perf: optimize user-list, xauth, and xdmcp paths

[GermanAizek]

@jpeisach,

Using perf benchmarks from this PR (#477), we were able to carry out useful algorithmic optimization, namely:

User List

• Replace O(N^2) per-entry g_list_insert_sorted() during passwd loading with O(1) prepend and a single O(N log N) g_list_sort() merge sort
• Replace O(N^2) nested list traversals for existing and removed users with O(1) hash tables

Xauthority

• Buffer all serialized records into a GByteArray in memory before writing to disk in a single atomic write() syscall, replacing up to 350 separate syscalls per update
• Replace byte-by-byte loops in read_data() with block memcpy()
• Replace O(N^2) g_list_append() with O(N) prepend + reverse

XDMCP

• Replace byte-by-byte loops in read_data(), read_string(), write_data(), and write_string() with single-check block memcpy() operations

I post visual graphs my benchmarks on NUMA server motherboard 2x Xeon E5-2699v3 (Haswell 2014 year - 36 core 72 threads)

Two commits were compared:

• integrate-benchmarks branch 4bf60ba (w/o optimizations)
• optimize branch 1dd9c81

Attention on lookup_by_name from user list, change math algorithm speedup in 371x times !!!

[jpeisach]

So the thing with the user lists.. I imagine most systems don't more than like, eight users at most? So I'm not sure how much this helps. Unless there are people using LightDM on systems with tons of users (in which this would be good)

I'm definitely interested in Xauthority and XDMCP optimizations.

I haven't decided what balance between space and time complexity we should have..

[GermanAizek]

Unless there are people using LightDM on systems with tons of users

This is easy to calculate as a proportion.

371x - 1000 users lookup
37.1x - 100 users
3.7x - 10 users performance up

It’s actually more logical to use a hashtable.

[jpeisach]

By that calculation.. for a one user system that's a slowdown (0.37)