Re: [PATCH v5 1/1] Created tunable to force small pages on stack allocation.

[Cupertino Miranda <cupertino.miranda@oracle.com>]

[-- Attachment #1: Type: text/plain, Size: 828 bytes --]


Hi Wilco,

I had a test application demonstrating the problem.
I include it in attach.
From my particular interest in the tunable this is the difference:

# GLIBC_TUNABLES=glibc.pthread.stack_hugetlb=0 ./tststackalloc 1
Page size: 4 kB, 2 MB huge pages
Will attempt to align allocations to make stacks eligible for huge pages
pid: 3482023 (/proc/3482023/smaps)
stack_size = 2097152, 0x200000
Creating 128 threads...
RSS: 448 pages (1835008 bytes = 1 MB)
Press enter to exit...

# GLIBC_TUNABLES=glibc.pthread.stack_hugetlb=1 ./tststackalloc 1
Page size: 4 kB, 2 MB huge pages
Will attempt to align allocations to make stacks eligible for huge pages
pid: 3482254 (/proc/3482254/smaps)
stack_size = 2097152, 0x200000
Creating 128 threads...
RSS: 65891 pages (269889536 bytes = 257 MB)
Press enter to exit...

Regards,
Cupertino


[-- Warning: decoded text below may be mangled, UTF-8 assumed --]
[-- Attachment #2: tststackalloc.c --]
[-- Type: text/x-csrc, Size: 4666 bytes --]

// Compile & run:
//    gcc -Wall -g -o tststackalloc tststackalloc.c $< -lpthread
//    ./tststackalloc 1     # Attempt to use huge pages for stacks -> RSS bloat
//    ./tststackalloc 0     # Do not attempt to use huge pages -> No RSS bloat

#define _GNU_SOURCE
#include <pthread.h>
#include <stdio.h>
#include <stdlib.h>
#include <stdint.h>
#include <string.h>
#include <unistd.h>
#include <inttypes.h>
#include <sys/mman.h>
#include <fcntl.h>

// Number of threads to create
#define NOOF_THREADS (128)

// Size of a small page (hard-coded)
#define SMALL_PAGE_SIZE (4*1024)

// Size of a huge page (hard-coded)
#define HUGE_PAGE_SIZE (2*1024*1024)

// Total size of the thread stack, including the guard page(s)
#define STACK_SIZE_TOTAL (HUGE_PAGE_SIZE)

// Size of the guard page(s)
#define GUARD_SIZE (SMALL_PAGE_SIZE)

//#define PRINT_STACK_RANGES
//#define PRINT_PROC_SMAPS

// When enabled (set to non-zero), tries to align thread stacks on
// huge page boundaries, making them eligible for huge pages
static int huge_page_align_stacks;

static volatile int exit_thread = 0;

#if defined(PRINT_STACK_RANGES)
static void print_stack_range(void) {
  pthread_attr_t attr;
  void* bottom;
  size_t size;
  int err;

  err = pthread_getattr_np(pthread_self(), &attr);
  if (err != 0) {
    fprintf(stderr, "Error looking up attr\n");
    exit(1);
  }

  err = pthread_attr_getstack(&attr, &bottom, &size);
  if (err != 0) {
    fprintf(stderr, "Cannot locate current stack attributes!\n");
    exit(1);
  }

  pthread_attr_destroy(&attr);

  fprintf(stderr, "Stack: %p-%p (0x%zx/%zd)\n", bottom, bottom + size, size, size);
}
#endif

static void* start(void* arg) {
#if defined(PRINT_STACK_RANGES)
  print_stack_range();
#endif

  while(!exit_thread) {
    sleep(1);
  }
  exit(2);
}

#if defined(PRINT_PROC_SMAPS)
static void print_proc_file(const char* file) {
  char path[128];
  snprintf(path, sizeof(path), "/proc/self/%s", file);
  int smap = open(path, O_RDONLY);
  char buf[4096];
  int x = 0;
  while ((x = read(smap, buf, sizeof(buf))) > 0) {
    write(1, buf, x);
  }
  close(smap);
}
#endif

static size_t get_rss(void) {
  FILE* stat = fopen("/proc/self/statm", "r");
  long rss;
  fscanf(stat, "%*d %ld", &rss);
  return rss;
}

uintptr_t align_down(uintptr_t value, uintptr_t alignment) {
  return value & ~(alignment - 1);
}

// Do a series of small, single page mmap calls to attempt to set
// everything up so that the next mmap call (glibc allocating the
// stack) returns a 2MB aligned range. The kernel "expands" vmas from
// higher to lower addresses (subsequent calls return ranges starting
// at lower addresses), so this function keeps calling mmap until it a
// huge page aligned address is returned. The next range (the stack)
// will then end on that same address.
static void align_next_on(uintptr_t alignment) {
  uintptr_t p;
  do {
    p = (uintptr_t)mmap(NULL, SMALL_PAGE_SIZE, PROT_NONE, MAP_ANONYMOUS|MAP_PRIVATE|MAP_NORESERVE, -1, 0);
  } while (p != align_down(p, HUGE_PAGE_SIZE));
}

int main(int argc, char* argv[]) {
  pthread_t t[NOOF_THREADS];
  pthread_attr_t attr;
  int i;

  if (argc != 2) {
    printf("Usage: %s <huge page stacks>\n", argv[0]);
    printf("    huge page stacks = 1 - attempt to use huge pages for stacks\n");
    exit(1);
  }
  huge_page_align_stacks = atoi(argv[1]);

  void* dummy = malloc(1024);
  free(dummy);

  fprintf(stderr, "Page size: %d kB, %d MB huge pages\n", SMALL_PAGE_SIZE / 1024, HUGE_PAGE_SIZE / (1024 * 1024));
  if (huge_page_align_stacks) {
    fprintf(stderr, "Will attempt to align allocations to make stacks eligible for huge pages\n");
  }
  pid_t pid = getpid();
  fprintf(stderr, "pid: %d (/proc/%d/smaps)\n", pid, pid);

  size_t guard_size = GUARD_SIZE;
  size_t stack_size = STACK_SIZE_TOTAL;
  pthread_attr_init(&attr);
  pthread_attr_setstacksize(&attr, stack_size);
  pthread_attr_setguardsize(&attr, guard_size);

  fprintf(stderr, "stack_size = %d, 0x%x\n", stack_size, stack_size);
  fprintf(stderr, "Creating %d threads...\n", NOOF_THREADS);
  for (i = 0; i < NOOF_THREADS; i++) {
    if (huge_page_align_stacks) {
      // align (next) allocation on huge page boundary
      align_next_on(HUGE_PAGE_SIZE);
    }
    pthread_create(&t[i], &attr, start, NULL);
  }
  sleep(1);

#if defined(PRINT_PROC_SMAPS)
  print_proc_file("smaps");
#endif

  size_t rss = get_rss();
  fprintf(stderr, "RSS: %zd pages (%zd bytes = %zd MB)\n", rss, rss * SMALL_PAGE_SIZE, rss * SMALL_PAGE_SIZE / 1024 / 1024);

  fprintf(stderr, "Press enter to exit...\n");
  getchar();

  exit_thread = 1;
  for (i = 0; i < NOOF_THREADS; i++) {
    pthread_join(t[i], NULL);
  }
  return 0;
}

[-- Attachment #3: Type: text/plain, Size: 460 bytes --]



Wilco Dijkstra writes:

> Hi,
>
>> The next question is whether the splitting of one huge page causes the whole
>> stack mmap to use small pages too.
>
> Btw if it wasn't obvious, you can trivially check this by creating threads that use
> 3MB of stack. Then the RSS per thread should be either 3MB (only small pages,
> no RSS loss!), 4MB (2 large pages, no RSS loss), 5 MB (3MB small pages, 2MB loss)
> or 6MB (2 large pages, 2MB loss).
>
> Cheers,
> Wilco