Ve/alloc count

[vladimir-ea]

add a counting allocator with per thread counters exported to surfer

[vladimir-ea]

changed to using CountingAllocator in this PR - possibly counting allocator should be put behind a feature flag ..

[alexX512]

Do we need that big macro? It feels like for our use case we can use small counter in shared memory with specific file name like in allocator file

const BYTES: usize = std::mem::size_of::<AtomicI64>();

thread_local! {
    static BASE: Cell<*mut AtomicI64> = const { Cell::new(null_mut()) };
}

#[inline]
fn cell() -> &'static AtomicI64 {
    let p = BASE.with(|c| {
        let p = c.get();
        if !p.is_null() {
            return p;             
        }
        let p = map_this_thread();   
        c.set(p);
        p
    });
    unsafe { &*p }
}

#[cold]
fn map_this_thread() -> *mut AtomicI64 {
    let t = std::thread::current();
    let file = match t.name() {
        Some(n) => format!("{n}_allocator"),
        None => format!("{:?}_allocator", t.id()),
    };
    mmap_counters_file(local_share_dir().as_ref(), "silver", &file, BYTES)
        .map(|p| p.cast::<AtomicI64>())
        .expect("alloc counter mmap")
}

pub fn add(n: i64) { cell().fetch_add(n, Relaxed); }
pub fn sub(n: i64) { cell().fetch_sub(n, Relaxed); }
pub fn get() -> i64 { cell().load(Relaxed) }

[alexX512]

Sampled label was removed from timed macro as we never used it in prod and with it it might be hard to analyse results as in flamegraphs we track how many time a function was called