zircon/kernel/tests/string_tests.cc

// Copyright 2016 The Fuchsia Authors
// Copyright (c) 2008-2014 Travis Geiselbrecht
//
// Use of this source code is governed by a MIT-style
// license that can be found in the LICENSE file or at
// https://opensource.org/licenses/MIT

#include <align.h>
#include <inttypes.h>
#include <platform.h>
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <zircon/types.h>

#include <fbl/auto_call.h>
#include <kernel/thread.h>
#include <vm/physmap.h>
#include <vm/pmm.h>

static uint8_t* src;
static uint8_t* dst;

static uint8_t* src2;
static uint8_t* dst2;

#define BUFFER_SIZE (8 * 1024 * 1024)
#define ITERATIONS \
  (1024 * 1024 * 1024 / BUFFER_SIZE)  // enough iterations to have to copy/set 1GB of memory

#if 1
static inline void* mymemcpy(void* dst, const void* src, size_t len) {
  return memcpy(dst, src, len);
}
static inline void* mymemset(void* dst, int c, size_t len) { return memset(dst, c, len); }
#else
// if we're testing our own memcpy, use this
extern void* mymemcpy(void* dst, const void* src, size_t len);
extern void* mymemset(void* dst, int c, size_t len);
#endif

/* reference implementations of memmove/memcpy */
typedef long word;

#define lsize sizeof(word)
#define lmask (lsize - 1)

static void* c_memmove(void* dest, void const* src, size_t count) {
  char* d = (char*)dest;
  const char* s = (const char*)src;
  size_t len;

  if (count == 0 || dest == src)
    return dest;

  if ((long)d < (long)s) {
    if (((long)d | (long)s) & lmask) {
      // src and/or dest do not align on word boundary
      if ((((long)d ^ (long)s) & lmask) || (count < lsize))
        len = count;  // copy the rest of the buffer with the byte mover
      else
        len = lsize - ((long)d & lmask);  // move the ptrs up to a word boundary

      count -= len;
      for (; len > 0; len--)
        *d++ = *s++;
    }
    for (len = count / lsize; len > 0; len--) {
      *(word*)d = *(word*)s;
      d += lsize;
      s += lsize;
    }
    for (len = count & lmask; len > 0; len--)
      *d++ = *s++;
  } else {
    d += count;
    s += count;
    if (((long)d | (long)s) & lmask) {
      // src and/or dest do not align on word boundary
      if ((((long)d ^ (long)s) & lmask) || (count <= lsize))
        len = count;
      else
        len = ((long)d & lmask);

      count -= len;
      for (; len > 0; len--)
        *--d = *--s;
    }
    for (len = count / lsize; len > 0; len--) {
      d -= lsize;
      s -= lsize;
      *(word*)d = *(word*)s;
    }
    for (len = count & lmask; len > 0; len--)
      *--d = *--s;
  }

  return dest;
}

static void* c_memset(void* s, int c, size_t count) {
  char* xs = (char*)s;
  size_t len = (-(size_t)s) & lmask;
  word cc = c & 0xff;

  if (count > len) {
    count -= len;
    cc |= cc << 8;
    cc |= cc << 16;
    if (sizeof(word) == 8)
      cc |= (uint64_t)cc << 32;  // should be optimized out on 32 bit machines

    // write to non-aligned memory byte-wise
    for (; len > 0; len--)
      *xs++ = (char)c;

    // write to aligned memory dword-wise
    for (len = count / lsize; len > 0; len--) {
      *((word*)xs) = (word)cc;
      xs += lsize;
    }

    count &= lmask;
  }

  // write remaining bytes
  for (; count > 0; count--)
    *xs++ = (char)c;

  return s;
}

static void* null_memcpy(void* dst, const void* src, size_t len) { return dst; }

static zx_duration_t bench_memcpy_routine(void* memcpy_routine(void*, const void*, size_t),
                                          size_t srcalign, size_t dstalign) {
  int i;
  zx_time_t t0;

  t0 = current_time();
  for (i = 0; i < ITERATIONS; i++) {
    memcpy_routine(dst + dstalign, src + srcalign, BUFFER_SIZE);
  }
  return current_time() - t0;
}

static void bench_memcpy(void) {
  zx_duration_t null, c, libc, mine;
  size_t srcalign, dstalign;

  printf("memcpy speed test\n");
  Thread::Current::SleepRelative(ZX_MSEC(200));  // let the debug string clear the serial port

  for (srcalign = 0; srcalign < 64;) {
    for (dstalign = 0; dstalign < 64;) {
      {
        InterruptDisableGuard irqd;

        null = bench_memcpy_routine(&null_memcpy, srcalign, dstalign) / ZX_MSEC(1);
        c = bench_memcpy_routine(&c_memmove, srcalign, dstalign) / ZX_MSEC(1);
        libc = bench_memcpy_routine(&memcpy, srcalign, dstalign) / ZX_MSEC(1);
        mine = bench_memcpy_routine(&mymemcpy, srcalign, dstalign) / ZX_MSEC(1);
      }

      printf("srcalign %zu, dstalign %zu: ", srcalign, dstalign);
      printf("   null memcpy %" PRIi64 " msecs\n", null);
      printf("c %" PRIi64 " msecs, %llu bytes/sec; ", c,
             (uint64_t)BUFFER_SIZE * ITERATIONS * 1000ULL / c);
      printf("libc %" PRIi64 " msecs, %llu bytes/sec; ", libc,
             (uint64_t)BUFFER_SIZE * ITERATIONS * 1000ULL / libc);
      printf("my %" PRIi64 " msecs, %llu bytes/sec; ", mine,
             (uint64_t)BUFFER_SIZE * ITERATIONS * 1000ULL / mine);
      printf("\n");

      if (dstalign < 8)
        dstalign++;
      else
        dstalign <<= 1;
    }
    if (srcalign < 8)
      srcalign++;
    else
      srcalign <<= 1;
  }
}

static void fillbuf(void* ptr, size_t len, uint32_t seed) {
  size_t i;
  uint8_t* cptr = (uint8_t*)ptr;

  for (i = 0; i < len; i++) {
    cptr[i] = (uint8_t)seed;
    seed *= 0x1234567;
  }
}

static void validate_memcpy(void) {
  size_t srcalign, dstalign, size;
  const size_t maxsize = 256;

  printf("testing memcpy for correctness\n");

  /*
   * do the simple tests to make sure that memcpy doesn't color outside
   * the lines for all alignment cases
   */
  for (srcalign = 0; srcalign < 64; srcalign++) {
    printf("srcalign %zu\n", srcalign);
    for (dstalign = 0; dstalign < 64; dstalign++) {
      // printf("\tdstalign %zu\n", dstalign);
      for (size = 0; size < maxsize; size++) {
        // printf("srcalign %zu, dstalign %zu, size %zu\n", srcalign, dstalign, size);

        fillbuf(src, maxsize * 2, 567);
        fillbuf(src2, maxsize * 2, 567);
        fillbuf(dst, maxsize * 2, 123514);
        fillbuf(dst2, maxsize * 2, 123514);

        c_memmove(dst + dstalign, src + srcalign, size);
        memcpy(dst2 + dstalign, src2 + srcalign, size);

        int comp = memcmp(dst, dst2, maxsize * 2);
        if (comp != 0) {
          printf("error! srcalign %zu, dstalign %zu, size %zu\n", srcalign, dstalign, size);
        }
      }
    }
  }
}

static zx_duration_t bench_memset_routine(void* memset_routine(void*, int, size_t), size_t dstalign,
                                          size_t len) {
  int i;
  zx_time_t t0;

  t0 = current_time();
  for (i = 0; i < ITERATIONS; i++) {
    memset_routine(dst + dstalign, 0, len);
  }
  return current_time() - t0;
}

static void bench_memset(void) {
  zx_duration_t c, libc, mine;
  size_t dstalign;

  printf("memset speed test\n");
  Thread::Current::SleepRelative(ZX_MSEC(200));  // let the debug string clear the serial port

  for (dstalign = 0; dstalign < 64; dstalign++) {
    {
      InterruptDisableGuard irqd;

      c = bench_memset_routine(&c_memset, dstalign, BUFFER_SIZE) / ZX_MSEC(1);
      libc = bench_memset_routine(&memset, dstalign, BUFFER_SIZE) / ZX_MSEC(1);
      mine = bench_memset_routine(&mymemset, dstalign, BUFFER_SIZE) / ZX_MSEC(1);
    }

    printf("dstalign %zu: ", dstalign);
    printf("c %" PRIi64 " msecs, %llu bytes/sec; ", c,
           (uint64_t)BUFFER_SIZE * ITERATIONS * 1000ULL / c);
    printf("libc %" PRIi64 " msecs, %llu bytes/sec; ", libc,
           (uint64_t)BUFFER_SIZE * ITERATIONS * 1000ULL / libc);
    printf("my %" PRIi64 " msecs, %llu bytes/sec; ", mine,
           (uint64_t)BUFFER_SIZE * ITERATIONS * 1000ULL / mine);
    printf("\n");
  }
}

static void validate_memset(void) {
  size_t dstalign, size;
  int c;
  const size_t maxsize = 256;

  printf("testing memset for correctness\n");

  for (dstalign = 0; dstalign < 64; dstalign++) {
    printf("align %zu\n", dstalign);
    for (size = 0; size < maxsize; size++) {
      for (c = -1; c < 257; c++) {
        fillbuf(dst, maxsize * 2, 123514);
        fillbuf(dst2, maxsize * 2, 123514);

        c_memset(dst + dstalign, c, size);
        memset(dst2 + dstalign, c, size);

        int comp = memcmp(dst, dst2, maxsize * 2);
        if (comp != 0) {
          printf("error! align %zu, c 0x%hhx, size %zu\n", dstalign, (unsigned char)c, size);
        }
      }
    }
  }
}

#include <lib/console.h>

static int string_tests(int argc, const cmd_args* argv, uint32_t flags) {
  list_node list;
  list_initialize(&list);

  // free the physical pages on exit
  auto free_pages = fbl::MakeAutoCall([&list]() {
    pmm_free(&list);
    src = dst = src2 = dst2 = nullptr;
  });

  // allocate a large run of physically contiguous pages and get the address out
  // of the physmap
  size_t total_size = (BUFFER_SIZE + 256) * 4;
  size_t page_count = ROUNDUP(total_size, PAGE_SIZE) / PAGE_SIZE;
  paddr_t pa;
  if (pmm_alloc_contiguous(page_count, 0, PAGE_SIZE_SHIFT, &pa, &list) != ZX_OK) {
    printf("failed to allocate %zu bytes of contiguous memory for test\n", total_size);
    return -1;
  }

  uint8_t* base = (uint8_t*)paddr_to_physmap(pa);
  src = base;
  dst = base + BUFFER_SIZE + 256;
  src2 = base + (BUFFER_SIZE + 256) * 2;
  dst2 = base + (BUFFER_SIZE + 256) * 3;

  printf("src %p, dst %p\n", src, dst);
  printf("src2 %p, dst2 %p\n", src2, dst2);

  if (argc < 3) {
    printf("not enough arguments:\n");
  usage:
    printf("%s validate <routine>\n", argv[0].str);
    printf("%s bench <routine>\n", argv[0].str);
    return -1;
  }

  if (!strcmp(argv[1].str, "validate")) {
    if (!strcmp(argv[2].str, "memcpy")) {
      validate_memcpy();
    } else if (!strcmp(argv[2].str, "memset")) {
      validate_memset();
    }
  } else if (!strcmp(argv[1].str, "bench")) {
    if (!strcmp(argv[2].str, "memcpy")) {
      bench_memcpy();
    } else if (!strcmp(argv[2].str, "memset")) {
      bench_memset();
    }
  } else {
    goto usage;
  }

  return 0;
}

STATIC_COMMAND_START
STATIC_COMMAND("string", "memcpy tests", &string_tests)
STATIC_COMMAND_END(stringtests)