Randy reported that objtool got stuck in an infinite loop when
processing drivers/i2c/busses/i2c-parport.o. It was caused by the
following code:
00000000000001fd <line_set>:
1fd: 48 b8 00 00 00 00 00 movabs $0x0,%rax
204: 00 00 00
1ff: R_X86_64_64 .rodata-0x8
207: 41 55 push %r13
209: 41 89 f5 mov %esi,%r13d
20c: 41 54 push %r12
20e: 49 89 fc mov %rdi,%r12
211: 55 push %rbp
212: 48 89 d5 mov %rdx,%rbp
215: 53 push %rbx
216: 0f b6 5a 01 movzbl 0x1(%rdx),%ebx
21a: 48 8d 34 dd 00 00 00 lea 0x0(,%rbx,8),%rsi
221: 00
21e: R_X86_64_32S .rodata
222: 48 89 f1 mov %rsi,%rcx
225: 48 29 c1 sub %rax,%rcx
find_jump_table() saw the .rodata reference and tried to find a jump
table associated with it (though there wasn't one). The -0x8 rela
addend is unusual. It caused find_jump_table() to send a negative
table_offset (unsigned 0xfffffffffffffff8) to find_rela_by_dest().
The negative offset should have been harmless, but it actually threw
for_offset_range() for a loop... literally. When the mask value got
incremented past the end value, it also wrapped to zero, causing the
loop exit condition to remain true forever.
Prevent this scenario from happening by ensuring the incremented value
is always >= the starting value.
Fixes: 74b873e49d92 ("objtool: Optimize find_rela_by_dest_range()")
Reported-by: Randy Dunlap <[email protected]>
Signed-off-by: Josh Poimboeuf <[email protected]>
---
tools/objtool/elf.h | 7 ++++---
1 file changed, 4 insertions(+), 3 deletions(-)
diff --git a/tools/objtool/elf.h b/tools/objtool/elf.h
index 5e76ac38cf99..f753148f5dac 100644
--- a/tools/objtool/elf.h
+++ b/tools/objtool/elf.h
@@ -89,9 +89,10 @@ struct elf {
#define OFFSET_STRIDE (1UL << OFFSET_STRIDE_BITS)
#define OFFSET_STRIDE_MASK (~(OFFSET_STRIDE - 1))
-#define for_offset_range(_offset, _start, _end) \
- for (_offset = ((_start) & OFFSET_STRIDE_MASK); \
- _offset <= ((_end) & OFFSET_STRIDE_MASK); \
+#define for_offset_range(_offset, _start, _end) \
+ for (_offset = ((_start) & OFFSET_STRIDE_MASK); \
+ _offset >= ((_start) & OFFSET_STRIDE_MASK) && \
+ _offset <= ((_end) & OFFSET_STRIDE_MASK); \
_offset += OFFSET_STRIDE)
static inline u32 sec_offset_hash(struct section *sec, unsigned long offset)
--
2.21.1
On 4/25/20 12:19 PM, Josh Poimboeuf wrote:
> Randy reported that objtool got stuck in an infinite loop when
> processing drivers/i2c/busses/i2c-parport.o. It was caused by the
> following code:
>
> 00000000000001fd <line_set>:
> 1fd: 48 b8 00 00 00 00 00 movabs $0x0,%rax
> 204: 00 00 00
> 1ff: R_X86_64_64 .rodata-0x8
> 207: 41 55 push %r13
> 209: 41 89 f5 mov %esi,%r13d
> 20c: 41 54 push %r12
> 20e: 49 89 fc mov %rdi,%r12
> 211: 55 push %rbp
> 212: 48 89 d5 mov %rdx,%rbp
> 215: 53 push %rbx
> 216: 0f b6 5a 01 movzbl 0x1(%rdx),%ebx
> 21a: 48 8d 34 dd 00 00 00 lea 0x0(,%rbx,8),%rsi
> 221: 00
> 21e: R_X86_64_32S .rodata
> 222: 48 89 f1 mov %rsi,%rcx
> 225: 48 29 c1 sub %rax,%rcx
>
> find_jump_table() saw the .rodata reference and tried to find a jump
> table associated with it (though there wasn't one). The -0x8 rela
> addend is unusual. It caused find_jump_table() to send a negative
> table_offset (unsigned 0xfffffffffffffff8) to find_rela_by_dest().
>
> The negative offset should have been harmless, but it actually threw
> for_offset_range() for a loop... literally. When the mask value got
> incremented past the end value, it also wrapped to zero, causing the
> loop exit condition to remain true forever.
>
> Prevent this scenario from happening by ensuring the incremented value
> is always >= the starting value.
>
> Fixes: 74b873e49d92 ("objtool: Optimize find_rela_by_dest_range()")
> Reported-by: Randy Dunlap <[email protected]>
> Signed-off-by: Josh Poimboeuf <[email protected]>
Hi Josh,
When applied to linux-next 20200414 (where it was reported) and using
config-r2092, objtool still loops (I killed it after 6 minutes of CPU time).
When applied to linux-next 20200424 and using the same config-r2092 file,
objtool appears to terminate normally and the entire build does also.
Acked-by: Randy Dunlap <[email protected]>
Tested-by: Randy Dunlap <[email protected]>
thanks.
> ---
> tools/objtool/elf.h | 7 ++++---
> 1 file changed, 4 insertions(+), 3 deletions(-)
>
> diff --git a/tools/objtool/elf.h b/tools/objtool/elf.h
> index 5e76ac38cf99..f753148f5dac 100644
> --- a/tools/objtool/elf.h
> +++ b/tools/objtool/elf.h
> @@ -89,9 +89,10 @@ struct elf {
> #define OFFSET_STRIDE (1UL << OFFSET_STRIDE_BITS)
> #define OFFSET_STRIDE_MASK (~(OFFSET_STRIDE - 1))
>
> -#define for_offset_range(_offset, _start, _end) \
> - for (_offset = ((_start) & OFFSET_STRIDE_MASK); \
> - _offset <= ((_end) & OFFSET_STRIDE_MASK); \
> +#define for_offset_range(_offset, _start, _end) \
> + for (_offset = ((_start) & OFFSET_STRIDE_MASK); \
> + _offset >= ((_start) & OFFSET_STRIDE_MASK) && \
> + _offset <= ((_end) & OFFSET_STRIDE_MASK); \
> _offset += OFFSET_STRIDE)
>
> static inline u32 sec_offset_hash(struct section *sec, unsigned long offset)
>
--
~Randy
On 4/25/20 1:04 PM, Randy Dunlap wrote:
> On 4/25/20 12:19 PM, Josh Poimboeuf wrote:
>> Randy reported that objtool got stuck in an infinite loop when
>> processing drivers/i2c/busses/i2c-parport.o. It was caused by the
>> following code:
>>
>> 00000000000001fd <line_set>:
>> 1fd: 48 b8 00 00 00 00 00 movabs $0x0,%rax
>> 204: 00 00 00
>> 1ff: R_X86_64_64 .rodata-0x8
>> 207: 41 55 push %r13
>> 209: 41 89 f5 mov %esi,%r13d
>> 20c: 41 54 push %r12
>> 20e: 49 89 fc mov %rdi,%r12
>> 211: 55 push %rbp
>> 212: 48 89 d5 mov %rdx,%rbp
>> 215: 53 push %rbx
>> 216: 0f b6 5a 01 movzbl 0x1(%rdx),%ebx
>> 21a: 48 8d 34 dd 00 00 00 lea 0x0(,%rbx,8),%rsi
>> 221: 00
>> 21e: R_X86_64_32S .rodata
>> 222: 48 89 f1 mov %rsi,%rcx
>> 225: 48 29 c1 sub %rax,%rcx
>>
>> find_jump_table() saw the .rodata reference and tried to find a jump
>> table associated with it (though there wasn't one). The -0x8 rela
>> addend is unusual. It caused find_jump_table() to send a negative
>> table_offset (unsigned 0xfffffffffffffff8) to find_rela_by_dest().
>>
>> The negative offset should have been harmless, but it actually threw
>> for_offset_range() for a loop... literally. When the mask value got
>> incremented past the end value, it also wrapped to zero, causing the
>> loop exit condition to remain true forever.
>>
>> Prevent this scenario from happening by ensuring the incremented value
>> is always >= the starting value.
>>
>> Fixes: 74b873e49d92 ("objtool: Optimize find_rela_by_dest_range()")
>> Reported-by: Randy Dunlap <[email protected]>
>> Signed-off-by: Josh Poimboeuf <[email protected]>
>
> Hi Josh,
>
> When applied to linux-next 20200414 (where it was reported) and using
> config-r2092, objtool still loops (I killed it after 6 minutes of CPU time).
Nope, scratch that. Operator error. Sorry.
> When applied to linux-next 20200424 and using the same config-r2092 file,
> objtool appears to terminate normally and the entire build does also.
>
> Acked-by: Randy Dunlap <[email protected]>
> Tested-by: Randy Dunlap <[email protected]>
>
> thanks.
>
>> ---
--
~Randy
The following commit has been merged into the x86/urgent branch of tip:
Commit-ID: 53fb6e990d782ded62d7c76d566e107c03393b74
Gitweb: https://git.kernel.org/tip/53fb6e990d782ded62d7c76d566e107c03393b74
Author: Josh Poimboeuf <[email protected]>
AuthorDate: Sat, 25 Apr 2020 14:19:01 -05:00
Committer: Ingo Molnar <[email protected]>
CommitterDate: Sun, 26 Apr 2020 09:28:14 +02:00
objtool: Fix infinite loop in for_offset_range()
Randy reported that objtool got stuck in an infinite loop when
processing drivers/i2c/busses/i2c-parport.o. It was caused by the
following code:
00000000000001fd <line_set>:
1fd: 48 b8 00 00 00 00 00 movabs $0x0,%rax
204: 00 00 00
1ff: R_X86_64_64 .rodata-0x8
207: 41 55 push %r13
209: 41 89 f5 mov %esi,%r13d
20c: 41 54 push %r12
20e: 49 89 fc mov %rdi,%r12
211: 55 push %rbp
212: 48 89 d5 mov %rdx,%rbp
215: 53 push %rbx
216: 0f b6 5a 01 movzbl 0x1(%rdx),%ebx
21a: 48 8d 34 dd 00 00 00 lea 0x0(,%rbx,8),%rsi
221: 00
21e: R_X86_64_32S .rodata
222: 48 89 f1 mov %rsi,%rcx
225: 48 29 c1 sub %rax,%rcx
find_jump_table() saw the .rodata reference and tried to find a jump
table associated with it (though there wasn't one). The -0x8 rela
addend is unusual. It caused find_jump_table() to send a negative
table_offset (unsigned 0xfffffffffffffff8) to find_rela_by_dest().
The negative offset should have been harmless, but it actually threw
for_offset_range() for a loop... literally. When the mask value got
incremented past the end value, it also wrapped to zero, causing the
loop exit condition to remain true forever.
Prevent this scenario from happening by ensuring the incremented value
is always >= the starting value.
Fixes: 74b873e49d92 ("objtool: Optimize find_rela_by_dest_range()")
Reported-by: Randy Dunlap <[email protected]>
Tested-by: Randy Dunlap <[email protected]>
Acked-by: Randy Dunlap <[email protected]>
Signed-off-by: Josh Poimboeuf <[email protected]>
Signed-off-by: Ingo Molnar <[email protected]>
Cc: Julien Thierry <[email protected]>
Cc: Miroslav Benes <[email protected]>
Cc: Peter Zijlstra <[email protected]>
Link: https://lore.kernel.org/r/02b719674b031800b61e33c30b2e823183627c19.1587842122.git.jpoimboe@redhat.com
---
tools/objtool/elf.h | 7 ++++---
1 file changed, 4 insertions(+), 3 deletions(-)
diff --git a/tools/objtool/elf.h b/tools/objtool/elf.h
index ebbb10c..c227a2e 100644
--- a/tools/objtool/elf.h
+++ b/tools/objtool/elf.h
@@ -87,9 +87,10 @@ struct elf {
#define OFFSET_STRIDE (1UL << OFFSET_STRIDE_BITS)
#define OFFSET_STRIDE_MASK (~(OFFSET_STRIDE - 1))
-#define for_offset_range(_offset, _start, _end) \
- for (_offset = ((_start) & OFFSET_STRIDE_MASK); \
- _offset <= ((_end) & OFFSET_STRIDE_MASK); \
+#define for_offset_range(_offset, _start, _end) \
+ for (_offset = ((_start) & OFFSET_STRIDE_MASK); \
+ _offset >= ((_start) & OFFSET_STRIDE_MASK) && \
+ _offset <= ((_end) & OFFSET_STRIDE_MASK); \
_offset += OFFSET_STRIDE)
static inline u32 sec_offset_hash(struct section *sec, unsigned long offset)