Skip to contentSkip to conversation

Some TIF_IA32 fixes

PATCH In review Andy Lutomirski 3 participants 7 messages Feb 25, 2015

Subsystems (2)
lore
Share & tools
lore
#0 lore raw

cover Some TIF_IA32 fixes

At some point I'd like to remove TIF_IA32 entirely and replace it with just user_64bit_mode and a per-mm configuration. In the mean time, this fixes some incorrect uses.

Ingo, Thomas, Peter, if you like these, please ack the first one and apply the other two. The first one will conflict with the pile of entry changes that I'll send a pull request for in a few days. (Alternatively, you can take all three and fix the conflict yourself -- it's trivial.)

Andy Lutomirski (3):
x86_64, entry: Remove a bogus ret_from_fork optimization x86_64, ptrace: Remove checks for TIF_IA32 when changing cs and ss x86_64, copy_thread: Fix CLONE_SETTLS bitness

 arch/x86/kernel/entry_64.S   | 13 ++++++++-----
 arch/x86/kernel/process_64.c |  2 +-
 arch/x86/kernel/ptrace.c     | 10 ++--------
 3 files changed, 11 insertions(+), 14 deletions(-)
Signature · 1 line
2.3.0
#1 lore raw

PATCH 1/3 x86_64, entry: Remove a bogus ret_from_fork optimization

ret_from_fork checks TIF_IA32 to determine whether pt_regs and the related state make sense for ret_from_sys_call. This is entirely the wrong check. TS_COMPAT would make a little more sense, but there's really no point in keeping this optimization at all.

This fixes a return to the wrong user CS if we came from int 0x80 in a 64-bit task.

---
 arch/x86/kernel/entry_64.S | 13 ++++++++-----
 1 file changed, 8 insertions(+), 5 deletions(-)
arch/x86/kernel/entry_64.S +8 -5
index 9ebaf63ba182..4ee9a2315f80 100644
@@ -334,11 +334,14 @@ ENTRY(ret_from_fork)
 	testl $3, CS-ARGOFFSET(%rsp)		# from kernel_thread?
 	jz   1f
 
-	testl $_TIF_IA32, TI_flags(%rcx)	# 32-bit compat task needs IRET
-	jnz  int_ret_from_sys_call
-
-	RESTORE_TOP_OF_STACK %rdi, -ARGOFFSET
-	jmp ret_from_sys_call			# go to the SYSRET fastpath
+	/*
+	 * By the time we get here, we have no idea whether our pt_regs,
+	 * ti flags, and ti status came from the 64-bit SYSCALL fast path,
+	 * the slow path, or one of the ia32entry paths.
+	 * Use int_ret_from_sys_call to return, since it can safely handle
+	 * all of the above.
+	 */
+	jmp  int_ret_from_sys_call
 
 1:
 	subq $REST_SKIP, %rsp	# leave space for volatiles
Signature · 1 line
2.3.0
Acked-by on v1 1/3
#2 lore raw

Re: [PATCH 1/3] x86_64, entry: Remove a bogus ret_from_fork optimization

Quoted 9 lines · Andy Lutomirski

ret_from_fork checks TIF_IA32 to determine whether pt_regs and the related state make sense for ret_from_sys_call. This is entirely the wrong check. TS_COMPAT would make a little more sense, but there's really no point in keeping this optimization at all.

This fixes a return to the wrong user CS if we came from int 0x80 in a 64-bit task.

Thanks,

Ingo

#3 lore raw

PATCH 2/3 x86_64, ptrace: Remove checks for TIF_IA32 when changing cs and ss

In reply to Andy Lutomirski

The ability for modified cs and/or ss to be useful has nothing to do with TIF_IA32. Similarly, if there's an exploit involving changing cs or ss, it's exploitable with or without a TIF_IA32 check.

So just delete the check.

---
 arch/x86/kernel/ptrace.c | 10 ++--------
 1 file changed, 2 insertions(+), 8 deletions(-)
arch/x86/kernel/ptrace.c +2 -8
index e510618b2e91..1e125817cf9f 100644
@@ -364,18 +364,12 @@ static int set_segment_reg(struct task_struct *task,
 	case offsetof(struct user_regs_struct,cs):
 		if (unlikely(value == 0))
 			return -EIO;
-#ifdef CONFIG_IA32_EMULATION
-		if (test_tsk_thread_flag(task, TIF_IA32))
-			task_pt_regs(task)->cs = value;
-#endif
+		task_pt_regs(task)->cs = value;
 		break;
 	case offsetof(struct user_regs_struct,ss):
 		if (unlikely(value == 0))
 			return -EIO;
-#ifdef CONFIG_IA32_EMULATION
-		if (test_tsk_thread_flag(task, TIF_IA32))
-			task_pt_regs(task)->ss = value;
-#endif
+		task_pt_regs(task)->ss = value;
 		break;
 	}
 
Signature · 1 line
2.3.0
#4 lore raw

[tip:x86/asm] x86/ptrace: Remove checks for TIF_IA32 when changing CS and SS

x86/ptrace: Remove checks for TIF_IA32 when changing CS and SS
Commit-ID:  08571f1ae327bfb631cb7171bde5ea605df626c6
Gitweb:     http://git.kernel.org/tip/08571f1ae327bfb631cb7171bde5ea605df626c6
Author:     Andy Lutomirski <[email protected]>
AuthorDate: Tue, 24 Feb 2015 16:01:38 -0800
Committer:  Ingo Molnar <[email protected]>
CommitDate: Wed, 25 Feb 2015 08:27:49 +0100

x86/ptrace: Remove checks for TIF_IA32 when changing CS and SS

The ability for modified CS and/or SS to be useful has nothing to do with TIF_IA32. Similarly, if there's an exploit involving changing CS or SS, it's exploitable with or without a TIF_IA32 check.

So just delete the check.

---
 arch/x86/kernel/ptrace.c | 10 ++--------
 1 file changed, 2 insertions(+), 8 deletions(-)
arch/x86/kernel/ptrace.c +2 -8
index e510618..1e12581 100644
@@ -364,18 +364,12 @@ static int set_segment_reg(struct task_struct *task,
 	case offsetof(struct user_regs_struct,cs):
 		if (unlikely(value == 0))
 			return -EIO;
-#ifdef CONFIG_IA32_EMULATION
-		if (test_tsk_thread_flag(task, TIF_IA32))
-			task_pt_regs(task)->cs = value;
-#endif
+		task_pt_regs(task)->cs = value;
 		break;
 	case offsetof(struct user_regs_struct,ss):
 		if (unlikely(value == 0))
 			return -EIO;
-#ifdef CONFIG_IA32_EMULATION
-		if (test_tsk_thread_flag(task, TIF_IA32))
-			task_pt_regs(task)->ss = value;
-#endif
+		task_pt_regs(task)->ss = value;
 		break;
 	}
 
#5 lore raw

PATCH 3/3 x86_64, copy_thread: Fix CLONE_SETTLS bitness

In reply to Andy Lutomirski

CLONE_SETTLS is expected to write a TLS entry in the GDT for 32-bit callers and to set fsbase for 64-bit callers.

The correct check is is_ia32_task(), which returns true in the context of a 32-bit syscall. TIF_IA32 is set if the task itself has a 32-bit personality, which is not the same thing.

---
 arch/x86/kernel/process_64.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
arch/x86/kernel/process_64.c +1 -1
index 5a2c02913af3..936d43461dca 100644
@@ -207,7 +207,7 @@ int copy_thread(unsigned long clone_flags, unsigned long sp,
 	 */
 	if (clone_flags & CLONE_SETTLS) {
 #ifdef CONFIG_IA32_EMULATION
-		if (test_thread_flag(TIF_IA32))
+		if (is_ia32_task())
 			err = do_set_thread_area(p, -1,
 				(struct user_desc __user *)childregs->si, 0);
 		else
Signature · 1 line
2.3.0
#6 lore raw

[tip:x86/asm] x86/ia32-compat: Fix CLONE_SETTLS bitness of copy_thread()

x86/ia32-compat: Fix CLONE_SETTLS bitness of copy_thread()
Commit-ID:  72c6fb4f74b6b3797f5b1abd6944d7a1d2adbf04
Gitweb:     http://git.kernel.org/tip/72c6fb4f74b6b3797f5b1abd6944d7a1d2adbf04
Author:     Andy Lutomirski <[email protected]>
AuthorDate: Tue, 24 Feb 2015 16:01:39 -0800
Committer:  Ingo Molnar <[email protected]>
CommitDate: Wed, 25 Feb 2015 08:27:50 +0100

x86/ia32-compat: Fix CLONE_SETTLS bitness of copy_thread()

CLONE_SETTLS is expected to write a TLS entry in the GDT for 32-bit callers and to set FSBASE for 64-bit callers.

The correct check is is_ia32_task(), which returns true in the context of a 32-bit syscall. TIF_IA32 is set if the task itself has a 32-bit personality, which is not the same thing.

---
 arch/x86/kernel/process_64.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
arch/x86/kernel/process_64.c +1 -1
index 5a2c029..936d434 100644
@@ -207,7 +207,7 @@ int copy_thread(unsigned long clone_flags, unsigned long sp,
 	 */
 	if (clone_flags & CLONE_SETTLS) {
 #ifdef CONFIG_IA32_EMULATION
-		if (test_thread_flag(TIF_IA32))
+		if (is_ia32_task())
 			err = do_set_thread_area(p, -1,
 				(struct user_desc __user *)childregs->si, 0);
 		else