Fixes #6020. Detect huge pages. The fractal tree layer notices when you...

Fixes #6020.  Detect huge pages.  The fractal tree layer notices when you create a logger (maybe there are other useful times to detect, but that's enough for now).  The ydb layer notices when you all {{{db_env->open()}}}.    The mysql handlerton notices the error in {{{tokudb_init_func()}}}.

git-svn-id: file:///svn/toku/tokudb@54498 c7de825b-a66e-492c-adef-691d508d4ae1

buildheader/make_tdb.cc

@@ -120,6 +120,7 @@ enum {
         TOKUDB_NEEDS_REPAIR            = -100013,
         TOKUDB_CURSOR_CONTINUE         = -100014,
         TOKUDB_BAD_CHECKSUM            = -100015,
+        TOKUDB_HUGE_PAGES_ENABLED      = -100016,
         DONTUSE_I_JUST_PUT_THIS_HERE_SO_I_COULD_HAVE_A_COMMA_AFTER_EACH_ITEM
 };
 
@@ -270,6 +271,7 @@ static void print_defines (void) {
     dodefine(TOKUDB_NEEDS_REPAIR);
     dodefine(TOKUDB_CURSOR_CONTINUE);
     dodefine(TOKUDB_BAD_CHECKSUM);
+    dodefine(TOKUDB_HUGE_PAGES_ENABLED);
 
     /* LOADER flags */
     printf("/* LOADER flags */\n");

db-benchmark-test/db-benchmark-test.cc

@@ -177,6 +177,7 @@ static void benchmark_setup (void) {
     }
 
     r = db_env_create(&dbenv, 0);
+    if (r!=0) fprintf(stderr, "Error on db_env_create: %s\n", db_strerror(r));
     assert(r == 0);
 
 #if DB_VERSION_MAJOR == 4 && DB_VERSION_MINOR <= 4

ft/logger.cc

@@ -13,6 +13,7 @@
 #include "log-internal.h"
 #include "txn_manager.h"
 #include "rollback_log_node_cache.h"
+#include "huge_page_detection.h"
 #include <util/partitioned_counter.h>
 
 static const int log_format_version=TOKU_LOG_VERSION;
@@ -78,6 +79,11 @@ static bool is_a_logfile (const char *name, long long *number_result) {
 
 // TODO: can't fail
 int toku_logger_create (TOKULOGGER *resultp) {
+    if (complain_and_return_true_if_huge_pages_are_enabled()) {
+        *resultp = NULL;
+        errno = TOKUDB_HUGE_PAGES_ENABLED;
+        return TOKUDB_HUGE_PAGES_ENABLED;
+    }
     TOKULOGGER CALLOC(result);
     if (result==0) return get_error_errno();
     result->is_open=false;

portability/CMakeLists.txt

 include_directories(${CMAKE_CURRENT_SOURCE_DIR})
 
 set(tokuportability_srcs
+  huge_page_detection
   file
   memory
   os_malloc

portability/huge_page_detection.cc

+/* -*- mode: C++; c-basic-offset: 4; indent-tabs-mode: nil -*- */
+// vim: ft=cpp:expandtab:ts=8:sw=4:softtabstop=4:
+#ident "$Id: memory.cc 52238 2013-01-18 20:21:22Z zardosht $"
+#ident "Copyright (c) 2007-2012 Tokutek Inc.  All rights reserved."
+
+#include <sys/mman.h>
+#include <stdlib.h>
+#include <stdio.h>
+#include <string.h>
+
+#include <toku_assert.h>
+
+#include "huge_page_detection.h"
+
+extern "C" {
+
+static bool check_huge_pages_config_file(const char *fname)
+// Effect: Return true if huge pages are there.  If so, print diagnostics.
+{
+    FILE *f=fopen(fname, "r");
+    if (f) {
+        // It's redhat and the feature appears to be there.  Is it enabled?
+        char buf[1000];
+        char *r = fgets(buf, sizeof(buf), f);
+        assert(r!=NULL);
+        if (strstr(buf, "[always]")) {
+            fprintf(stderr, "Transparent huge pages are enabled, according to %s\n", fname);
+            return true;
+        } else {
+            return false;
+        }
+    }
+    return false;
+}
+
+/* struct mapinfo { */
+/*     void *addr; */
+/*     size_t size; */
+/* }; */
+
+/* static void* map_it(size_t size, struct mapinfo *mi, int *n_maps) { */
+/*     void *r = mmap(NULL, size, PROT_READ|PROT_WRITE, MAP_PRIVATE|MAP_ANONYMOUS, -1, 0); */
+/*     if ((long)r==-1) perror("mmap failed"); */
+/*     mi[*n_maps].addr = r; */
+/*     mi[*n_maps].size = size; */
+/*     (*n_maps)++; */
+/*     return r; */
+/* } */
+
+static bool check_huge_pages_in_practice(void)
+// Effect: Return true if huge pages appear to be defined in practice.
+{
+    const size_t TWO_MB = 2UL*1024UL*1024UL;
+
+    void *first = mmap(NULL, 2*TWO_MB, PROT_READ|PROT_WRITE, MAP_PRIVATE|MAP_ANONYMOUS, -1, 0);
+    if ((long)first==-1) perror("mmap failed");
+    {
+        int r = munmap(first, 2*TWO_MB);
+        assert(r==0);
+    }
+
+    void *second_addr = (void*)(((unsigned long)first + TWO_MB) & ~(TWO_MB -1));
+    void *second = mmap(second_addr, TWO_MB, PROT_READ|PROT_WRITE, MAP_FIXED|MAP_PRIVATE|MAP_ANONYMOUS, -1, 0);
+    if ((long)second==-1) perror("mmap failed");
+    assert((long)second%TWO_MB == 0);
+
+    const long pagesize = 4096;
+    const long n_pages = TWO_MB/pagesize;
+    unsigned char vec[n_pages];
+    {
+        int r = mincore(second, TWO_MB, vec);
+        assert(r==0);
+    }
+    for (long i=0; i<n_pages; i++) {
+        assert(!vec[i]);
+    }
+    ((char*)second)[0] = 1;
+    {
+        int r = mincore(second, TWO_MB, vec);
+        assert(r==0);
+    }
+    assert(vec[0]);
+    {
+        int r = munmap(second, TWO_MB);
+        assert(r==0);
+    }
+    if (vec[1]) {
+        fprintf(stderr, "Transparent huge pages appear to be enabled according to mincore()\n");
+        return true;
+    } else {
+        return false;
+    }
+}
+
+bool complain_and_return_true_if_huge_pages_are_enabled(void)
+// Effect: Return true if huge pages appear to be enabled.  If so, print some diagnostics to stderr.
+//  If environment variable TOKU_HUGE_PAGES_OK is set, then don't complain.
+{
+    char *toku_huge_pages_ok = getenv("TOKU_HUGE_PAGES_OK");
+    if (toku_huge_pages_ok) {
+        return false;
+    } else {
+        bool conf1 = check_huge_pages_config_file("/sys/kernel/mm/redhat_transparent_hugepage/enabled");
+        bool conf2 = check_huge_pages_config_file("/sys/kernel/mm/transparent_hugepage/enabled");
+        bool prac = check_huge_pages_in_practice();
+        return conf1|conf2|prac;
+    }
+}
+}

portability/huge_page_detection.h

+extern "C" bool complain_and_return_true_if_huge_pages_are_enabled(void);
+// Effect: Return true if huge pages appear to be enabled.  If so, print some diagnostics to stderr.
+//  If environment variable TOKU_HUGE_PAGES_OK is set, then don't complain.

src/ydb.cc

@@ -2275,7 +2275,7 @@ toku_env_create(DB_ENV ** envp, uint32_t flags) {
     result->i->bt_compare = toku_builtin_compare_fun;
 
     r = toku_logger_create(&result->i->logger);
-    assert_zero(r);
+    if (r!=0) goto cleanup; // In particular, logger_create can return the huge page error.
     assert(result->i->logger);
 
     // Create the locktree manager, passing in the create/destroy/escalate callbacks.
@@ -2659,6 +2659,8 @@ db_strerror(int error) {
             return "User cancelled operation";
         case TOKUDB_NO_DATA:
             return "Ran out of data (not EOF)";
+        case TOKUDB_HUGE_PAGES_ENABLED:
+            return "Transparent huge pages are enabled but TokuDB's memory allocator will oversubscribe main memory with transparent huge pages.  This check can be disabled by setting the environment variable TOKU_HUGE_PAGES_OK.";
     }
 
     static char unknown_result[100];    // Race condition if two threads call this at the same time. However even in a bad case, it should be some sort of null-terminated string.