[PATCH v2] x86/ptrace: Clean up PTRACE_GETREGS/PTRACE_PUTREGS regset selection

From: Andy Lutomirski
Date: Wed Feb 03 2021 - 13:11:04 EST

task_user_regset_view() has nonsensical semantics, but those semantics
appear to be relied on by existing users of PTRACE_GETREGSET and
PTRACE_SETREGSET. (See added comments in this patch for details.)

It shouldn't be used for PTRACE_GETREGS or PTRACE_SETREGS, though. A
native 64-bit ptrace() call and an x32 ptrace() call using GETREGS or
SETREGS wants the 64-bit regset views, and a 32-bit ptrace() call
(native or compat) should use the 32-bit regset.
task_user_regset_view() almost does this except that it will
malfunction if a ptracer is itself ptraced and the outer ptracer
modifies CS on entry to a ptrace() syscall. Hopefully that has never
happened. (The compat ptrace() code already hardcoded the 32-bit
regset, so this patch has no effect on that path.)

Improve the situation and deobfuscate the code by hardcoding the
64-bit view in the x32 ptrace() and selecting the view based on the
kernel config in the native ptrace().

I tried to figure out the history behind this API. I naïvely assumed
that PTRAGE_GETREGSET and PTRACE_SETREGSET were ancient APIs that
predated compat, but no. They were introduced by commit 2225a122ae26
("ptrace: Add support for generic PTRACE_GETREGSET/PTRACE_SETREGSET")
in 2010, and they are simply a poor design. ELF core dumps have the
ELF e_machine field and a bunch of register sets in ELF notes, and the
pair (e_machine, NT_XXX) indicates the format of the regset blob. But
the new PTRACE_GET/SETREGSET API coopted the NT_XXX numbering without
any way to specify which e_machine was in effect. This is especially
bad on x86, where a process can freely switch between 32-bit and
64-bit mode, and, in fact, the PTRAGE_SETREGSET call itself can cause
this switch to happen. Oops.

Signed-off-by: Andy Lutomirski <luto@xxxxxxxxxx>
---
arch/x86/kernel/ptrace.c | 46 +++++++++++++++++++++++++++++++++-------
1 file changed, 38 insertions(+), 8 deletions(-)

diff --git a/arch/x86/kernel/ptrace.c b/arch/x86/kernel/ptrace.c
index bedca011459c..87a4143aa7d7 100644
--- a/arch/x86/kernel/ptrace.c
+++ b/arch/x86/kernel/ptrace.c
@@ -704,6 +704,9 @@ void ptrace_disable(struct task_struct *child)
#if defined CONFIG_X86_32 || defined CONFIG_IA32_EMULATION
static const struct user_regset_view user_x86_32_view; /* Initialized below. */
#endif
+#ifdef CONFIG_X86_64
+static const struct user_regset_view user_x86_64_view; /* Initialized below. */
+#endif

long arch_ptrace(struct task_struct *child, long request,
unsigned long addr, unsigned long data)
@@ -711,6 +714,14 @@ long arch_ptrace(struct task_struct *child, long request,
int ret;
unsigned long __user *datap = (unsigned long __user *)data;

+#ifdef CONFIG_X86_64
+ /* This is native 64-bit ptrace() */
+ const struct user_regset_view *regset_view = &user_x86_64_view;
+#else
+ /* This is native 32-bit ptrace() */
+ const struct user_regset_view *regset_view = &user_x86_32_view;
+#endif
+
switch (request) {
/* read the word at location addr in the USER area. */
case PTRACE_PEEKUSR: {
@@ -749,28 +760,28 @@ long arch_ptrace(struct task_struct *child, long request,

case PTRACE_GETREGS: /* Get all gp regs from the child. */
return copy_regset_to_user(child,
- task_user_regset_view(current),
+ regset_view,
REGSET_GENERAL,
0, sizeof(struct user_regs_struct),
datap);

case PTRACE_SETREGS: /* Set all gp regs in the child. */
return copy_regset_from_user(child,
- task_user_regset_view(current),
+ regset_view,
REGSET_GENERAL,
0, sizeof(struct user_regs_struct),
datap);

case PTRACE_GETFPREGS: /* Get the child FPU state. */
return copy_regset_to_user(child,
- task_user_regset_view(current),
+ regset_view,
REGSET_FP,
0, sizeof(struct user_i387_struct),
datap);

case PTRACE_SETFPREGS: /* Set the child FPU state. */
return copy_regset_from_user(child,
- task_user_regset_view(current),
+ regset_view,
REGSET_FP,
0, sizeof(struct user_i387_struct),
datap);
@@ -1152,28 +1163,28 @@ static long x32_arch_ptrace(struct task_struct *child,

case PTRACE_GETREGS: /* Get all gp regs from the child. */
return copy_regset_to_user(child,
- task_user_regset_view(current),
+ &user_x86_64_view,
REGSET_GENERAL,
0, sizeof(struct user_regs_struct),
datap);

case PTRACE_SETREGS: /* Set all gp regs in the child. */
return copy_regset_from_user(child,
- task_user_regset_view(current),
+ &user_x86_64_view,
REGSET_GENERAL,
0, sizeof(struct user_regs_struct),
datap);

case PTRACE_GETFPREGS: /* Get the child FPU state. */
return copy_regset_to_user(child,
- task_user_regset_view(current),
+ &user_x86_64_view,
REGSET_FP,
0, sizeof(struct user_i387_struct),
datap);

case PTRACE_SETFPREGS: /* Set the child FPU state. */
return copy_regset_from_user(child,
- task_user_regset_view(current),
+ &user_x86_64_view,
REGSET_FP,
0, sizeof(struct user_i387_struct),
datap);
@@ -1309,6 +1320,25 @@ void __init update_regset_xstate_info(unsigned int size, u64 xstate_mask)
xstate_fx_sw_bytes[USER_XSTATE_XCR0_WORD] = xstate_mask;
}

+/*
+ * This is used by the core dump code to decide which regset to dump. The
+ * core dump code writes out the resulting .e_machine and the corresponding
+ * regsets. This is suboptimal if the task is messing around with its CS.L
+ * field, but at worst the core dump will end up missing some information.
+ *
+ * Unfortunately, it is also used by the broken PTRACE_GETREGSET and
+ * PTRACE_SETREGSET APIs. These APIs look at the .regsets field but have
+ * no way to make sure that the e_machine they use matches the caller's
+ * expectations. The result is that the data format returned by
+ * PTRACE_GETREGSET depends on the returned CS field (and even the offset
+ * of the returned CS field depends on its value!) and the data format
+ * accepted by PTRACE_SETREGSET is determined by the old CS value. The
+ * upshot is that it is basically impossible to use these APIs correctly.
+ *
+ * The best way to fix it in the long run would probably be to add new
+ * improved ptrace() APIs to read and write registers reliably, possibly by
+ * allowing userspace to select the ELF e_machine variant that they expect.
+ */
const struct user_regset_view *task_user_regset_view(struct task_struct *task)
{
#ifdef CONFIG_IA32_EMULATION
--
2.29.2