arm64: ptrace: fix empty registers set in prstatus of aarch32 process core

Hi Catalin, Will,

I've run into situation when core of aarch32 (V7) process created by V8 kernel has empty registers set saved in prstatus notes. As result core file are broken. Below are details on the issues. Proposed patch follows this cover letter.

Test Case For Problem
---------------------

V8 kernel runs with v7 rootfs. Below output for BE system but LE has the same issue. hello test case compiled as v7 code. Notice that all registers in core file are zeros.

root at genericarmv7ab:~# cat hello.c
#include <stdio.h>

int main (void)

{
    printf("Hello, world\n");
    *(int *) 0 = 0;
}

root at genericarmv7ab:~# ./hello
Hello, world
hello[825]: unhandled level 3 translation fault (11) at 0x00000000, esr 0x92000047 pgd = ffffffc07cb02000
[00000000] *pgd=00000000fcad2003, *pmd=00000000fcad9003, *pte=0000000000000000

CPU: 0 PID: 825 Comm: hello Not tainted 3.15.0-rc5+ #1
task: ffffffc07dbf0000 ti: ffffffc07c978000 task.ti: ffffffc07c978000
PC is at 0x8434
LR is at 0xf74e03c4
pc : [<0000000000008434>] lr : [<00000000f74e03c4>] pstate: 600f0210

sp : 00000000ff9ce128
x12: 000000000000000d
x11: 00000000ff9ce12c x10: 00000000f75d2000
x9 : 0000000000000000 x8 : 0000000000000000
x7 : 0000000000000000 x6 : 00000000000082f4
x5 : 0000000000000000 x4 : 00000000ff9ce148
x3 : 0000000000000000 x2 : 0000000000000000
x1 : 0000000000000000 x0 : 000000000000000d

Segmentation fault (core dumped)
root at genericarmv7ab:~# gdb hello
GNU gdb (Linaro GDB) 7.6.1-2013.10
Copyright (C) 2013 Free Software Foundation, Inc.
License GPLv3+: GNU GPL version 3 or later <http://gnu.org/licenses/gpl.html>
This is free software: you are free to change and redistribute it. There is NO WARRANTY, to the extent permitted by law. Type "show copying" and "show warranty" for details.
This GDB was configured as "armeb-oe-linux-gnueabi".
For bug reporting instructions, please see:
<http://bugs.launchpad.net/gdb-linaro/>...
Reading symbols from /home/root/hello...done.
(gdb) target core core
[New LWP 825]
Program terminated with signal 11, Segmentation fault.
#0 0x00000000 in ?? ()
(gdb) info reg

r0             0x0      0
r1             0x0      0
r2             0x0      0
r3             0x0      0
r4             0x0      0
r5             0x0      0
r6             0x0      0
r7             0x0      0
r8             0x0      0
r9             0x0      0
r10            0x0      0
r11            0x0      0
r12            0x0      0
sp             0x0      0x0
lr             0x0      0
pc             0x0      0x0
cpsr           0x0      0

(gdb)

If one look at core file notes it can see that all registers in PRSTATUS note are zeros:

[root at kamensky-w530 root]# eu-readelf --notes ~kamensky/v8v7-core/v8_v7_le_core

Note segment of 984 bytes at offset 0x1f4:
Owner Data size Type
CORE 148 PRSTATUS

    info.si_signo: 11, info.si_code: 0, info.si_errno: 0, cursig: 11
    sigpend: <>
    sighold: <>
    pid: 1358, ppid: 1351, pgrp: 1358, sid: 1349
    utime: 0.000000, stime: 0.020000, cutime: 0.000000, cstime: 0.000000
    orig_r0: 0, fpvalid: 0
    r0:             0  r1:             0  r2:             0
    r3:             0  r4:             0  r5:             0
    r6:             0  r7:             0  r8:             0
    r9:             0  r10:            0  r11:            0
    r12:            0  sp:    0x00000000  lr:    0x00000000
    pc:    0x00000000  spsr:  0x00000000

Writing prstatus code under debugger
------------------------------------

Here is backtrace of code that writes PRSTATUS note. compat_gpr_get called to fill in registers value. Note kbuf has valid address, ubuf is 0.

#0 compat_gpr_get( target = (struct task_struct*) 0xFFFFFFC07CB39400, regset = (const struct user_regset*) 0xFFFFFFC0003D5290, pos = 0, count = 72, kbuf = (void*) 0xFFFFFFC07CFE4158, ubuf = (void*) 0x0 ) at ptrace.c:622
#1 fill_note( data = (void*) 0xFFFFFFC07CFE4110, sz = 148, type = 1, name = 0xFFFFFFC0004C4670 "CORE", note = (struct memelfnote*) 0xFFFFFFC07CFE41A8 ) at binfmt_elf.c:1279
#2 fill_thread_core_info( view = (const struct user_regset_view*) 0xFFFFFFC0003D5300, view = (const struct user_regset_view*) 0xFFFFFFC0003D5300, total = (size_t*) 0xFFFFFFC07C8B3B18, signr = <Value currently has no location>, t = (struct elf_thread_core_info*) 0xFFFFFFC07CFE4100 ) at binfmt_elf.c:1541 #3 elf_core_dump( cprm = (struct coredump_params*) 0xFFFFFFC07C8B3C38 ) at binfmt_elf.c:2123
#4 write_note_info( cprm = (struct coredump_params*) 0xFFFFFFC07C8B3C38, info = (struct elf_note_info*) 0xFFFFFFFFFFFFFFF8 ) at binfmt_elf.c:1698
#5 do_coredump( siginfo = <Value optimised away by compiler> ) at coredump.c:667
#6 get_signal_to_deliver( info = (siginfo_t*) 0xFFFFFFC07C8B3DF0, return_ka = (struct k_sigaction*) 0xFFFFFFC07C8B3E70, regs = <Value optimised away by compiler>, cookie = <Value optimised away by compiler> ) at signal.c:2372
#7 do_signal( regs = (struct pt_regs*) 0xFFFFFFC07C8B3ED0 ) at signal.c:377
#8 do_notify_resume( regs = <Value optimised away by compiler>, thread_flags = <Value optimised away by compiler> ) at signal.c:413
#9 [/wd1/linaro/linux-build/_le_64_linus_062/vmlinux EL1N:0xFFFFFFC00008428C ]
p ubuf
$30 = (void*) 0x0
p kbuf
$31 = (void*) 0xFFFFFFC07CFE4158

Fix
---

In compat_gpr_get and compat_gpr_set functions
handle kbuf and ubuf as two separate cases as other similar functions do, i.e user_regset_copyout, user_regset_copyin functions

Tests
-----

Tested core files or single threaded and multithread aarch32 processes. Able to see registers value, execute backtrace command, etc.

Thanks,
Victor

Victor Kamensky (1):
arm64: ptrace: fix empty registers set in prstatus of aarch32 process
core

 arch/arm64/kernel/ptrace.c | 28 +++++++++++++++++++---------
 1 file changed, 19 insertions(+), 9 deletions(-)

1.8.1.4

Currently core file of aarch32 process prstatus note has empty registers set. As result aarch32 core files create by V8 kernel are not very useful.

It happens because compat_gpr_get and compat_gpr_set functions can copy registers values to/from either kbuf or ubuf. ELF core file collection function fill_thread_core_info calls compat_gpr_get with kbuf set and ubuf set to 0. But current compat_gpr_get and compat_gpr_set function handle copy to/from only ubuf case.

Fix is to handle kbuf and ubuf as two separate cases in similar way as other functions like user_regset_copyout, user_regset_copyin do.

---
 arch/arm64/kernel/ptrace.c | 28 +++++++++++++++++++---------
 1 file changed, 19 insertions(+), 9 deletions(-)

index 6a8928b..9c9c2b9 100644
@@ -650,11 +650,16 @@ static int compat_gpr_get(struct task_struct *target,
 			reg = task_pt_regs(target)->regs[idx];
 		}
 
-		ret = copy_to_user(ubuf, &reg, sizeof(reg));
-		if (ret)
-			break;
-
-		ubuf += sizeof(reg);
+		if (kbuf) {
+			memcpy(kbuf, &reg, sizeof(reg));
+			kbuf += sizeof(reg);
+		} else {
+			ret = copy_to_user(ubuf, &reg, sizeof(reg));
+			if (ret)
+				break;
+
+			ubuf += sizeof(reg);
+		}
 	}
 
 	return ret;

@@ -684,11 +689,16 @@ static int compat_gpr_set(struct task_struct *target,
 		unsigned int idx = start + i;
 		compat_ulong_t reg;
 
-		ret = copy_from_user(&reg, ubuf, sizeof(reg));
-		if (ret)
-			return ret;
+		if (kbuf) {
+			memcpy(&reg, kbuf, sizeof(reg));
+			kbuf += sizeof(reg);
+		} else {
+			ret = copy_from_user(&reg, ubuf, sizeof(reg));
+			if (ret)
+				return ret;
 
-		ubuf += sizeof(reg);
+			ubuf += sizeof(reg);
+		}
 
 		switch (idx) {
 		case 15:

1.8.1.4

Hi Victor,

Thanks for both the fix and the detailed explanation!

An alternative is to use set_fs when kbuf is set, then use copy_{to,from}_user for everything. However, given how ugly I find set_fs to start with, your patch looks good to me:

We probably want a CC stable too.

Cheers,

Will

Hi Will,

Thank you for review. Please forgive my naive question, I've tried to google it, but does not look I do a good job. Is there any special thing I need to do, so you or Catalin would pick this up? I've posted patches to Russell's patch system before but never dealt with arm64 patches.

Or I just need to repost the patch to linux-arm-kernel with your 'Acked-by' and Cc: stable at vger.kernel.org in it?

Thanks,
Victor

Yes, please. Just repost the patch with those tags and we'll pick it up after the merge window. If you notice that we've forgotten, please give us a prod!

Will

Also don't forget to cc us on the email, it's quicker this way.

Thanks.

Catalin