On 2022/10/26 1:53, Luis Chamberlain wrote: > On Wed, Oct 19, 2022 at 10:11:58PM +0800, Leizhen (ThunderTown) wrote: >> >> >> On 2022/10/19 20:01, Luis Chamberlain wrote: >>> On Mon, Oct 17, 2022 at 02:49:39PM +0800, 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 is very slow. >>>> >>>> In fact, we can first compress the name being looked up and then use >>>> it for comparison when traversing 'kallsyms_names'. >>>> >>>> This patch series optimizes the performance of function kallsyms_lookup_name(), >>>> and function klp_find_object_symbol() in the livepatch module. Based on the >>>> test results, the performance overhead is reduced to 5%. That is, the >>>> performance of these functions is improved by 20 times. >>> >>> Stupid question, is a hash table in order? >> >> No hash table. >> >> All symbols are arranged in ascending order of address. For example: cat /proc/kallsyms >> >> The addresses of all symbols are stored in kallsyms_addresses[], and names of all symbols >> are stored in kallsyms_names[]. The elements in these two arrays are in a one-to-one >> relationship. For any symbol, it has the same index in both arrays. >> >> Therefore, when we look up a symbolic name based on an address, we use a binary lookup. >> However, when we look up an address based on a symbol name, we can only traverse array >> kallsyms_names[] in sequence. I think the reason why hash is not used is to save memory. > > This answers how we don't use a hash table, the question was *should* we > use one? I'm not the original author, and I can only answer now based on my understanding. Maybe the original author didn't think of the hash method, or he has weighed it out. Hash is a good solution if only performance is required and memory overhead is not considered. Using hash will increase the memory size by up to "4 * kallsyms_num_syms + 4 * ARRAY_SIZE(hashtable)" bytes, kallsyms_num_syms is about 1-2 million. Because I don't know what hash algorithm will be used, the cost of generating the hash value corresponding to the symbol name is unknown now. But I think it's gonna be small. But it definitely needs a simpler algorithm, the tool needs to implement the same hash algorithm. If the hash is not very uniform or ARRAY_SIZE(hashtable) is small, then my current approach still makes sense. So maybe hash can be deferred to the next phase of improvement. > > Luis > . > -- Regards, Zhen Lei
