Skip to contentSkip to conversation

pipe: handle zero sized F_SETPIPE_SZ fcntl

PATCH In review Nicolai Stange 2 participants 2 messages Feb 19, 2017 – Feb 20, 2017 VFS

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 pipe: handle zero sized F_SETPIPE_SZ fcntl

Syzkaller reported this

  UBSAN: Undefined behaviour in include/linux/log2.h:63:13
  shift exponent 64 is too large for 64-bit type 'long unsigned int'
  Call Trace:
   __dump_stack lib/dump_stack.c:15 [inline]
   dump_stack+0x108/0x19b lib/dump_stack.c:51
   ubsan_epilogue+0x12/0x8f lib/ubsan.c:164
   __ubsan_handle_shift_out_of_bounds+0x29c/0x300 lib/ubsan.c:421
   __roundup_pow_of_two include/linux/log2.h:63 [inline]
   round_pipe_size fs/pipe.c:1049 [inline]
   pipe_set_size fs/pipe.c:1042 [inline]
   pipe_fcntl+0x12a/0x6e0 fs/pipe.c:1159
   do_fcntl fs/fcntl.c:332 [inline]
   SYSC_fcntl fs/fcntl.c:372 [inline]
   SyS_fcntl+0xa90/0xb80 fs/fcntl.c:380
   entry_SYSCALL_64_fastpath+0x1f/0xc2

after it had tried to do this:

  fcntl(fd, F_SETPIPE_SZ, 0);

The reason is that round_pipe_size() rounds the requested size towards the next multiple of PAGE_SIZE and passes the thus resulting number of pages to roundup_pow_of_two(). Now, roundup_pow_of_two(0) is defined to be undefined. More specifically,

roundup_pow_of_two(0) == 1UL << fls_long(-1) == 1UL << BITS_PER_LONG

which evaluates to 1 on x86 (good), but to zero on ARMv7 (not good), for example.

Whatever the resulting value is, pipe_set_size() will hand it to kcalloc(), so it should better be nonzero.

Make round_pipe_size() handle the size == 0 case explicitly and let it return PAGE_SIZE then.

---
 fs/pipe.c | 3 +++
 1 file changed, 3 insertions(+)
fs/pipe.c +3 -0
index 73b84baf58f8..2c88fa8cbe3b 100644
@@ -1024,6 +1024,9 @@ static inline unsigned int round_pipe_size(unsigned int size)
 {
 	unsigned long nr_pages;
 
+	if (!size)
+		return PAGE_SIZE;
+
 	nr_pages = (size + PAGE_SIZE - 1) >> PAGE_SHIFT;
 	return roundup_pow_of_two(nr_pages) << PAGE_SHIFT;
 }
Signature · 1 line
2.11.1