SchOS/kernel/acpi/host.c

// SPDX-License-Identifier: GPL-2.0-or-later
//
// What uACPI asks of the kernel it runs in: memory, mapping, port and PCI
// configuration I/O, time, locks, the SCI, and somewhere to run deferred
// work. Every function here is one uacpi_kernel_* entry point from
// <uacpi/kernel_api.h>; the interpreter calls nothing else.
//
// Two things make this harder than it looks.
//
// The namespace is loaded from kmain(), before the scheduler, with no
// process to sleep in -- and AML sleeps, takes mutexes and waits on events
// while it initializes devices. So every blocking primitive here checks
// myproc() and spins when there is nothing to put to sleep, the same rule
// the AHCI and NVMe drivers follow.
//
// And the SCI handler runs AML's notifications and GPE methods through
// uacpi_kernel_schedule_work(), from interrupt context, where AML must not
// run. That is the "acpi" kernel thread's whole job.

#include "core/types.h"
#include "core/defs.h"
#include "core/param.h"
#include "arch/memlayout.h"
#include "arch/mmu.h"
#include "arch/x86.h"
#include "core/spinlock.h"
#include "core/proc.h"
#include "core/irq.h"
#include "core/dma.h"
#include "core/kmem.h"
#include "dev/acpi.h"

#include <uacpi/kernel_api.h>

// --- memory ------------------------------------------------------------------
//
// kmalloc stops at 2 KB and kalloc is one page, and uACPI's allocations range
// from a few bytes to buffers AML sizes however it likes. So each allocation
// carries a header saying where it came from: the slab caches, a page, or --
// for the rare one bigger than a page -- the contiguous DMA arena, which is
// the only multi-page allocator the kernel has.

#define HDR 16
enum { FROM_SLAB = 1, FROM_PAGE, FROM_DMA };

void *
uacpi_kernel_alloc(uacpi_size size)
{
  uint64 n = size + HDR;
  uint64 *h;
  int kind;

  if (n <= 2048) {
    h = kmalloc(n);
    kind = FROM_SLAB;
  } else if (n <= PGSIZE) {
    h = (uint64 *)kalloc();
    kind = FROM_PAGE;
  } else {
    h = dma_alloc(n, 16);
    kind = FROM_DMA;
  }
  if (h == 0)
    return 0;
  h[0] = n;
  h[1] = kind;
  return (uchar *)h + HDR;
}

void
uacpi_kernel_free(void *mem)
{
  uint64 *h;

  if (mem == 0)
    return;
  h = (uint64 *)((uchar *)mem - HDR);
  switch (h[1]) {
  case FROM_SLAB: kmfree(h); break;
  case FROM_PAGE: kfree((char *)h); break;
  case FROM_DMA:  dma_free(h, h[0]); break;
  default:        panic("uacpi_kernel_free: not ours");
  }
}

// Physical memory: tables, and AML's SystemMemory regions. The direct map
// covers the first 4 GB; above that, ioremap().
void *
uacpi_kernel_map(uacpi_phys_addr addr, uacpi_size len)
{
  if (addr + len <= 4ULL * 1024 * 1024 * 1024)
    return P2V(addr);
  void *p = ioremap(addr, len);
  return p ? p : UACPI_MAP_FAILED;
}

void
uacpi_kernel_unmap(void *addr, uacpi_size len)
{
  // Nothing to undo: the direct map is permanent, and ioremap's 1 GB windows
  // are shared and never torn down.
  (void)addr;
  (void)len;
}

uacpi_status
uacpi_kernel_get_rsdp(uacpi_phys_addr *out)
{
  uint64 pa = acpi_rsdp_phys();
  if (pa == 0)
    return UACPI_STATUS_NOT_FOUND;
  *out = pa;
  return UACPI_STATUS_OK;
}

// --- logging -----------------------------------------------------------------

void
uacpi_kernel_log(uacpi_log_level level, const uacpi_char *msg)
{
  switch (level) {
  case UACPI_LOG_ERROR:
  case UACPI_LOG_WARN:
    cprintf("acpi: %s", msg);
    break;
  case UACPI_LOG_INFO:
    kinfo("acpi: %s", msg);
    break;
  default:
    break;                     // trace and debug: per-opcode noise
  }
}

uacpi_status
uacpi_kernel_handle_firmware_request(uacpi_firmware_request *req)
{
  if (req->type == UACPI_FIRMWARE_REQUEST_TYPE_FATAL)
    cprintf("acpi: firmware reports a fatal error (type %x code %x arg %p)\n",
            req->fatal.type, req->fatal.code, req->fatal.arg);
  return UACPI_STATUS_OK;
}

// --- port I/O ----------------------------------------------------------------
//
// The handle is the base port itself; there is nothing to map.

uacpi_status
uacpi_kernel_io_map(uacpi_io_addr base, uacpi_size len, uacpi_handle *out)
{
  (void)len;
  *out = (uacpi_handle)(addr_t)base;
  return UACPI_STATUS_OK;
}

void uacpi_kernel_io_unmap(uacpi_handle h) { (void)h; }

#define PORT(h, off) ((ushort)((addr_t)(h) + (off)))

uacpi_status uacpi_kernel_io_read8(uacpi_handle h, uacpi_size o, uacpi_u8 *v)   { *v = inb(PORT(h, o)); return UACPI_STATUS_OK; }
uacpi_status uacpi_kernel_io_read16(uacpi_handle h, uacpi_size o, uacpi_u16 *v) { *v = inw(PORT(h, o)); return UACPI_STATUS_OK; }
uacpi_status uacpi_kernel_io_read32(uacpi_handle h, uacpi_size o, uacpi_u32 *v) { *v = inl(PORT(h, o)); return UACPI_STATUS_OK; }
uacpi_status uacpi_kernel_io_write8(uacpi_handle h, uacpi_size o, uacpi_u8 v)   { outb(PORT(h, o), v); return UACPI_STATUS_OK; }
uacpi_status uacpi_kernel_io_write16(uacpi_handle h, uacpi_size o, uacpi_u16 v) { outw(PORT(h, o), v); return UACPI_STATUS_OK; }
uacpi_status uacpi_kernel_io_write32(uacpi_handle h, uacpi_size o, uacpi_u32 v) { outl(PORT(h, o), v); return UACPI_STATUS_OK; }

// --- PCI configuration space -------------------------------------------------
//
// Through the 0xCF8 mechanism, which reaches segment 0 and the first 256
// bytes of each function. AML that wants PCIe extended space is refused
// rather than handed bytes from somewhere else.

struct pcihandle {
  uint8 bus, dev, func;
};

uacpi_status
uacpi_kernel_pci_device_open(uacpi_pci_address a, uacpi_handle *out)
{
  struct pcihandle *p;

  if (a.segment != 0)
    return UACPI_STATUS_UNIMPLEMENTED;
  p = uacpi_kernel_alloc(sizeof(*p));
  if (p == 0)
    return UACPI_STATUS_OUT_OF_MEMORY;
  p->bus = a.bus;
  p->dev = a.device;
  p->func = a.function;
  *out = p;
  return UACPI_STATUS_OK;
}

void
uacpi_kernel_pci_device_close(uacpi_handle h)
{
  uacpi_kernel_free(h);
}

#define PCIH ((struct pcihandle *)h)
#define CHECK(o, w) if ((o) + (w) > 256) return UACPI_STATUS_UNIMPLEMENTED

uacpi_status
uacpi_kernel_pci_read8(uacpi_handle h, uacpi_size o, uacpi_u8 *v)
{
  CHECK(o, 1);
  *v = pci_read8(PCIH->bus, PCIH->dev, PCIH->func, o);
  return UACPI_STATUS_OK;
}

uacpi_status
uacpi_kernel_pci_read16(uacpi_handle h, uacpi_size o, uacpi_u16 *v)
{
  CHECK(o, 2);
  *v = pci_read16(PCIH->bus, PCIH->dev, PCIH->func, o);
  return UACPI_STATUS_OK;
}

uacpi_status
uacpi_kernel_pci_read32(uacpi_handle h, uacpi_size o, uacpi_u32 *v)
{
  CHECK(o, 4);
  *v = pci_read32(PCIH->bus, PCIH->dev, PCIH->func, o);
  return UACPI_STATUS_OK;
}

// Narrow writes as read-modify-write of the dword: the kernel has only a
// 32-bit config write, which is all 0xCF8 strictly offers anyway.
static void
pci_write_part(uacpi_handle h, uacpi_size o, uint32 v, int width)
{
  uint32 shift = (o & 3) * 8;
  uint32 mask = (width == 1 ? 0xFFu : 0xFFFFu) << shift;
  uint32 d = pci_read32(PCIH->bus, PCIH->dev, PCIH->func, o & ~3);
  d = (d & ~mask) | ((v << shift) & mask);
  pci_write32(PCIH->bus, PCIH->dev, PCIH->func, o & ~3, d);
}

uacpi_status
uacpi_kernel_pci_write8(uacpi_handle h, uacpi_size o, uacpi_u8 v)
{
  CHECK(o, 1);
  pci_write_part(h, o, v, 1);
  return UACPI_STATUS_OK;
}

uacpi_status
uacpi_kernel_pci_write16(uacpi_handle h, uacpi_size o, uacpi_u16 v)
{
  CHECK(o, 2);
  pci_write_part(h, o, v, 2);
  return UACPI_STATUS_OK;
}

uacpi_status
uacpi_kernel_pci_write32(uacpi_handle h, uacpi_size o, uacpi_u32 v)
{
  CHECK(o, 4);
  pci_write32(PCIH->bus, PCIH->dev, PCIH->func, o, v);
  return UACPI_STATUS_OK;
}

// --- time --------------------------------------------------------------------

uacpi_u64
uacpi_kernel_get_nanoseconds_since_boot(void)
{
  uint64 t = rdtsc(), hz = tsc_hz ? tsc_hz : 1000000000ULL;
  // Split so that the multiplication cannot overflow after a few hours.
  return (t / hz) * 1000000000ULL + (t % hz) * 1000000000ULL / hz;
}

void
uacpi_kernel_stall(uacpi_u8 usec)
{
  microdelay(usec);
}

// Wait for up to `ms`, or until *cond is nonzero when cond is given, sleeping
// a tick at a time when there is a process to sleep and spinning otherwise.
// Caller holds lk, which is released while waiting. Returns 0 if cond became
// true, -1 on timeout. 0xFFFF is "for ever", as uACPI spells it.
static int
wait_until(volatile int *cond, struct spinlock *lk, uint ms)
{
  uint64 start = rdtsc();
  uint64 limit = tsc_hz / 1000 * ms;
  uint t0 = ticks;

  while (!cond || !*cond) {
    if (ms != 0xFFFF) {
      if (myproc() ? (ticks - t0) * MSEC_PER_TICK >= ms
                   : rdtsc() - start >= limit)
        return cond ? -1 : 0;
    }
    if (myproc()) {
      sleep(&ticks, lk);
    } else {
      release(lk);
      microdelay(10);
      acquire(lk);
    }
  }
  return 0;
}

void
uacpi_kernel_sleep(uacpi_u64 msec)
{
  static struct spinlock lk;
  static int init;

  if (!init) {
    initlock(&lk, "acpisleep");
    init = 1;
  }
  if (msec >= 0xFFFF)
    msec = 0xFFFE;
  acquire(&lk);
  wait_until(0, &lk, msec);
  release(&lk);
}

// --- locks and events --------------------------------------------------------

struct kmutex {
  struct spinlock lk;
  volatile int free;          // 1 when nobody holds it
};

uacpi_handle
uacpi_kernel_create_mutex(void)
{
  struct kmutex *m = uacpi_kernel_alloc(sizeof(*m));
  if (m) {
    initlock(&m->lk, "acpimutex");
    m->free = 1;
  }
  return m;
}

void uacpi_kernel_free_mutex(uacpi_handle h) { uacpi_kernel_free(h); }

uacpi_status
uacpi_kernel_acquire_mutex(uacpi_handle h, uacpi_u16 timeout)
{
  struct kmutex *m = h;
  int rc = 0;

  acquire(&m->lk);
  if (!m->free)
    rc = timeout == 0 ? -1 : wait_until(&m->free, &m->lk, timeout);
  if (rc == 0)
    m->free = 0;
  release(&m->lk);
  return rc == 0 ? UACPI_STATUS_OK : UACPI_STATUS_TIMEOUT;
}

void
uacpi_kernel_release_mutex(uacpi_handle h)
{
  struct kmutex *m = h;
  acquire(&m->lk);
  m->free = 1;
  release(&m->lk);
}

struct kevent {
  struct spinlock lk;
  volatile int count;
};

uacpi_handle
uacpi_kernel_create_event(void)
{
  struct kevent *e = uacpi_kernel_alloc(sizeof(*e));
  if (e) {
    initlock(&e->lk, "acpievent");
    e->count = 0;
  }
  return e;
}

void uacpi_kernel_free_event(uacpi_handle h) { uacpi_kernel_free(h); }

uacpi_bool
uacpi_kernel_wait_for_event(uacpi_handle h, uacpi_u16 timeout)
{
  struct kevent *e = h;
  int rc = 0;

  acquire(&e->lk);
  if (e->count == 0)
    rc = timeout == 0 ? -1 : wait_until(&e->count, &e->lk, timeout);
  if (rc == 0)
    e->count--;
  release(&e->lk);
  return rc == 0;
}

void
uacpi_kernel_signal_event(uacpi_handle h)
{
  struct kevent *e = h;
  acquire(&e->lk);
  e->count++;
  release(&e->lk);
}

void
uacpi_kernel_reset_event(uacpi_handle h)
{
  struct kevent *e = h;
  acquire(&e->lk);
  e->count = 0;
  release(&e->lk);
}

uacpi_thread_id
uacpi_kernel_get_thread_id(void)
{
  // A process, or -- before there are any -- the CPU, numbered from 1 so that
  // it can never look like a null pointer.
  struct proc *p = myproc();
  return p ? (uacpi_thread_id)p : (uacpi_thread_id)(addr_t)(cpunum() + 1);
}

uacpi_interrupt_state
uacpi_kernel_disable_interrupts(void)
{
  uacpi_interrupt_state s = readeflags() & FL_IF;
  cli();
  return s;
}

void
uacpi_kernel_restore_interrupts(uacpi_interrupt_state s)
{
  if (s)
    sti();
}

uacpi_handle
uacpi_kernel_create_spinlock(void)
{
  struct spinlock *l = uacpi_kernel_alloc(sizeof(*l));
  if (l)
    initlock(l, "acpispin");
  return l;
}

void uacpi_kernel_free_spinlock(uacpi_handle h) { uacpi_kernel_free(h); }

uacpi_cpu_flags
uacpi_kernel_lock_spinlock(uacpi_handle h)
{
  acquire(h);                  // which disables interrupts, as asked
  return 0;
}

void
uacpi_kernel_unlock_spinlock(uacpi_handle h, uacpi_cpu_flags f)
{
  (void)f;
  release(h);
}

// --- the SCI -----------------------------------------------------------------

struct sci {
  uacpi_interrupt_handler fn;
  uacpi_handle ctx;
  int irq;
};

static void
sci_trampoline(void *arg)
{
  struct sci *s = arg;
  s->fn(s->ctx);
}

uacpi_status
uacpi_kernel_install_interrupt_handler(uacpi_u32 irq,
                                       uacpi_interrupt_handler fn,
                                       uacpi_handle ctx, uacpi_handle *out)
{
  struct sci *s;
  int mode, gsi;

  // The FADT names the SCI as an ISA IRQ. ACPI says it is level/low unless
  // an override says otherwise -- and on most machines one does, making it
  // level/high, which is the case that went wrong when only ISA's and PCI's
  // signalling could be expressed.
  gsi = acpi_isa_gsi(irq, IRQ_LEVEL_LOW, &mode);
  if (gsi < 0 || gsi >= NIRQ_IOAPIC)
    return UACPI_STATUS_UNIMPLEMENTED;

  s = uacpi_kernel_alloc(sizeof(*s));
  if (s == 0)
    return UACPI_STATUS_OUT_OF_MEMORY;
  s->fn = fn;
  s->ctx = ctx;
  s->irq = gsi;
  if (irq_register(gsi, sci_trampoline, s, "acpi", mode | IRQ_QUIET) < 0) {
    uacpi_kernel_free(s);
    return UACPI_STATUS_INTERNAL_ERROR;
  }
  kinfo("acpi: SCI on irq %d (gsi %d, %s)\n", irq, gsi,
        IRQ_IS_LEVEL(mode) ? "level" : "edge");
  *out = s;
  return UACPI_STATUS_OK;
}

uacpi_status
uacpi_kernel_uninstall_interrupt_handler(uacpi_interrupt_handler fn,
                                         uacpi_handle h)
{
  struct sci *s = h;
  (void)fn;
  irq_unregister(s->irq, sci_trampoline, s);
  uacpi_kernel_free(s);
  return UACPI_STATUS_OK;
}

// --- deferred work -----------------------------------------------------------
//
// A ring of pending calls and one thread to make them. Work queued before the
// thread exists waits for it; nothing is queued before interrupts are on.

#define NWORK 64

static struct {
  struct spinlock lk;
  int init;
  struct { uacpi_work_handler fn; uacpi_handle ctx; } q[NWORK];
  uint r, w;
  int busy;
} work;

static void
work_init(void)
{
  if (!work.init) {
    initlock(&work.lk, "acpiwork");
    work.init = 1;
  }
}

uacpi_status
uacpi_kernel_schedule_work(uacpi_work_type type, uacpi_work_handler fn,
                           uacpi_handle ctx)
{
  (void)type;                  // GPEs and notifications alike: one CPU's worth
  work_init();
  acquire(&work.lk);
  if (work.w - work.r >= NWORK) {
    release(&work.lk);
    cprintf("acpi: work queue full; an event was dropped\n");
    return UACPI_STATUS_OUT_OF_MEMORY;
  }
  work.q[work.w % NWORK].fn = fn;
  work.q[work.w % NWORK].ctx = ctx;
  work.w++;
  release(&work.lk);
  wakeup(&work);
  return UACPI_STATUS_OK;
}

uacpi_status
uacpi_kernel_wait_for_work_completion(void)
{
  work_init();
  acquire(&work.lk);
  while (work.r != work.w || work.busy) {
    if (myproc()) {
      sleep(&ticks, &work.lk);
    } else {
      // No thread yet, so nothing would ever drain it: run it here.
      uacpi_work_handler fn = work.q[work.r % NWORK].fn;
      uacpi_handle ctx = work.q[work.r % NWORK].ctx;
      work.r++;
      release(&work.lk);
      fn(ctx);
      acquire(&work.lk);
    }
  }
  release(&work.lk);
  return UACPI_STATUS_OK;
}

static void
acpi_worker(void)
{
  work_init();
  for (;;) {
    acquire(&work.lk);
    while (work.r == work.w)
      sleep(&work, &work.lk);
    uacpi_work_handler fn = work.q[work.r % NWORK].fn;
    uacpi_handle ctx = work.q[work.r % NWORK].ctx;
    work.r++;
    work.busy = 1;
    release(&work.lk);

    fn(ctx);

    acquire(&work.lk);
    work.busy = 0;
    release(&work.lk);
  }
}

void
acpi_thread_start(void)
{
  if (acpi_rsdp_phys() == 0)
    return;
  if (kthread_create(acpi_worker, "acpi") < 0)
    cprintf("acpi: no worker thread; the power button will not work\n");
}