#include <pthread.h>
#include <stddef.h>
#include <string.h>
#include <sys/mman.h>
#include <zircon/process.h>
#include <zircon/syscalls.h>

#include <runtime/thread.h>
#include <runtime/tls.h>

#include "asan_impl.h"
#include "futex_impl.h"
#include "libc.h"
#include "stdio_impl.h"
#include "threads_impl.h"
#include "zircon_impl.h"

__NO_SAFESTACK NO_ASAN static inline pthread_t prestart(void* arg) {
  pthread_t self = arg;
#ifdef __aarch64__
  // Initialize the shadow call stack pointer, which grows up.
  __asm__ volatile("ldr x18, %0" : : "m"(self->shadow_call_stack.iov_base));
#endif
  zxr_tp_set(zxr_thread_get_handle(&self->zxr_thread), pthread_to_tp(self));
  __sanitizer_thread_start_hook(self->sanitizer_hook, (thrd_t)self);
  return self;
}

__NO_RETURN __NO_SAFESTACK NO_ASAN static void start_pthread(void* arg) {
  pthread_t self = prestart(arg);
  __pthread_exit(self->start(self->start_arg));
}

__NO_RETURN __NO_SAFESTACK NO_ASAN static void start_c11(void* arg) {
  pthread_t self = prestart(arg);
  int (*start)(void*) = (int (*)(void*))(uintptr_t)self->start;
  __pthread_exit((void*)(intptr_t)start(self->start_arg));
}

__NO_SAFESTACK static void deallocate_region(const struct iovec* region) {
  _zx_vmar_unmap(_zx_vmar_root_self(), (uintptr_t)region->iov_base, region->iov_len);
}

__NO_SAFESTACK static void deallocate_stack(struct iovec* stack, const struct iovec* region) {
  // Clear the pointers in the TCB before actually unmapping.  In case we get
  // suspended by __sanitizer_memory_snapshot, the TCB is always expected to
  // contain valid pointers.
  stack->iov_base = NULL;
  stack->iov_len = 0;
  atomic_signal_fence(memory_order_seq_cst);
  deallocate_region(region);
}

int __pthread_create(pthread_t* restrict res, const pthread_attr_t* restrict attrp,
                     void* (*entry)(void*), void* restrict arg) {
  pthread_attr_t attr = attrp == NULL ? DEFAULT_PTHREAD_ATTR : *attrp;

  // We do not support providing a stack via pthread attributes.
  if (attr._a_stackaddr != NULL)
    return ENOTSUP;

  char thread_name[ZX_MAX_NAME_LEN];
  thrd_t new = __allocate_thread(
      attr._a_guardsize, attr._a_stacksize,
      attr.__name != NULL ? attr.__name : attr.__c11 ? "thrd_t" : "pthread_t", thread_name);
  if (new == NULL)
    return EAGAIN;

  const char* name = attr.__name != NULL ? attr.__name : thread_name;
  zx_status_t status =
      zxr_thread_create(_zx_process_self(), name, attr._a_detach, &new->zxr_thread);
  if (status != ZX_OK)
    goto fail_after_alloc;

  zxr_thread_entry_t start = attr.__c11 ? start_c11 : start_pthread;

  new->start = entry;
  new->start_arg = arg;

  void* sanitizer_hook = __sanitizer_before_thread_create_hook(
      (thrd_t) new, attr._a_detach, name, new->safe_stack.iov_base, new->safe_stack.iov_len);
  new->sanitizer_hook = sanitizer_hook;

  // We have to publish the pointer now, and make sure it is
  // visible, as in C11 the end of thrd_create synchronizes with the
  // entry point of the new thread.
  *res = new;
  atomic_thread_fence(memory_order_release);

  atomic_fetch_add(&libc.thread_count, 1);

  // This will (hopefully) start the new thread. It could instantly
  // run to completion and deallocate it self. As such, we can't
  // access new->anything after this point.
  status = zxr_thread_start(&new->zxr_thread, (uintptr_t) new->safe_stack.iov_base,
                            new->safe_stack.iov_len, start, new);

  if (status == ZX_OK) {
    __sanitizer_thread_create_hook(sanitizer_hook, (thrd_t) new, thrd_success);
    return 0;
  }

  *res = NULL;
  atomic_fetch_sub(&libc.thread_count, 1);

  __sanitizer_thread_create_hook(sanitizer_hook, (thrd_t) new,
                                 status == ZX_ERR_ACCESS_DENIED ? thrd_error : thrd_nomem);

fail_after_alloc:
  __thread_list_erase(new);
  deallocate_stack(&new->safe_stack, &new->safe_stack_region);
  deallocate_stack(&new->unsafe_stack, &new->unsafe_stack_region);
#if HAVE_SHADOW_CALL_STACK
  deallocate_stack(&new->shadow_call_stack, &new->shadow_call_stack_region);
#endif
  deallocate_region(&new->tcb_region);
  return status == ZX_ERR_ACCESS_DENIED ? EPERM : EAGAIN;
}

static _Noreturn void final_exit(pthread_t self) __asm__("final_exit") __attribute__((used));

static __NO_SAFESTACK NO_ASAN void final_exit(pthread_t self) {
  deallocate_stack(&self->safe_stack, &self->safe_stack_region);
  deallocate_stack(&self->unsafe_stack, &self->unsafe_stack_region);
#if HAVE_SHADOW_CALL_STACK
  deallocate_stack(&self->shadow_call_stack, &self->shadow_call_stack_region);
#endif

  // This deallocates the TCB region too for the detached case.  If not
  // detached, pthread_join will deallocate it.  This always makes the
  // __thread_list_erase callback before deallocating the TCB, so
  // __sanitizer_memory_snapshot should not consider the thread to be "alive"
  // any more safely before the memory might be unmapped.
  zxr_thread_exit_unmap_if_detached(&self->zxr_thread, __thread_list_erase, self,
                                    _zx_vmar_root_self(), (uintptr_t)self->tcb_region.iov_base,
                                    self->tcb_region.iov_len);
}

static NO_ASAN _Noreturn void finish_exit(pthread_t self) {
  __sanitizer_thread_exit_hook(self->sanitizer_hook, (thrd_t)self);

  /* It's impossible to determine whether this is "the last thread"
   * until performing the atomic decrement, since multiple threads
   * could exit at the same time. For the last thread, revert the
   * decrement and unblock signals to give the atexit handlers and
   * stdio cleanup code a consistent state. */
  if (atomic_fetch_sub(&libc.thread_count, 1) == -1) {
    atomic_store(&libc.thread_count, 0);
    exit(0);
  }

  // Switch off the thread's normal stack so it can be freed.  The TCB
  // region stays alive so the pthread_t is still valid for pthread_join.
  // The rest of the region is no longer used for TLS, so it can serve
  // as the small amount of temporary stack needed for the exit calls.

#ifdef __x86_64__
  // The thread descriptor is at the end of the region, so the space
  // before it is available as the temporary stack.
  // The x86-64 ABI requires %rsp % 16 = 8 on entry.
  __asm__(
      "mov %[self], %%rsp\n"
      "and $-16, %%rsp\n"
      "call final_exit\n"
      "# Target receives %[self]"
      :
      : [self] "D"(self));
#elif defined(__aarch64__)
  // The thread descriptor is at the start of the region, so the rest of
  // the space up to the guard page is available as the temporary stack.
  __asm__(
      "add sp, %[base], %[len]\n"
      "mov x0, %[self]\n"
      "bl final_exit"
      :
      : [base] "r"(self->tcb_region.iov_base), [len] "r"(self->tcb_region.iov_len - PAGE_SIZE),
        [self] "r"(self));
#else
#error what architecture?
#endif
  __builtin_unreachable();
}

_Noreturn void __pthread_exit(void* result) {
  pthread_t self = __pthread_self();

  self->result = result;

  __tls_run_dtors();

  __thread_tsd_run_dtors();

  __dl_thread_cleanup();

  // After this point the sanitizer runtime will tear down its state,
  // so we cannot run any more sanitized code.
  finish_exit(self);
}

weak_alias(__pthread_create, pthread_create);
weak_alias(__pthread_exit, pthread_exit);