zircon/system/utest/vmo/vmo.cc

// Copyright 2016 The Fuchsia Authors. All rights reserved.
// Use of this source code is governed by a BSD-style license that can be
// found in the LICENSE file.

#include <ctype.h>
#include <inttypes.h>
#include <limits.h>
#include <threads.h>

#include <fbl/function.h>
#include <unittest/unittest.h>
#include <zircon/process.h>
#include <zircon/syscalls.h>

#include "bench.h"

bool vmo_cache_map_test() {
  BEGIN_TEST;

  auto maptest = [](uint32_t policy, const char *type) {
    zx_handle_t vmo;
    const size_t size = 256 * 1024;  // 256K

    EXPECT_EQ(ZX_OK, zx_vmo_create(size, 0, &vmo));

    // set the cache policy
    EXPECT_EQ(ZX_OK, zx_vmo_set_cache_policy(vmo, policy));

    // commit it
    EXPECT_EQ(ZX_OK, zx_vmo_op_range(vmo, ZX_VMO_OP_COMMIT, 0, size, nullptr, 0));

    // map it
    uintptr_t ptr;
    EXPECT_EQ(ZX_OK,
              zx_vmar_map(zx_vmar_root_self(), ZX_VM_PERM_READ | ZX_VM_PERM_WRITE | ZX_VM_MAP_RANGE,
                          0, vmo, 0, size, &ptr));

    volatile uint32_t *buf = (volatile uint32_t *)ptr;

    // write it once, priming the cache
    for (size_t i = 0; i < size / 4; i++)
      buf[i] = 0;

    // write to it
    zx_time_t wt = zx_clock_get_monotonic();
    for (size_t i = 0; i < size / 4; i++)
      buf[i] = 0;
    wt = zx_clock_get_monotonic() - wt;

    // read from it
    zx_time_t rt = zx_clock_get_monotonic();
    for (size_t i = 0; i < size / 4; i++)
      __UNUSED uint32_t hole = buf[i];
    rt = zx_clock_get_monotonic() - rt;

    printf("took %" PRIu64 " nsec to write %s memory\n", wt, type);
    printf("took %" PRIu64 " nsec to read %s memory\n", rt, type);

    EXPECT_EQ(ZX_OK, zx_vmar_unmap(zx_vmar_root_self(), ptr, size));
    EXPECT_EQ(ZX_OK, zx_handle_close(vmo));
  };

  printf("\n");
  maptest(ZX_CACHE_POLICY_CACHED, "cached");
  maptest(ZX_CACHE_POLICY_UNCACHED, "uncached");
  maptest(ZX_CACHE_POLICY_UNCACHED_DEVICE, "uncached device");
  maptest(ZX_CACHE_POLICY_WRITE_COMBINING, "write combining");

  END_TEST;
}

bool vmo_unmap_coherency() {
  BEGIN_TEST;

  // This is an expensive test to try to detect a multi-cpu coherency
  // problem with TLB flushing of unmap operations
  //
  // algorithm: map a relatively large committed VMO.
  // Create a worker thread that simply walks through the VMO writing to
  // each page.
  // In the main thread continually decommit the vmo with a little bit of
  // a gap between decommits to allow the worker thread to bring it all back in.
  // If the worker thread appears stuck by not making it through a loop in
  // a reasonable time, we have failed.

  // allocate a vmo
  const size_t len = 32 * 1024 * 1024;
  zx_handle_t vmo;
  zx_status_t status = zx_vmo_create(len, 0, &vmo);
  EXPECT_EQ(ZX_OK, status, "vm_object_create");

  // do a regular map
  uintptr_t ptr = 0;
  status =
      zx_vmar_map(zx_vmar_root_self(), ZX_VM_PERM_READ | ZX_VM_PERM_WRITE, 0, vmo, 0, len, &ptr);
  EXPECT_EQ(ZX_OK, status, "map");
  EXPECT_NE(0u, ptr, "map address");

  // create a worker thread
  struct worker_args {
    size_t len;
    uintptr_t ptr;
    std::atomic<bool> exit;
    std::atomic<bool> exited;
    std::atomic<size_t> count;
  } args = {};
  args.len = len;
  args.ptr = ptr;

  auto worker = [](void *_args) -> int {
    worker_args *a = (worker_args *)_args;

    unittest_printf("ptr %#" PRIxPTR " len %zu\n", a->ptr, a->len);

    while (!a->exit.load()) {
      // walk through the mapping, writing to every page
      for (size_t off = 0; off < a->len; off += PAGE_SIZE) {
        *(uint32_t *)(a->ptr + off) = 99;
      }

      a->count.fetch_add(1);
    }

    unittest_printf("exiting worker\n");

    a->exited.store(true);

    return 0;
  };

  thrd_t t;
  thrd_create(&t, worker, &args);

  const zx_time_t max_duration = ZX_SEC(30);
  const zx_time_t max_wait = ZX_SEC(1);
  zx_time_t start = zx_clock_get_monotonic();
  for (;;) {
    // wait for it to loop at least once
    zx_time_t t = zx_clock_get_monotonic();
    size_t last_count = args.count.load();
    while (args.count.load() <= last_count) {
      if (zx_clock_get_monotonic() - t > max_wait) {
        UNITTEST_FAIL_TRACEF("looper appears stuck!\n");
        break;
      }
    }

    // decommit the vmo
    status = zx_vmo_op_range(vmo, ZX_VMO_OP_DECOMMIT, 0, len, nullptr, 0);
    EXPECT_EQ(0, status, "vm decommit");

    if (zx_clock_get_monotonic() - start > max_duration)
      break;
  }

  // stop the thread and wait for it to exit
  args.exit.store(true);
  while (args.exited.load() == false)
    ;

  END_TEST;
}

BEGIN_TEST_CASE(vmo_tests)
RUN_TEST_PERFORMANCE(vmo_cache_map_test);
RUN_TEST_LARGE(vmo_unmap_coherency);
END_TEST_CASE(vmo_tests)

int main(int argc, char **argv) {
  bool run_bench = false;
  if (argc > 1) {
    if (!strcmp(argv[1], "bench")) {
      run_bench = true;
    }
  }

  if (!run_bench) {
    bool success = unittest_run_all_tests(argc, argv);
    return success ? 0 : -1;
  } else {
    return vmo_run_benchmark();
  }
}