[ This is -mm only until Andi acks it ]
The 32-bit sysenter entry point mangles the sixth system call argument
for both 32-bit and 64-bit ptrace. In both cases, strace shows the
frame pointer (ebp) as the sixth argument.
Here's a snippet of a 64-bit strace of a 32-bit test program which
calls mmap through sysenter:
mmap2(NULL, 4096, PROT_READ|PROT_WRITE, MAP_PRIVATE|MAP_ANONYMOUS, -1, 0xfff00fcc) = 0xfffffffff7f7a000
fstat64(0x1, 0xfff008d8) = 0
mmap2(NULL, 4096, PROT_READ|PROT_WRITE, MAP_PRIVATE|MAP_ANONYMOUS, -1, 0xfff0089c) = 0xfffffffff7f79000
write(1, "mmap returns 0xf7f7a000\n", 24mmap returns 0xf7f7a000
) = 24
Here's a 32-bit strace of the same program:
mmap2(NULL, 4096, PROT_READ|PROT_WRITE, MAP_PRIVATE|MAP_ANONYMOUS, -1, 0xffc224ec) = 0xf7fcb000
fstat64(1, {st_mode=S_IFCHR|0620, st_rdev=makedev(136, 1), ...}) = 0
mmap2(NULL, 4096, PROT_READ|PROT_WRITE, MAP_PRIVATE|MAP_ANONYMOUS, -1, 0xffc21dbc) = 0xf7fca000
write(1, "mmap returns 0xf7fcb000\n", 24mmap returns 0xf7fcb000
) = 24
The first mmap is the one made by the test - its final argument (the
offset) is 0, but strace shows 0xfff00fcc, which is the value of ebp.
The second is a guilty bystander which is also showing the bug.
The patch below copies %r9 (where the sixth argument has been
stashed) into the RBP slot of pt_regs before syscall_trace_enter is
called. This fixes ptrace.
To allow a successful return to userspace, the original value of rbp
must be restored. This is done by storing the current value of rbp
into the RBP slot of pt_regs before the RESTORE_REST.
With this patch, the straces now look like this:
64-bit:
mmap2(NULL, 4096, PROT_READ|PROT_WRITE, MAP_PRIVATE|MAP_ANONYMOUS, -1, 0) = 0xfffffffff7f5a000
fstat64(0x1, 0xff926ee8) = 0
mmap2(NULL, 4096, PROT_READ|PROT_WRITE, MAP_PRIVATE|MAP_ANONYMOUS, -1, 0) = 0xfffffffff7f59000
write(1, "mmap returns 0xf7f5a000\n", 24mmap returns 0xf7f5a000
) = 24
32-bit:
mmap2(NULL, 4096, PROT_READ|PROT_WRITE, MAP_PRIVATE|MAP_ANONYMOUS, -1, 0) = 0xf7fa9000
fstat64(1, {st_mode=S_IFCHR|0620, st_rdev=makedev(136, 1), ...}) = 0
mmap2(NULL, 4096, PROT_READ|PROT_WRITE, MAP_PRIVATE|MAP_ANONYMOUS, -1, 0) = 0xf7fa8000
write(1, "mmap returns 0xf7fa9000\n", 24mmap returns 0xf7fa9000
) = 24
Signed-off-by: Jeff Dike <[email protected]>
--
arch/x86_64/ia32/ia32entry.S | 12 ++++++++++++
1 file changed, 12 insertions(+)
Index: linux-2.6/arch/x86_64/ia32/ia32entry.S
===================================================================
--- linux-2.6.orig/arch/x86_64/ia32/ia32entry.S
+++ linux-2.6/arch/x86_64/ia32/ia32entry.S
@@ -148,11 +148,23 @@ sysenter_do_call:
sysenter_tracesys:
CFI_RESTORE_STATE
SAVE_REST
+ /*
+ * We need the 6th system call argument to be in regs->rbp at
+ * this point so that ptrace will see it. It's in r9 now, so copy
+ * it to the rbp slot now.
+ */
+ movq %r9, RBP(%rsp)
CLEAR_RREGS
movq $-ENOSYS,RAX(%rsp) /* really needed? */
movq %rsp,%rdi /* &pt_regs -> arg1 */
call syscall_trace_enter
LOAD_ARGS ARGOFFSET /* reload args from stack in case ptrace changed it */
+ /*
+ * Now, we need the correct value of rbp to be restored. It
+ * was never munged, so we can save it to the rbp slot and
+ * just have it restored.
+ */
+ movq %rbp, RBP(%rsp)
RESTORE_REST
movl %ebp, %ebp
/* no need to do an access_ok check here because rbp has been
--
Work email - jdike at linux dot intel dot com