fs/exec.c: fix minor memory leak

PATCH Applied Oleg Nesterov 2 participants 3 messages May 16, 2016 – May 17, 2016

Subsystems (3)

Memory management VFS exec and binfmt

lore

Share & tools

lore

Oleg Nesterov

author

2016-05-16 20:43 UTC #0 lore raw

Andrew, Vlastimil,

I found this patch by accident when I was looking at http://marc.info/?l=linux-mm
and I can't resist ;)

Quoted 17 lines · Andrew Morton

Quoted 2 lines
Could someone please double-check this?
Looks OK to me.
Quoted 11 lines
From: Andrew Morton <[email protected]>
Subject: fs/exec.c: fix minor memory leak
When the to-be-removed argument's trailing '\0' is the final byte in the page, remove_arg_zero()'s logic will avoid freeing the page, will break from the loop and will then advance bprm->p to point at the first byte in the next page. Net result: the final page for the zeroeth argument is unfreed.
It isn't a very important leak - that page will be freed later by the bprm-wide sweep in free_arg_pages().

And so I think we should just remove this free_arg_page(), it (and the patch) only adds the unnecessary confusion.

Note that today free_arg_page() is nop if CONFIG_MMU. At the same time, the only reason for this free_arg_page() was that (until the commit b6a2fea39) CONFIG_MMU did install_arg_page() for every page != NULL in bprm->page[].

So we simply do not need it today. And note that the caller is going to do copy_strings_kernel(), so if we do free_arg_page() with CONFIG_MMU=n we will likely have to re-allocate this page right after free.

And note that this code is actually wrong! remove_arg_zero() assumes that argv[0] is null-terminated but this is not necessarily true. copy_strings() does:

	len = strnlen_user(...);
	...
	copy_from_user(..., len);

another thread or debugger can change the memory in between. Fortunately nothing really bad can happen (afaics) even if CONFIG_MMU=n, bprm->filename must be always zero-terminated and it was copied by the 1st copy_strings_kernel(). Still perhaps it makes sense to check "bprm->p < bprm->exec" in the main loop.

Oleg.

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in the body to [email protected]. For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"[email protected]"> [email protected] </a>

Andrew Morton

maintainer

2016-05-16 20:55 UTC #1 lore raw

Quoted 20 lines · On Mon, 16 May 2016 22:43:39 +0200 Oleg Nesterov <[email protected]> wrote:

Andrew, Vlastimil,
I found this patch by accident when I was looking at http://marc.info/?l=linux-mm
and I can't resist ;)
Quoted 15 lines · Andrew Morton
Quoted 2 lines
Could someone please double-check this?
Looks OK to me.
Quoted 9 lines
From: Andrew Morton <[email protected]>
Subject: fs/exec.c: fix minor memory leak
When the to-be-removed argument's trailing '\0' is the final byte in the page, remove_arg_zero()'s logic will avoid freeing the page, will break from the loop and will then advance bprm->p to point at the first byte in the next page. Net result: the final page for the zeroeth argument is unfreed.

Quoted 6 lines

Quoted 2 lines
Quoted 2 lines
It isn't a very important leak - that page will be freed later by the bprm-wide sweep in free_arg_pages().
And so I think we should just remove this free_arg_page(), it (and the patch) only adds the unnecessary confusion.

Send patch :)

Oleg Nesterov

author

2016-05-17 15:53 UTC #2 lore raw

PATCH 1/1 exec: remove the no longer needed remove_arg_zero()->free_arg_page()

remove_arg_zero() does free_arg_page() for no reason. This was needed before and only if CONFIG_MMU=y: see the commit 4fc75ff4 ("exec: fix remove_arg_zero"), install_arg_page() was called for every page != NULL in bprm->page[] array. Today install_arg_page() has already gone and free_arg_page() is nop after another commit b6a2fea39 ("mm: variable length argument support").

CONFIG_MMU=n does free_arg_pages() in free_bprm() and thus it doesn't need remove_arg_zero()->free_arg_page() too; apart from get_arg_page() it never checks if the page in bprm->page[] was allocated or not, so the "extra" non-freed page is fine. OTOH, this free_arg_page() can add the minor pessimization, the caller is going to do copy_strings_kernel() right after remove_arg_zero() which will likely need to re-allocate the same page again.

And as Hujunjie pointed out, the "offset == PAGE_SIZE" check is wrong because we are going to increment bprm->p once again before return, so CONFIG_MMU=n "leaks" the page anyway if '\0' is the final byte in this page.

NOTE: remove_arg_zero() assumes that argv[0] is null-terminated but this is not necessarily true. copy_strings() does "len = strnlen_user(...)", then copy_from_user(len) but another thread or debuger can overwrite the trailing '\0' in between. Afaics nothing really bad can happen because we must always have the null-terminated bprm->filename copied by the 1st copy_strings_kernel(), but perhaps we should change this code to check "bprm->p < bprm->exec" anyway, and/or change copy_strings() to ensure that the last byte in string is always zero.

Reported by: hujunjie <[email protected]>

Signed-off-by Oleg Nesterov <[email protected]>

---
 fs/exec.c | 7 -------
 1 file changed, 7 deletions(-)

fs/exec.c +0 -7

index c4010b8..9b85c4d 100644
@@ -243,10 +243,6 @@ static void put_arg_page(struct page *page)
 	put_page(page);
 }
 
-static void free_arg_page(struct linux_binprm *bprm, int i)
-{
-}
-
 static void free_arg_pages(struct linux_binprm *bprm)
 {
 }

@@ -1481,9 +1477,6 @@ int remove_arg_zero(struct linux_binprm *bprm)
 
 		kunmap_atomic(kaddr);
 		put_arg_page(page);
-
-		if (offset == PAGE_SIZE)
-			free_arg_page(bprm, (bprm->p >> PAGE_SHIFT) - 1);
 	} while (offset == PAGE_SIZE);
 
 	bprm->p++;

Signature · 8 lines

2.5.0


--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to [email protected].  For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"[email protected]"> [email protected] </a>