Re: [PATCH v6] mm: Uninline copy_overflow()

[Christophe Leroy]

Le 14/02/2022 à 15:00, David Laight a écrit :
> From: Christophe Leroy
>> Sent: 14 February 2022 13:21
>>
>> Le 14/02/2022 à 12:31, David Laight a écrit :
>>> From: Anshuman Khandual
>>>> Sent: 14 February 2022 09:54
>>> ...
>>>>> With -Winline, GCC tells:
>>>>>
>>>>> 	/include/linux/thread_info.h:212:20: warning: inlining failed in call to 'copy_overflow': call
>>>> is unlikely and code size would grow [-Winline]
>>>>>
>>>>> copy_overflow() is a non conditional warning called by
>>>>> check_copy_size() on an error path.
>>>>>
>>>>> check_copy_size() have to remain inlined in order to benefit
>>>>> from constant folding, but copy_overflow() is not worth inlining.
>>>>>
>>>>> Uninline the warning when CONFIG_BUG is selected.
>>>>>
>>>>> When CONFIG_BUG is not selected, WARN() does nothing so skip it.
>>>>>
>>>>> This reduces the size of vmlinux by almost 4kbytes.
>>>>
>>>
>>>>> +void __copy_overflow(int size, unsigned long count);
>>>>> +
>>>>>    static inline void copy_overflow(int size, unsigned long count)
>>>>>    {
>>>>> -	WARN(1, "Buffer overflow detected (%d < %lu)!\n", size, count);
>>>>> +	if (IS_ENABLED(CONFIG_BUG))
>>>>> +		__copy_overflow(size, count);
>>>>>    }
>>>
>>>> Just wondering, is this the only such scenario which results in
>>>> an avoidable bloated vmlinux image ?
>>>
>>> The more interesting question is whether the call to __copy_overflow()
>>> is actually significantly smaller than the one to WARN()?
>>> And if so why.
>>>
>> unsigned long tst_copy_to_user(void __user *to, unsigned long n)
>> {
>> 	return copy_to_user(to, &jiffies_64, n);
>> }
>>
>> With the patch:
>>
>> 00003c78 <tst_copy_to_user>:
>>       3c78:	28 04 00 08 	cmplwi  r4,8
>>       3c7c:	7c 85 23 78 	mr      r5,r4
>>       3c80:	41 81 00 10 	bgt     3c90 <tst_copy_to_user+0x18>
>>       3c84:	3c 80 00 00 	lis     r4,0
>> 			3c86: R_PPC_ADDR16_HA	jiffies_64
>>       3c88:	38 84 00 00 	addi    r4,r4,0
>> 			3c8a: R_PPC_ADDR16_LO	jiffies_64
>>       3c8c:	48 00 00 00 	b       3c8c <tst_copy_to_user+0x14>
>> 			3c8c: R_PPC_REL24	_copy_to_user
>>
>>       3c90:	94 21 ff f0 	stwu    r1,-16(r1)
>>       3c94:	7c 08 02 a6 	mflr    r0
>>       3c98:	38 60 00 08 	li      r3,8
>>       3c9c:	90 01 00 14 	stw     r0,20(r1)
>>       3ca0:	90 81 00 08 	stw     r4,8(r1)
>>       3ca4:	48 00 00 01 	bl      3ca4 <tst_copy_to_user+0x2c>
>> 			3ca4: R_PPC_REL24	__copy_overflow
>>       3ca8:	80 a1 00 08 	lwz     r5,8(r1)
>>       3cac:	80 01 00 14 	lwz     r0,20(r1)
>>       3cb0:	7c a3 2b 78 	mr      r3,r5
>>       3cb4:	7c 08 03 a6 	mtlr    r0
>>       3cb8:	38 21 00 10 	addi    r1,r1,16
>>       3cbc:	4e 80 00 20 	blr
>>
>>
>> Without the patch:
>>
>> 00003c88 <tst_copy_to_user>:
>>       3c88:	28 04 00 08 	cmplwi  r4,8
>>       3c8c:	7c 85 23 78 	mr      r5,r4
>>       3c90:	41 81 00 10 	bgt     3ca0 <tst_copy_to_user+0x18>
>>       3c94:	3c 80 00 00 	lis     r4,0
>> 			3c96: R_PPC_ADDR16_HA	jiffies_64
>>       3c98:	38 84 00 00 	addi    r4,r4,0
>> 			3c9a: R_PPC_ADDR16_LO	jiffies_64
>>       3c9c:	48 00 00 00 	b       3c9c <tst_copy_to_user+0x14>
>> 			3c9c: R_PPC_REL24	_copy_to_user
>>
>>       3ca0:	94 21 ff f0 	stwu    r1,-16(r1)
>>       3ca4:	3c 60 00 00 	lis     r3,0
>> 			3ca6: R_PPC_ADDR16_HA	.rodata.str1.4+0x30
>>       3ca8:	90 81 00 08 	stw     r4,8(r1)
>>       3cac:	7c 08 02 a6 	mflr    r0
>>       3cb0:	38 63 00 00 	addi    r3,r3,0
>> 			3cb2: R_PPC_ADDR16_LO	.rodata.str1.4+0x30
>>       3cb4:	38 80 00 08 	li      r4,8
>>       3cb8:	90 01 00 14 	stw     r0,20(r1)
>>       3cbc:	48 00 00 01 	bl      3cbc <tst_copy_to_user+0x34>
>> 			3cbc: R_PPC_REL24	__warn_printk
>>       3cc0:	80 a1 00 08 	lwz     r5,8(r1)
>>       3cc4:	0f e0 00 00 	twui    r0,0
>>       3cc8:	80 01 00 14 	lwz     r0,20(r1)
>>       3ccc:	7c a3 2b 78 	mr      r3,r5
>>       3cd0:	7c 08 03 a6 	mtlr    r0
>>       3cd4:	38 21 00 10 	addi    r1,r1,16
>>       3cd8:	4e 80 00 20 	blr
> 
> I make that 3 extra instructions.
> Two are needed to load the format string.
> Not sure why the third gets added.

Third instruction is 'twui', to 'trap' and get the warning oops.

> 
> Not really significant in the 12-15 the error call actually takes.
> Although a lot of those are just generating the stack frame
> in order to call the error function - and wouldn't be there in
> a less trivial example.


Yes, after looking once more, maybe making it __always_inline would be 
enough.

The starting point was that I got almost 50 times copy_overflow() in my 
vmlinux, each having its own format string as well.

So my patch reduced vmlinux size by 3908 bytes.

But with __always_inline I get a reduction by 3560 which is almost the same.

So if you prefer, I can just make copy_overflow() __always_inline and voila.


> 
> More interesting would be changing copy_overflow() to return the size.
> So copy_to_user() becomes:
> 
> 	if (size_valid())
> 		return _copy_to_user();
> 	return copy_overflow()

Yes that's something to try, allthough it means changing all callers of 
check_copy_size()

Christophe