//! Hardware self-test for the ESP-Hosted port table, with no ESP-Hosted C linked in.
//!
//! Run with:  zig build -Dapp=src/portcheck.zig run -Dseconds=8
//!
//! Every line is a claim this file can actually make from the die. What it does *not* do is bring
//! the radio up: that needs the ESP-Hosted C linked beside it, which is the parent's build step.
//! What it proves is that the seam works - that the table's layout is what C measured, that the
//! heap survives the allocation pattern ESP-Hosted subjects it to, that the OS objects behave under
//! the real `std.Io` on this chip rather than under `Threaded` on the host, and that the reset pin
//! moves the way the radio needs.

const std = @import("std");
const hal = @import("hal");

const net = @import("net");
const port = net.port;
const hheap = net.heap;
const os = net.os;
const p4 = @import("io");

extern fn ets_printf(fmt: [*:0]const u8, ...) c_int;

fn print(comptime fmt: [*:0]const u8, args: anytype) void {
    _ = @call(.auto, ets_printf, .{fmt} ++ args);
}

pub const panic = std.debug.FullPanic(struct {
    fn call(msg: []const u8, _: ?usize) noreturn {
        print("MARK PORT_PANIC %s\r\n", .{msg.ptr});
        while (true) {}
    }
}.call);

/// Required in **every** app root that implements `std.Io.VTable` on this target, and it has to be
/// here rather than in the runtime: std reads `std_options` from `@import("root")` only
/// (`/usr/lib/zig/std/std.zig:112`), so the same declaration inside src/io/p4.zig is ignored.
///
/// The reason it is needed at all: defining `fileMemoryMapCreate` forces `Io.File.MemoryMap` to be
/// laid out, its `memory` field is `[]align(std.heap.page_size_min) u8`
/// (`/usr/lib/zig/std/Io/File/MemoryMap.zig:18`), and `page_size_min` has no default for
/// freestanding (`/usr/lib/zig/std/heap.zig:48`). It is the *return type* that does it, so no stub
/// body can avoid it. 4096 is arbitrary and honest: nothing in this image pages, and that one
/// field's alignment is the only thing that reads it. `page_size_max` is not reached.
pub const std_options: std.Options = .{
    .page_size_min = 4096,
};

/// The heap ESP-Hosted allocates from. 48 KiB is a starting point, not a measurement: the honest
/// number comes from `port.stats().peak_reserved` after a run with the C linked in, and the
/// dominant term is the transport queues - `CONFIG_ESP_HOSTED_SDIO_TX_Q_SIZE` and `..._RX_Q_SIZE`
/// are both 20 in the working IDF build, and 40 in-flight buffers at 1536 bytes is 60 KB on its
/// own. Those depths will have to come down for this memory budget; see the report.
var heap_buffer: [48 * 1024]u8 align(hheap.Heap.granule) = undefined;

/// Nine task slots: ESP-Hosted's seven, the port's timer service, and this context.
/// 4 KiB each = 36 KiB. `port.requested_stack_bytes` is 5 KiB, which is FreeRTOS's number for tasks
/// that call `printf`; these do not, and the real number wants a painted-stack watermark.
var runtime_storage: p4.Static(9, 4 * 1024) = .{};

var events_seen: u32 = 0;

fn onEvent(e: port.Event) void {
    events_seen += 1;
    switch (e.base) {
        .wifi => print("MARK PORT_EVENT wifi id=%d\r\n", .{e.id}),
        .named => |n| print("MARK PORT_EVENT %s id=%d\r\n", .{ n, e.id }),
    }
}

export fn zig_main() noreturn {
    hal.intr.init();
    hal.systimer.init();
    print("\r\nMARK PORT_START\r\n", .{});

    // -------------------------------------------------------------- 1. the ABI the C side measured
    // Three numbers, and if any of them is wrong the table is a set of calls to the wrong
    // functions. C's own offsetof, with the force-include in place, gives 284 / 148 / 280.
    print("MARK PORT_ABI sizeof=%u config_gpio=%u event_post=%u fields=%u expect=284,148,280,71\r\n", .{
        @as(u32, @sizeOf(port.HostedOsiFuncs)),
        @as(u32, @offsetOf(port.HostedOsiFuncs, "config_gpio")),
        @as(u32, @offsetOf(port.HostedOsiFuncs, "event_post")),
        @as(u32, std.meta.fields(port.HostedOsiFuncs).len),
    });
    // g_h must point at the table before anything runs; C reads `g_h.funcs->...` directly.
    print("MARK PORT_GH funcs_is_table=%u stubs=%u real=%u\r\n", .{
        @as(u32, @intFromBool(port.g_h.funcs == &port.g_hosted_osi_funcs)),
        @as(u32, port.stubbed.len),
        @as(u32, std.meta.fields(port.HostedOsiFuncs).len - port.stubbed.len),
    });

    // -------------------------------------------------------------- 2. install
    var heap = hheap.Heap.init(&heap_buffer);
    const rt = runtime_storage.init(.{});
    const io = rt.io();
    port.install(io, heap.allocator());
    port.setEventHandler(onEvent);
    print("MARK PORT_INSTALL heap=%u tasks=%u stack=%u\r\n", .{
        @as(u32, heap_buffer.len),
        @as(u32, 9),
        @as(u32, 4 * 1024),
    });

    // -------------------------------------------------------------- 3. the timebase
    // _h_get_time_ms off hal.systimer's 16 MHz. Two reads a known delay apart: the difference is
    // the claim, and it is checked against the counter that produced it.
    const t0 = port.g_h.funcs.get_time_ms();
    hal.systimer.delayMicros(50_000);
    const t1 = port.g_h.funcs.get_time_ms();
    print("MARK PORT_TIME t0=%u t1=%u delta_ms=%u expect~50\r\n", .{
        @as(u32, @intCast(t0)), @as(u32, @intCast(t1)), @as(u32, @intCast(t1 - t0)),
    });

    // -------------------------------------------------------------- 4. memory, through the table
    // The exact pattern an arena cannot serve: allocate, free out of order, reallocate. This is
    // mempool.c's churn, done through the C entry points rather than through Zig.
    const f = port.g_h.funcs;
    var held: [12]?*anyopaque = @splat(null);
    for (&held) |*h| h.* = f.malloc_align(1536, 64);
    var aligned_ok: u32 = 0;
    for (held) |h| {
        if (h) |p| if (@intFromPtr(p) % 64 == 0) {
            aligned_ok += 1;
        };
    }
    const after_alloc = port.stats();
    const order = [_]usize{ 7, 0, 11, 3, 9, 1, 5, 10, 2, 8, 4, 6 };
    for (order) |i| f.free_align(held[i]);
    const after_free = port.stats();
    for (&held) |*h| h.* = f.malloc_align(1536, 64);
    const after_realloc = port.stats();
    for (order) |i| f.free_align(held[i]);

    print("MARK PORT_HEAP aligned=%u/12 live_after_alloc=%u live_after_free=%u peak=%u fail=%u\r\n", .{
        aligned_ok,
        @as(u32, @intCast(after_alloc.blocks_live)),
        @as(u32, @intCast(after_free.blocks_live)),
        @as(u32, @intCast(after_realloc.peak_reserved)),
        @as(u32, @intCast(after_realloc.alloc_failures)),
    });
    // The whole point: the second round must not need more memory than the first.
    print("MARK PORT_HEAP_REUSE round1=%u round2=%u expect_equal\r\n", .{
        @as(u32, @intCast(after_alloc.bytes_reserved)),
        @as(u32, @intCast(after_realloc.bytes_reserved)),
    });
    const s = heap.stats();
    print("MARK PORT_HEAP_FREELIST total=%u free=%u largest=%u blocks=%u expect free==total,blocks==1\r\n", .{
        s.total, s.free, s.largest_free, s.free_blocks,
    });
    heap.check();

    // -------------------------------------------------------------- 5. the C-visible sync objects
    // Created and driven through the table, so the handles and the return codes are the C ones.
    const mtx = f.create_mutex().?;
    const lock_ok = f.lock_mutex(mtx, -1);
    const relock_busy = f.lock_mutex(mtx, 0);
    const unlock_ok = f.unlock_mutex(mtx);
    print("MARK PORT_MUTEX lock=%d try_while_held=%d unlock=%d expect=0,-1,0\r\n", .{ lock_ok, relock_busy, unlock_ok });
    _ = f.destroy_mutex(mtx);

    // A FreeRTOS semaphore arrives with one permit already given; sdio_drv.c:1504 depends on it.
    const sem = f.create_semaphore(4).?;
    const initial_take = f.get_semaphore(sem, 0);
    const empty_take = f.get_semaphore(sem, 0);
    _ = f.post_semaphore(sem);
    const after_post = f.get_semaphore(sem, 0);
    print("MARK PORT_SEM initial=%d empty=%d after_post=%d expect=0,-5,0\r\n", .{ initial_take, empty_take, after_post });
    _ = f.destroy_semaphore(sem);

    // A queue of 24-byte records, which is sizeof(interface_buffer_handle_t) on rv32.
    const q = f.create_queue(4, 24).?;
    var rec: [24]u8 = @splat(0xA5);
    var out: [24]u8 = @splat(0);
    const empty_deq = f.dequeue_item(q, &out, 0);
    var sent: c_int = 0;
    for (0..4) |_| sent += f.queue_item(q, &rec, -1);
    const full_send = f.queue_item(q, &rec, 0);
    const waiting = f.queue_msg_waiting(q);
    const deq = f.dequeue_item(q, &out, -1);
    print("MARK PORT_QUEUE empty=%d sent=%d full=%d waiting=%d deq=%d roundtrip=%u expect=-1,0,-1,4,0,1\r\n", .{
        empty_deq, sent, full_send, waiting, deq, @as(u32, @intFromBool(out[0] == 0xA5 and out[23] == 0xA5)),
    });
    _ = f.destroy_queue(q);

    // -------------------------------------------------------------- 6. timers, through the table
    const timer = f.timer_start("portcheck_oneshot", 30, 0, timerFired, null);
    print("MARK PORT_TIMER_ARMED handle=%u\r\n", .{@as(u32, @intFromBool(timer != null))});
    // The timer service task only runs when this context blocks. Sleeping is what starts it.
    _ = f.msleep(120);
    print("MARK PORT_TIMER fired=%u expect=1\r\n", .{timer_fires});
    // Stopping an expired one-shot reports failure, as esp_timer_stop does.
    if (timer) |t| print("MARK PORT_TIMER_STOP %d expect=-1\r\n", .{f.timer_stop(t)});

    // -------------------------------------------------------------- 7. events
    _ = f.event_post("PORTCHECK_EVENT", 7, null, 0, 0);
    _ = f.event_wifi_post(4, null, 0, 0);
    print("MARK PORT_EVENTS seen=%u expect=2\r\n", .{events_seen});

    // -------------------------------------------------------------- 8. the reset pin
    // GPIO54 has an external pull-up, so released means high. ESP-Hosted's sequence
    // (sdio_drv.c:1651-1657) is active, inactive, active, and with this board's configuration
    // active is HIGH - so it ends released. Driven here through the table's own GPIO entries, with
    // readback, because getting this backwards holds the radio in reset for ever.
    const pin: u32 = port.config.reset_pin;
    _ = f.config_gpio(null, pin, 1 | 2); // H_GPIO_MODE_INPUT_OUTPUT: drive and read back
    _ = f.write_gpio(null, pin, 1);
    const high1 = f.read_gpio(null, pin);
    _ = f.msleep(10);
    _ = f.write_gpio(null, pin, 0);
    const low = f.read_gpio(null, pin);
    _ = f.msleep(10);
    _ = f.write_gpio(null, pin, 1);
    const high2 = f.read_gpio(null, pin);
    print("MARK PORT_RESET pin=%u high=%d low=%d released=%d expect=1,0,1\r\n", .{ pin, high1, low, high2 });

    // The pull entries, on the same pad: enable a pull-down, then disable it, and check the
    // internal pull does not end up fighting the external one.
    _ = f.config_gpio(null, pin, 1); // input only, so the pull is what drives the pad
    _ = f.pull_gpio(null, pin, 0, 1); // H_GPIO_PULL_DOWN, enable
    hal.systimer.delayMicros(200);
    const pulled_down = f.read_gpio(null, pin);
    _ = f.pull_gpio(null, pin, 0, 0); // disable it again
    hal.systimer.delayMicros(200);
    const released = f.read_gpio(null, pin);
    print("MARK PORT_PULL down=%d released=%d expect=0,1\r\n", .{ pulled_down, released });
    // Leave the radio out of reset, whatever the test did to the pad.
    _ = f.config_gpio(null, pin, 1 | 2);
    _ = f.write_gpio(null, pin, 1);

    // -------------------------------------------------------------- 9. what was never implemented
    const final = port.stats();
    print("MARK PORT_STUBS_HIT %u\r\n", .{final.stub_calls});
    print("MARK PORT_DONE live=%u reserved=%u peak=%u blocks=%u fail=%u\r\n", .{
        @as(u32, @intCast(final.bytes_live)),
        @as(u32, @intCast(final.bytes_reserved)),
        @as(u32, @intCast(final.peak_reserved)),
        @as(u32, @intCast(final.blocks_live)),
        @as(u32, @intCast(final.alloc_failures)),
    });
    while (true) {}
}

var timer_fires: u32 = 0;

fn timerFired(_: ?*anyopaque) callconv(.c) void {
    timer_fires += 1;
}

/// Reset entry, verbatim from `src/main.zig:79-95`: enable the F extension, establish a stack in
/// L2MEM, clear .bss, jump to `zig_main`. Every app in this repo carries its own copy because the
/// linker script's entry symbol is per-image.
export fn _start() linksection(".text.entry") callconv(.naked) noreturn {
    asm volatile (
        \\ li t0, 1 << 13
        \\ csrs mstatus, t0
        \\ la sp, __stack_top
        \\ mv fp, sp
        \\ la t0, __bss_start
        \\ la t1, __bss_end
        \\ bgeu t0, t1, 2f
        \\1:
        \\ sw zero, 0(t0)
        \\ addi t0, t0, 4
        \\ bltu t0, t1, 1b
        \\2:
        \\ j zig_main
    );
}