Re: [PATCH v4 4/8] kallsyms: Improve the performance of kallsyms_lookup_name()

[Leizhen (ThunderTown)]

On 2022/9/22 10:15, Leizhen (ThunderTown) wrote:
> 
> 
> On 2022/9/21 23:25, Petr Mladek wrote:
>> On Tue 2022-09-20 15:13:13, Zhen Lei wrote:
>>> Currently, to search for a symbol, we need to expand the symbols in
>>> 'kallsyms_names' one by one, and then use the expanded string for
>>> comparison. This process can be optimized.
>>>
>>> And now scripts/kallsyms no longer compresses the symbol types, each
>>> symbol type always occupies one byte. So we can first compress the
>>> searched symbol and then make a quick comparison based on the compressed
>>> length and content. In this way, for entries with mismatched lengths,
>>> there is no need to expand and compare strings. And for those matching
>>> lengths, there's no need to expand the symbol. This saves a lot of time.
>>> According to my test results, the average performance of
>>> kallsyms_lookup_name() can be improved by 20 to 30 times.
>>>
>>> The pseudo code of the test case is as follows:
>>> static int stat_find_name(...)
>>> {
>>> 	start = sched_clock();
>>> 	(void)kallsyms_lookup_name(name);
>>> 	end = sched_clock();
>>> 	//Update min, max, cnt, sum
>>> }
>>>
>>> /*
>>>  * Traverse all symbols in sequence and collect statistics on the time
>>>  * taken by kallsyms_lookup_name() to lookup each symbol.
>>>  */
>>> kallsyms_on_each_symbol(stat_find_name, NULL);
>>>
>>> The test results are as follows (twice):
>>> After : min=5250, max=  726560, avg= 302132
>>> After : min=5320, max=  726850, avg= 301978
>>> Before: min=170,  max=15949190, avg=7553906
>>> Before: min=160,  max=15877280, avg=7517784
>>>
>>> The average time consumed is only 4.01% and the maximum time consumed is
>>> only 4.57% of the time consumed before optimization.
>>>
>>> Signed-off-by: Zhen Lei <thunder.leizhen@xxxxxxxxxx>
>>> ---
>>>  kernel/kallsyms.c | 79 +++++++++++++++++++++++++++++++++++++++++++++--
>>>  1 file changed, 76 insertions(+), 3 deletions(-)
>>>
>>> diff --git a/kernel/kallsyms.c b/kernel/kallsyms.c
>>> index 3e7e2c2ad2f75ef..2d76196cfe89f34 100644
>>> --- a/kernel/kallsyms.c
>>> +++ b/kernel/kallsyms.c
>>> @@ -87,6 +87,71 @@ static unsigned int kallsyms_expand_symbol(unsigned int off,
>>>  	return off;
>>>  }
>>>  
>>> +static int kallsyms_name_to_tokens(const char *name, char *buf)
>>
>> This is not safe API. It is always needed to pass the size of the
>> buffer.
> 
> OK
> 
>>
>> Also it should be called "compress". "token" is just an implementation
>> detail.
>>
>> I would do:
>>
>> static int kallsyms_compress_symbol_name(const char *name,
>> 					 char *buf, size_t size)
> 
> This's a wonderful name. Thanks.
> 
>>
>>
>>> +{
>>> +	int i, j, k, n;
>>> +	int len, token_len;
>>> +	const char *token;
>>> +	unsigned char token_idx[KSYM_NAME_LEN];
>>> +	unsigned char token_bak[KSYM_NAME_LEN];
>>
>> Why do we need two buffers? It should be possible to compress the name
>> in the same buffer as it is done in compress_symbols() in scripts/callsyms.c.
> 
> Because the performance would be a little better. Now this function takes
> just over a microsecond. Currently, it takes about 250 microseconds on
> average to lookup a symbol, so adding a little more time to this function
> doesn't affect the overall picture. I'll modify and test it as you suggest
> below.
> 
>>
>>> +
>>> +	/*
>>> +	 * n, number of tokens in the string name.
>>> +	 * token_idx[i], the start index of the ith token.
>>> +	 * token_idx[n] is used to calculate the length of the last token.
>>> +	 */
>>> +	n = strlen(name);
>>> +	if (n >= KSYM_NAME_LEN) {
>>> +		buf[0] = 0;
>>> +		return 0;
>>> +	}
>>> +	for (i = 0; i <= n; i++)
>>> +		token_idx[i] = (unsigned char)i;
>>> +
>>> +	/*
>>> +	 * For tokens whose token_len >= 2, a larger index value indicates
>>> +	 * a higher occurrence frequency. See scripts/kallsyms.c
>>> +	 */
>>> +	for (i = 255; i >= 0; i--) {
>>> +		token = &kallsyms_token_table[kallsyms_token_index[i]];
>>> +		token_len = strlen(token);
>>> +		if (token_len <= 1)
>>> +			continue;
>>> +
>>> +		/*
>>> +		 * Find and merge two tokens into one.
>>> +		 *
>>> +		 *                |<-- new_token -->|
>>> +		 *                | token1 | token2 |
>>> +		 * token_idx[]:   j       j+1      j+2
>>> +		 */
>>> +		for (j = 0; j < n - 1; j++) {
>>> +			len = token_idx[j + 2] - token_idx[j];
>>> +			if (len == token_len &&
>>> +			    !strncmp(name + token_idx[j], token, len)) {
>>> +				token_bak[token_idx[j]] = (unsigned char)i;
>>> +				for (k = j + 1; k < n; k++)
>>> +					token_idx[k] = token_idx[k + 1];
>>> +				n--;
>>> +			}
>>> +		}
>>> +	}
>>> +
>>> +	for (j = 0; j < n; j++) {
>>> +		len = token_idx[j + 1] - token_idx[j];
>>> +		if (len <= 1) {
>>> +			buf[j] = name[token_idx[j]];
>>> +			continue;
>>> +		}
>>> +
>>> +		buf[j] = token_bak[token_idx[j]];
>>
>> Maybe, I do not understand the compression format correctly but
>> this code looks too complicated. Honestly, I even did not try to
>> understand it.
>>
>> My understanding is the we just need to find all tokens and
>> replace them with index.
>>
>> It should be even easier than compress_symbols() in scripts/callsyms.c.
>> The token_table already exists and we do not need to handle the token_profit...
>>
>> The following looks more strigtforward (not even compile tested):
> 
> OK, I will try this one. Or refer to compress_symbols() in scripts/callsyms.c.

This method won't work. Because the tokens in kallsyms_token_table[] have
been expanded. For example: name = "nfs_fs_proc_net_init"
kallsyms_token_table[0xf3] = "s_",  raw token = 73 5f
kallsyms_token_table[0x9f] = "fs_", raw token = 66 f3

After "s_" have been replaced with f3, there is no "fs_" substring in namebuf.
That's why I wrote a new piece of code. Due to I didn't want to add a variable
like kallsyms_token_table[].

Now, I will add kallsyms_best_token_table[], kallsyms_best_token_table_len;
kallsyms_best_token_table[] does not store tokens that contain only one
character. And index=0 is the token with the highest frequency.


> 
>>
>> 	ssize_t len, size;
>>
>> 	len = strscpy(buf, symname, size);
>> 	if (WARN_ON_ONCE(len < 0))
>> 		return -EINVAL;
>>
>> 	/* the tokens with higher index are used first */
>> 	for (idx = 255; idx >= 0; idx--) {
>> 		token =	&kallsyms_token_table[kallsyms_token_index[i]];
>> 		token_len = strlen(token);
>>
>> 		p1 = buf;
>> 		/* length of the remaining symname including the trailing '\0' */
>> 		remaining = len + 1;
>>
>> 		/* find the token in the symbol name */
>> 		p2 = strstr(token, p1);
>>
>> 		while (p2) {
>> 			/* replace token with index */
>> 			*p2 = idx;
>> 			remaining -= ((p2 - p1) + token_len);
>> 			memmove(p2 + 1, p2 + token_len, remaining);
>> 			len -= (token_len - 1);
>> 			p1 = p2;
>>
>> 			/* find the token in the rest of the symbol name */
>> 			p2 = strstr(token, p1);
>> 		}
>> 	}
>>
>> 	return len;
>>
>>> +	}
>>> +	buf[n] = 0;
>>> +
>>> +	return n;
>>> +}
>>> +
>>>  /*
>>>   * Get symbol type information. This is encoded as a single char at the
>>>   * beginning of the symbol name.
>>> @@ -192,20 +257,28 @@ unsigned long kallsyms_lookup_name(const char *name)
>>>  	char namebuf[KSYM_NAME_LEN];
>>>  	unsigned long i;
>>>  	unsigned int off;
>>> +	int len;
>>>  
>>>  	/* Skip the search for empty string. */
>>>  	if (!*name)
>>>  		return 0;
>>>  
>>> +	len = kallsyms_name_to_tokens(name, namebuf);
>>> +	for (i = 0, off = 0; len && i < kallsyms_num_syms; i++) {
>>> +		if (kallsyms_names[off] == len + 1 &&
>>> +		    !memcmp(&kallsyms_names[off + 2], namebuf, len))
>>> +			return kallsyms_sym_address(i);
>>> +
>>> +		off += kallsyms_names[off] + 1;
>>
>> These complicated checks are hard to review. The following looks much
>> more readable to me:
> 
> Yes, it looks well.
> 
>>
>> 	for (i = 0, off = 0; len && i < kallsyms_num_syms; i++) {
>> 		/* the stored symbol name is prefixed by symbol type */
>> 		name_len = kallsyms_names[off] - 1;
>> 		name = &kallsyms_names[off + 2];
>> 		off += name_len + 2;
>>
>> 		if (name_len != len)
>> 			continue;
>>
>> 		if (!memcmp(name, namebuf, len))
>> 			return kallsyms_sym_address(i);
>> 	}
>>
>>
>>> +	}
>>> +
>>>  	for (i = 0, off = 0; i < kallsyms_num_syms; i++) {
>>>  		off = kallsyms_expand_symbol(off, namebuf, ARRAY_SIZE(namebuf));
>>>  
>>> -		if (strcmp(namebuf, name) == 0)
>>> -			return kallsyms_sym_address(i);
>>> -
>>>  		if (cleanup_symbol_name(namebuf) && strcmp(namebuf, name) == 0)
>>>  			return kallsyms_sym_address(i);
>>
>> Hmm, it means that the speedup is not usable when kernel is compiled LLVM?
>> It might actually slow down the search because we would need to use
>> both fast and slow search?
> 
> Theoretically, I don't think so. A string comparison was removed from the
> slow search. "if (name_len != len)" is faster than
> "if (strcmp(namebuf, name) == 0)". Even if they're equal,
> kallsyms_compress_symbol_name() only takes 1-2us, it doesn't affect the
> overall picture. The average lookup time before optimization is
> millisecond-level. To allay your concerns, I can run a test.
> 
> Before: min=170,  max=15949190, avg=7553906
> 
>>
>> Best Regards,
>> Petr
>> .
>>
> 

-- 
Regards,
  Zhen Lei