selftests/vm/transhuge-stress: stress test for memory compaction
authorKonstantin Khlebnikov <koct9i@gmail.com>
Thu, 9 Oct 2014 22:29:34 +0000 (15:29 -0700)
committerLinus Torvalds <torvalds@linux-foundation.org>
Fri, 10 Oct 2014 02:26:01 +0000 (22:26 -0400)
This tool induces memory fragmentation via sequential allocation of
transparent huge pages and splitting off everything except their last
sub-pages.  It easily generates pressure to the memory compaction code.

$ perf stat -e 'compaction:*' -e 'migrate:*' ./transhuge-stress
transhuge-stress: allocate 7858 transhuge pages, using 15716 MiB virtual memory and 61 MiB of ram
transhuge-stress: 1.653 s/loop, 0.210 ms/page,   9504.828 MiB/s 7858 succeed,    0 failed, 2439 different pages
transhuge-stress: 1.537 s/loop, 0.196 ms/page,  10226.227 MiB/s 7858 succeed,    0 failed, 2364 different pages
transhuge-stress: 1.658 s/loop, 0.211 ms/page,   9479.215 MiB/s 7858 succeed,    0 failed, 2179 different pages
transhuge-stress: 1.617 s/loop, 0.206 ms/page,   9716.992 MiB/s 7858 succeed,    0 failed, 2421 different pages
^C./transhuge-stress: Interrupt

 Performance counter stats for './transhuge-stress':

         1.744.051      compaction:mm_compaction_isolate_migratepages
             1.014      compaction:mm_compaction_isolate_freepages
         1.744.051      compaction:mm_compaction_migratepages
             1.647      compaction:mm_compaction_begin
             1.647      compaction:mm_compaction_end
         1.744.051      migrate:mm_migrate_pages
                 0      migrate:mm_numa_migrate_ratelimit

       7,964696835 seconds time elapsed

Signed-off-by: Konstantin Khlebnikov <koct9i@gmail.com>
Cc: Rafael Aquini <aquini@redhat.com>
Cc: Andrey Ryabinin <ryabinin.a.a@gmail.com>
Cc: Shuah Khan <shuahkh@osg.samsung.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
tools/testing/selftests/vm/Makefile
tools/testing/selftests/vm/transhuge-stress.c [new file with mode: 0644]

index 3f94e1afd6cfdf88b906d9806b7873633f5523cb..4c4b1f631ecf61f6e3048d746c23be39ea4f2ef9 100644 (file)
@@ -3,6 +3,7 @@
 CC = $(CROSS_COMPILE)gcc
 CFLAGS = -Wall
 BINARIES = hugepage-mmap hugepage-shm map_hugetlb thuge-gen hugetlbfstest
+BINARIES += transhuge-stress
 
 all: $(BINARIES)
 %: %.c
diff --git a/tools/testing/selftests/vm/transhuge-stress.c b/tools/testing/selftests/vm/transhuge-stress.c
new file mode 100644 (file)
index 0000000..fd7f1b4
--- /dev/null
@@ -0,0 +1,144 @@
+/*
+ * Stress test for transparent huge pages, memory compaction and migration.
+ *
+ * Authors: Konstantin Khlebnikov <koct9i@gmail.com>
+ *
+ * This is free and unencumbered software released into the public domain.
+ */
+
+#include <stdlib.h>
+#include <stdio.h>
+#include <stdint.h>
+#include <err.h>
+#include <time.h>
+#include <unistd.h>
+#include <fcntl.h>
+#include <string.h>
+#include <sys/mman.h>
+
+#define PAGE_SHIFT 12
+#define HPAGE_SHIFT 21
+
+#define PAGE_SIZE (1 << PAGE_SHIFT)
+#define HPAGE_SIZE (1 << HPAGE_SHIFT)
+
+#define PAGEMAP_PRESENT(ent)   (((ent) & (1ull << 63)) != 0)
+#define PAGEMAP_PFN(ent)       ((ent) & ((1ull << 55) - 1))
+
+int pagemap_fd;
+
+int64_t allocate_transhuge(void *ptr)
+{
+       uint64_t ent[2];
+
+       /* drop pmd */
+       if (mmap(ptr, HPAGE_SIZE, PROT_READ | PROT_WRITE,
+                               MAP_FIXED | MAP_ANONYMOUS |
+                               MAP_NORESERVE | MAP_PRIVATE, -1, 0) != ptr)
+               errx(2, "mmap transhuge");
+
+       if (madvise(ptr, HPAGE_SIZE, MADV_HUGEPAGE))
+               err(2, "MADV_HUGEPAGE");
+
+       /* allocate transparent huge page */
+       *(volatile void **)ptr = ptr;
+
+       if (pread(pagemap_fd, ent, sizeof(ent),
+                       (uintptr_t)ptr >> (PAGE_SHIFT - 3)) != sizeof(ent))
+               err(2, "read pagemap");
+
+       if (PAGEMAP_PRESENT(ent[0]) && PAGEMAP_PRESENT(ent[1]) &&
+           PAGEMAP_PFN(ent[0]) + 1 == PAGEMAP_PFN(ent[1]) &&
+           !(PAGEMAP_PFN(ent[0]) & ((1 << (HPAGE_SHIFT - PAGE_SHIFT)) - 1)))
+               return PAGEMAP_PFN(ent[0]);
+
+       return -1;
+}
+
+int main(int argc, char **argv)
+{
+       size_t ram, len;
+       void *ptr, *p;
+       struct timespec a, b;
+       double s;
+       uint8_t *map;
+       size_t map_len;
+
+       ram = sysconf(_SC_PHYS_PAGES);
+       if (ram > SIZE_MAX / sysconf(_SC_PAGESIZE) / 4)
+               ram = SIZE_MAX / 4;
+       else
+               ram *= sysconf(_SC_PAGESIZE);
+
+       if (argc == 1)
+               len = ram;
+       else if (!strcmp(argv[1], "-h"))
+               errx(1, "usage: %s [size in MiB]", argv[0]);
+       else
+               len = atoll(argv[1]) << 20;
+
+       warnx("allocate %zd transhuge pages, using %zd MiB virtual memory"
+             " and %zd MiB of ram", len >> HPAGE_SHIFT, len >> 20,
+             len >> (20 + HPAGE_SHIFT - PAGE_SHIFT - 1));
+
+       pagemap_fd = open("/proc/self/pagemap", O_RDONLY);
+       if (pagemap_fd < 0)
+               err(2, "open pagemap");
+
+       len -= len % HPAGE_SIZE;
+       ptr = mmap(NULL, len + HPAGE_SIZE, PROT_READ | PROT_WRITE,
+                       MAP_ANONYMOUS | MAP_NORESERVE | MAP_PRIVATE, -1, 0);
+       if (ptr == MAP_FAILED)
+               err(2, "initial mmap");
+       ptr += HPAGE_SIZE - (uintptr_t)ptr % HPAGE_SIZE;
+
+       if (madvise(ptr, len, MADV_HUGEPAGE))
+               err(2, "MADV_HUGEPAGE");
+
+       map_len = ram >> (HPAGE_SHIFT - 1);
+       map = malloc(map_len);
+       if (!map)
+               errx(2, "map malloc");
+
+       while (1) {
+               int nr_succeed = 0, nr_failed = 0, nr_pages = 0;
+
+               memset(map, 0, map_len);
+
+               clock_gettime(CLOCK_MONOTONIC, &a);
+               for (p = ptr; p < ptr + len; p += HPAGE_SIZE) {
+                       int64_t pfn;
+
+                       pfn = allocate_transhuge(p);
+
+                       if (pfn < 0) {
+                               nr_failed++;
+                       } else {
+                               size_t idx = pfn >> (HPAGE_SHIFT - PAGE_SHIFT);
+
+                               nr_succeed++;
+                               if (idx >= map_len) {
+                                       map = realloc(map, idx + 1);
+                                       if (!map)
+                                               errx(2, "map realloc");
+                                       memset(map + map_len, 0, idx + 1 - map_len);
+                                       map_len = idx + 1;
+                               }
+                               if (!map[idx])
+                                       nr_pages++;
+                               map[idx] = 1;
+                       }
+
+                       /* split transhuge page, keep last page */
+                       if (madvise(p, HPAGE_SIZE - PAGE_SIZE, MADV_DONTNEED))
+                               err(2, "MADV_DONTNEED");
+               }
+               clock_gettime(CLOCK_MONOTONIC, &b);
+               s = b.tv_sec - a.tv_sec + (b.tv_nsec - a.tv_nsec) / 1000000000.;
+
+               warnx("%.3f s/loop, %.3f ms/page, %10.3f MiB/s\t"
+                     "%4d succeed, %4d failed, %4d different pages",
+                     s, s * 1000 / (len >> HPAGE_SHIFT), len / s / (1 << 20),
+                     nr_succeed, nr_failed, nr_pages);
+       }
+}