SubmissionQueue

pub const SubmissionQueue = struct {
    head: *u32,
    tail: *u32,
    mask: u32,
    flags: *u32,
    dropped: *u32,
    array: []u32,
    sqes: []linux.io_uring_sqe,
    mmap: []align(page_size_min) u8,
    mmap_sqes: []align(page_size_min) u8,

    // We use `sqe_head` and `sqe_tail` in the same way as liburing:
    // We increment `sqe_tail` (but not `tail`) for each call to `get_sqe()`.
    // We then set `tail` to `sqe_tail` once, only when these events are actually submitted.
    // This allows us to amortize the cost of the @atomicStore to `tail` across multiple SQEs.
    sqe_head: u32 = 0,
    sqe_tail: u32 = 0,

    pub fn init(fd: linux.fd_t, p: linux.io_uring_params) !SubmissionQueue {
        assert(fd >= 0);
        assert((p.features & linux.IORING_FEAT_SINGLE_MMAP) != 0);
        const size = @max(
            p.sq_off.array + p.sq_entries * @sizeOf(u32),
            p.cq_off.cqes + p.cq_entries * @sizeOf(linux.io_uring_cqe),
        );
        const mmap = try posix.mmap(
            null,
            size,
            .{ .READ = true, .WRITE = true },
            .{ .TYPE = .SHARED, .POPULATE = true },
            fd,
            linux.IORING_OFF_SQ_RING,
        );
        errdefer posix.munmap(mmap);
        assert(mmap.len == size);

        // The motivation for the `sqes` and `array` indirection is to make it possible for the
        // application to preallocate static linux.io_uring_sqe entries and then replay them when needed.
        const size_sqes = p.sq_entries * @sizeOf(linux.io_uring_sqe);
        const mmap_sqes = try posix.mmap(
            null,
            size_sqes,
            .{ .READ = true, .WRITE = true },
            .{ .TYPE = .SHARED, .POPULATE = true },
            fd,
            linux.IORING_OFF_SQES,
        );
        errdefer posix.munmap(mmap_sqes);
        assert(mmap_sqes.len == size_sqes);

        const array: [*]u32 = @ptrCast(@alignCast(&mmap[p.sq_off.array]));
        const sqes: [*]linux.io_uring_sqe = @ptrCast(@alignCast(&mmap_sqes[0]));
        // We expect the kernel copies p.sq_entries to the u32 pointed to by p.sq_off.ring_entries,
        // see https://github.com/torvalds/linux/blob/v5.8/fs/io_uring.c#L7843-L7844.
        assert(p.sq_entries == @as(*u32, @ptrCast(@alignCast(&mmap[p.sq_off.ring_entries]))).*);
        return SubmissionQueue{
            .head = @ptrCast(@alignCast(&mmap[p.sq_off.head])),
            .tail = @ptrCast(@alignCast(&mmap[p.sq_off.tail])),
            .mask = @as(*u32, @ptrCast(@alignCast(&mmap[p.sq_off.ring_mask]))).*,
            .flags = @ptrCast(@alignCast(&mmap[p.sq_off.flags])),
            .dropped = @ptrCast(@alignCast(&mmap[p.sq_off.dropped])),
            .array = array[0..p.sq_entries],
            .sqes = sqes[0..p.sq_entries],
            .mmap = mmap,
            .mmap_sqes = mmap_sqes,
        };
    }

    pub fn deinit(self: *SubmissionQueue) void {
        posix.munmap(self.mmap_sqes);
        posix.munmap(self.mmap);
    }
}