec path are protected by mmap_sem, so
conservatively, protect also this one. Besides that, explain why we omit
mm_struct.arg_lock in the exec(2) path.

Cc: Cyrill Gorcunov <gorcunov@gmail.com>
Signed-off-by: Michal Koutný <mkoutny@suse.com>
---

When I was attempting to reduce usage of mmap_sem I came across this
unprotected access and increased number of its holders :-/

I'm not sure whether there is a real concurrent writer at this early
stages (I considered khugepaged especially as setup_arg_pages invokes
khugepaged_enter_vma_merge but we're lucky because khugepaged skips it
because of VM_STACK_INCOMPLETE_SETUP).

A nicer approach would perhaps be to do all this exec setup when the
mm_struct is still not exposed via current->mm (and hence no need to
synchronize via mmap_sem). But I didn't look enough into binfmt specific
whether it is even doable and worth it.

So I'm sending this for a discussion.

fs/binfmt_elf.c | 10 +++++++++-
fs/exec.c | 3 ++-
2 files changed, 11 insertions(+), 2 deletions(-)

diff --git a/fs/binfmt_elf.c b/fs/binfmt_elf.c
index 8264b468f283..48e169760a9c 100644
--- a/fs/binfmt_elf.c
+++ b/fs/binfmt_elf.c
@@ -299,7 +299,11 @@ create_elf_tables(struct linux_binprm *bprm, struct elfhdr *exec,
* Grow the stack manually; some architectures have a limit on how
* far ahead a user-space access may be in order to grow the stack.
*/
+ if (down_read_killable(&current->mm->mmap_sem))
+ return -EINTR;
vma = find_extend_vma(current->mm, bprm->p);
+ up_read(&current->mm->mmap_sem);
+
if (!vma)
return -EFAULT;

@@ -1123,11 +1127,15 @@ static int load_elf_binary(struct linux_binprm *bprm)
goto out;
#endif /* ARCH_HAS_SETUP_ADDITIONAL_PAGES */

+ /*
+ * Don't take mm->arg_lock. The concurrent change might happen only
+ * from prctl_set_mm but after de_thread we are certainly alone here.
+ */
retval = create_elf_tables(bprm, &loc->elf_ex,
load_addr, interp_load_addr);
if (retval < 0)
goto out;
- /* N.B. passed_fileno might not be initialized? */
+
current->mm->end_code = end_code;
current->mm->start_code = start_code;
current->mm->start_data = start_data;
diff --git a/fs/exec.c b/fs/exec.c
index 89a500bb897a..d5b55c92019a 100644
--- a/fs/exec.c
+++ b/fs/exec.c
@@ -212,7 +212,8 @@ static struct page *get_arg_page(struct linux_binprm *bprm, unsigned long pos,

/*
* We are doing an exec(). 'current' is the process
- * doing the exec and bprm->mm is the new process's mm.
+ * doing the exec and bprm->mm is the new process's mm that is not
+ * shared yet, so no synchronization on mmap_sem.
*/
ret = get_user_pages_remote(current, bprm->mm, pos, 1, gup_flags,
&page, NULL, NULL);
--
2.21.0
\
 
 \ /
  Last update: 2019-06-12 16:29    [W:0.049 / U:13.612 seconds]
©2003-2018 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site