1 /*
   2  * Copyright (c) 2015, 2020, Oracle and/or its affiliates. All rights reserved.
   3  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   4  *
   5  * This code is free software; you can redistribute it and/or modify it
   6  * under the terms of the GNU General Public License version 2 only, as
   7  * published by the Free Software Foundation.
   8  *
   9  * This code is distributed in the hope that it will be useful, but WITHOUT
  10  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
  11  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License
  12  * version 2 for more details (a copy is included in the LICENSE file that
  13  * accompanied this code).
  14  *
  15  * You should have received a copy of the GNU General Public License version
  16  * 2 along with this work; if not, write to the Free Software Foundation,
  17  * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
  18  *
  19  * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
  20  * or visit www.oracle.com if you need additional information or have any
  21  * questions.
  22  */
  23 
  24 #include "precompiled.hpp"
  25 #include "gc/z/zArray.inline.hpp"
  26 #include "gc/z/zErrno.hpp"
  27 #include "gc/z/zGlobals.hpp"
  28 #include "gc/z/zLargePages.inline.hpp"
  29 #include "gc/z/zMountPoint_linux.hpp"
  30 #include "gc/z/zNUMA.inline.hpp"
  31 #include "gc/z/zPhysicalMemoryBacking_linux.hpp"
  32 #include "gc/z/zSyscall_linux.hpp"
  33 #include "logging/log.hpp"
  34 #include "runtime/init.hpp"
  35 #include "runtime/os.hpp"
  36 #include "utilities/align.hpp"
  37 #include "utilities/debug.hpp"
  38 #include "utilities/growableArray.hpp"
  39 
  40 #include <fcntl.h>
  41 #include <stdio.h>
  42 #include <sys/mman.h>
  43 #include <sys/stat.h>
  44 #include <sys/statfs.h>
  45 #include <sys/types.h>
  46 #include <unistd.h>
  47 
  48 //
  49 // Support for building on older Linux systems
  50 //
  51 
  52 // memfd_create(2) flags
  53 #ifndef MFD_CLOEXEC
  54 #define MFD_CLOEXEC                      0x0001U
  55 #endif
  56 #ifndef MFD_HUGETLB
  57 #define MFD_HUGETLB                      0x0004U
  58 #endif
  59 
  60 // open(2) flags
  61 #ifndef O_CLOEXEC
  62 #define O_CLOEXEC                        02000000
  63 #endif
  64 #ifndef O_TMPFILE
  65 #define O_TMPFILE                        (020000000 | O_DIRECTORY)
  66 #endif
  67 
  68 // fallocate(2) flags
  69 #ifndef FALLOC_FL_KEEP_SIZE
  70 #define FALLOC_FL_KEEP_SIZE              0x01
  71 #endif
  72 #ifndef FALLOC_FL_PUNCH_HOLE
  73 #define FALLOC_FL_PUNCH_HOLE             0x02
  74 #endif
  75 
  76 // Filesystem types, see statfs(2)
  77 #ifndef TMPFS_MAGIC
  78 #define TMPFS_MAGIC                      0x01021994
  79 #endif
  80 #ifndef HUGETLBFS_MAGIC
  81 #define HUGETLBFS_MAGIC                  0x958458f6
  82 #endif
  83 
  84 // Filesystem names
  85 #define ZFILESYSTEM_TMPFS                "tmpfs"
  86 #define ZFILESYSTEM_HUGETLBFS            "hugetlbfs"
  87 
  88 // Proc file entry for max map mount
  89 #define ZFILENAME_PROC_MAX_MAP_COUNT     "/proc/sys/vm/max_map_count"
  90 
  91 // Sysfs file for transparent huge page on tmpfs
  92 #define ZFILENAME_SHMEM_ENABLED          "/sys/kernel/mm/transparent_hugepage/shmem_enabled"
  93 
  94 // Java heap filename
  95 #define ZFILENAME_HEAP                   "java_heap"
  96 
  97 // Preferred tmpfs mount points, ordered by priority
  98 static const char* z_preferred_tmpfs_mountpoints[] = {
  99   "/dev/shm",
 100   "/run/shm",
 101   NULL
 102 };
 103 
 104 // Preferred hugetlbfs mount points, ordered by priority
 105 static const char* z_preferred_hugetlbfs_mountpoints[] = {
 106   "/dev/hugepages",
 107   "/hugepages",
 108   NULL
 109 };
 110 
 111 static int z_fallocate_hugetlbfs_attempts = 3;
 112 static bool z_fallocate_supported = true;
 113 
 114 ZPhysicalMemoryBacking::ZPhysicalMemoryBacking() :
 115     _fd(-1),
 116     _size(0),
 117     _filesystem(0),
 118     _block_size(0),
 119     _available(0),
 120     _initialized(false) {
 121 
 122   // Create backing file
 123   _fd = create_fd(ZFILENAME_HEAP);
 124   if (_fd == -1) {
 125     return;
 126   }
 127 
 128   // Get filesystem statistics
 129   struct statfs buf;
 130   if (fstatfs(_fd, &buf) == -1) {
 131     ZErrno err;
 132     log_error(gc)("Failed to determine filesystem type for backing file (%s)", err.to_string());
 133     return;
 134   }
 135 
 136   _filesystem = buf.f_type;
 137   _block_size = buf.f_bsize;
 138   _available = buf.f_bavail * _block_size;
 139 
 140   // Make sure we're on a supported filesystem
 141   if (!is_tmpfs() && !is_hugetlbfs()) {
 142     log_error(gc)("Backing file must be located on a %s or a %s filesystem",
 143                   ZFILESYSTEM_TMPFS, ZFILESYSTEM_HUGETLBFS);
 144     return;
 145   }
 146 
 147   // Make sure the filesystem type matches requested large page type
 148   if (ZLargePages::is_transparent() && !is_tmpfs()) {
 149     log_error(gc)("-XX:+UseTransparentHugePages can only be enable when using a %s filesystem",
 150                   ZFILESYSTEM_TMPFS);
 151     return;
 152   }
 153 
 154   if (ZLargePages::is_transparent() && !tmpfs_supports_transparent_huge_pages()) {
 155     log_error(gc)("-XX:+UseTransparentHugePages on a %s filesystem not supported by kernel",
 156                   ZFILESYSTEM_TMPFS);
 157     return;
 158   }
 159 
 160   if (ZLargePages::is_explicit() && !is_hugetlbfs()) {
 161     log_error(gc)("-XX:+UseLargePages (without -XX:+UseTransparentHugePages) can only be enabled "
 162                   "when using a %s filesystem", ZFILESYSTEM_HUGETLBFS);
 163     return;
 164   }
 165 
 166   if (!ZLargePages::is_explicit() && is_hugetlbfs()) {
 167     log_error(gc)("-XX:+UseLargePages must be enabled when using a %s filesystem",
 168                   ZFILESYSTEM_HUGETLBFS);
 169     return;
 170   }
 171 
 172   const size_t expected_block_size = is_tmpfs() ? os::vm_page_size() : os::large_page_size();
 173   if (expected_block_size != _block_size) {












 174     log_error(gc)("%s filesystem has unexpected block size " SIZE_FORMAT " (expected " SIZE_FORMAT ")",
 175                   is_tmpfs() ? ZFILESYSTEM_TMPFS : ZFILESYSTEM_HUGETLBFS, _block_size, expected_block_size);
 176     return;
 177   }
 178 
 179   // Successfully initialized
 180   _initialized = true;
 181 }
 182 
 183 int ZPhysicalMemoryBacking::create_mem_fd(const char* name) const {
 184   // Create file name
 185   char filename[PATH_MAX];
 186   snprintf(filename, sizeof(filename), "%s%s", name, ZLargePages::is_explicit() ? ".hugetlb" : "");
 187 
 188   // Create file
 189   const int extra_flags = ZLargePages::is_explicit() ? MFD_HUGETLB : 0;
 190   const int fd = ZSyscall::memfd_create(filename, MFD_CLOEXEC | extra_flags);
 191   if (fd == -1) {
 192     ZErrno err;
 193     log_debug(gc, init)("Failed to create memfd file (%s)",
 194                         ((ZLargePages::is_explicit() && err == EINVAL) ? "Hugepages not supported" : err.to_string()));
 195     return -1;
 196   }
 197 
 198   log_info(gc, init)("Heap backed by file: /memfd:%s", filename);
 199 
 200   return fd;
 201 }
 202 
 203 int ZPhysicalMemoryBacking::create_file_fd(const char* name) const {
 204   const char* const filesystem = ZLargePages::is_explicit()
 205                                  ? ZFILESYSTEM_HUGETLBFS
 206                                  : ZFILESYSTEM_TMPFS;
 207   const char** const preferred_mountpoints = ZLargePages::is_explicit()
 208                                              ? z_preferred_hugetlbfs_mountpoints
 209                                              : z_preferred_tmpfs_mountpoints;
 210 
 211   // Find mountpoint
 212   ZMountPoint mountpoint(filesystem, preferred_mountpoints);
 213   if (mountpoint.get() == NULL) {
 214     log_error(gc)("Use -XX:AllocateHeapAt to specify the path to a %s filesystem", filesystem);
 215     return -1;
 216   }
 217 
 218   // Try to create an anonymous file using the O_TMPFILE flag. Note that this
 219   // flag requires kernel >= 3.11. If this fails we fall back to open/unlink.
 220   const int fd_anon = os::open(mountpoint.get(), O_TMPFILE|O_EXCL|O_RDWR|O_CLOEXEC, S_IRUSR|S_IWUSR);
 221   if (fd_anon == -1) {
 222     ZErrno err;
 223     log_debug(gc, init)("Failed to create anonymous file in %s (%s)", mountpoint.get(),
 224                         (err == EINVAL ? "Not supported" : err.to_string()));
 225   } else {
 226     // Get inode number for anonymous file
 227     struct stat stat_buf;
 228     if (fstat(fd_anon, &stat_buf) == -1) {
 229       ZErrno err;
 230       log_error(gc)("Failed to determine inode number for anonymous file (%s)", err.to_string());
 231       return -1;
 232     }
 233 
 234     log_info(gc, init)("Heap backed by file: %s/#" UINT64_FORMAT, mountpoint.get(), (uint64_t)stat_buf.st_ino);
 235 
 236     return fd_anon;
 237   }
 238 
 239   log_debug(gc, init)("Falling back to open/unlink");
 240 
 241   // Create file name
 242   char filename[PATH_MAX];
 243   snprintf(filename, sizeof(filename), "%s/%s.%d", mountpoint.get(), name, os::current_process_id());
 244 
 245   // Create file
 246   const int fd = os::open(filename, O_CREAT|O_EXCL|O_RDWR|O_CLOEXEC, S_IRUSR|S_IWUSR);
 247   if (fd == -1) {
 248     ZErrno err;
 249     log_error(gc)("Failed to create file %s (%s)", filename, err.to_string());
 250     return -1;
 251   }
 252 
 253   // Unlink file
 254   if (unlink(filename) == -1) {
 255     ZErrno err;
 256     log_error(gc)("Failed to unlink file %s (%s)", filename, err.to_string());
 257     return -1;
 258   }
 259 
 260   log_info(gc, init)("Heap backed by file: %s", filename);
 261 
 262   return fd;
 263 }
 264 
 265 int ZPhysicalMemoryBacking::create_fd(const char* name) const {
 266   if (AllocateHeapAt == NULL) {
 267     // If the path is not explicitly specified, then we first try to create a memfd file
 268     // instead of looking for a tmpfd/hugetlbfs mount point. Note that memfd_create() might
 269     // not be supported at all (requires kernel >= 3.17), or it might not support large
 270     // pages (requires kernel >= 4.14). If memfd_create() fails, then we try to create a
 271     // file on an accessible tmpfs or hugetlbfs mount point.
 272     const int fd = create_mem_fd(name);
 273     if (fd != -1) {
 274       return fd;
 275     }
 276 
 277     log_debug(gc, init)("Falling back to searching for an accessible mount point");
 278   }
 279 
 280   return create_file_fd(name);
 281 }
 282 
 283 bool ZPhysicalMemoryBacking::is_initialized() const {
 284   return _initialized;
 285 }
 286 
 287 void ZPhysicalMemoryBacking::warn_available_space(size_t max) const {
 288   // Note that the available space on a tmpfs or a hugetlbfs filesystem
 289   // will be zero if no size limit was specified when it was mounted.
 290   if (_available == 0) {
 291     // No size limit set, skip check
 292     log_info(gc, init)("Available space on backing filesystem: N/A");
 293     return;
 294   }
 295 
 296   log_info(gc, init)("Available space on backing filesystem: " SIZE_FORMAT "M", _available / M);
 297 
 298   // Warn if the filesystem doesn't currently have enough space available to hold
 299   // the max heap size. The max heap size will be capped if we later hit this limit
 300   // when trying to expand the heap.
 301   if (_available < max) {
 302     log_warning(gc)("***** WARNING! INCORRECT SYSTEM CONFIGURATION DETECTED! *****");
 303     log_warning(gc)("Not enough space available on the backing filesystem to hold the current max Java heap");
 304     log_warning(gc)("size (" SIZE_FORMAT "M). Please adjust the size of the backing filesystem accordingly "
 305                     "(available", max / M);
 306     log_warning(gc)("space is currently " SIZE_FORMAT "M). Continuing execution with the current filesystem "
 307                     "size could", _available / M);
 308     log_warning(gc)("lead to a premature OutOfMemoryError being thrown, due to failure to map memory.");
 309   }
 310 }
 311 
 312 void ZPhysicalMemoryBacking::warn_max_map_count(size_t max) const {
 313   const char* const filename = ZFILENAME_PROC_MAX_MAP_COUNT;
 314   FILE* const file = fopen(filename, "r");
 315   if (file == NULL) {
 316     // Failed to open file, skip check
 317     log_debug(gc, init)("Failed to open %s", filename);
 318     return;
 319   }
 320 
 321   size_t actual_max_map_count = 0;
 322   const int result = fscanf(file, SIZE_FORMAT, &actual_max_map_count);
 323   fclose(file);
 324   if (result != 1) {
 325     // Failed to read file, skip check
 326     log_debug(gc, init)("Failed to read %s", filename);
 327     return;
 328   }
 329 
 330   // The required max map count is impossible to calculate exactly since subsystems
 331   // other than ZGC are also creating memory mappings, and we have no control over that.
 332   // However, ZGC tends to create the most mappings and dominate the total count.
 333   // In the worst cases, ZGC will map each granule three times, i.e. once per heap view.
 334   // We speculate that we need another 20% to allow for non-ZGC subsystems to map memory.
 335   const size_t required_max_map_count = (max / ZGranuleSize) * 3 * 1.2;
 336   if (actual_max_map_count < required_max_map_count) {
 337     log_warning(gc)("***** WARNING! INCORRECT SYSTEM CONFIGURATION DETECTED! *****");
 338     log_warning(gc)("The system limit on number of memory mappings per process might be too low for the given");
 339     log_warning(gc)("max Java heap size (" SIZE_FORMAT "M). Please adjust %s to allow for at",
 340                     max / M, filename);
 341     log_warning(gc)("least " SIZE_FORMAT " mappings (current limit is " SIZE_FORMAT "). Continuing execution "
 342                     "with the current", required_max_map_count, actual_max_map_count);
 343     log_warning(gc)("limit could lead to a fatal error, due to failure to map memory.");
 344   }
 345 }
 346 
 347 void ZPhysicalMemoryBacking::warn_commit_limits(size_t max) const {
 348   // Warn if available space is too low
 349   warn_available_space(max);
 350 
 351   // Warn if max map count is too low
 352   warn_max_map_count(max);
 353 }
 354 
 355 size_t ZPhysicalMemoryBacking::size() const {
 356   return _size;
 357 }
 358 
 359 bool ZPhysicalMemoryBacking::is_tmpfs() const {
 360   return _filesystem == TMPFS_MAGIC;
 361 }
 362 
 363 bool ZPhysicalMemoryBacking::is_hugetlbfs() const {
 364   return _filesystem == HUGETLBFS_MAGIC;
 365 }
 366 
 367 bool ZPhysicalMemoryBacking::tmpfs_supports_transparent_huge_pages() const {
 368   // If the shmem_enabled file exists and is readable then we
 369   // know the kernel supports transparent huge pages for tmpfs.
 370   return access(ZFILENAME_SHMEM_ENABLED, R_OK) == 0;
 371 }
 372 
 373 ZErrno ZPhysicalMemoryBacking::fallocate_compat_ftruncate(size_t size) const {
 374   while (ftruncate(_fd, size) == -1) {
 375     if (errno != EINTR) {
 376       // Failed
 377       return errno;
 378     }
 379   }
 380 
 381   // Success
 382   return 0;
 383 }
 384 
 385 ZErrno ZPhysicalMemoryBacking::fallocate_compat_mmap(size_t offset, size_t length, bool touch) const {
 386   // On hugetlbfs, mapping a file segment will fail immediately, without
 387   // the need to touch the mapped pages first, if there aren't enough huge
 388   // pages available to back the mapping.
 389   void* const addr = mmap(0, length, PROT_READ|PROT_WRITE, MAP_SHARED, _fd, offset);
 390   if (addr == MAP_FAILED) {
 391     // Failed
 392     return errno;
 393   }
 394 
 395   // Once mapped, the huge pages are only reserved. We need to touch them
 396   // to associate them with the file segment. Note that we can not punch
 397   // hole in file segments which only have reserved pages.
 398   if (touch) {
 399     char* const start = (char*)addr;
 400     char* const end = start + length;
 401     os::pretouch_memory(start, end, _block_size);
 402   }
 403 
 404   // Unmap again. From now on, the huge pages that were mapped are allocated
 405   // to this file. There's no risk in getting SIGBUS when touching them.
 406   if (munmap(addr, length) == -1) {
 407     // Failed
 408     return errno;
 409   }
 410 
 411   // Success
 412   return 0;
 413 }
 414 
 415 ZErrno ZPhysicalMemoryBacking::fallocate_compat_pwrite(size_t offset, size_t length) const {
 416   uint8_t data = 0;
 417 
 418   // Allocate backing memory by writing to each block
 419   for (size_t pos = offset; pos < offset + length; pos += _block_size) {
 420     if (pwrite(_fd, &data, sizeof(data), pos) == -1) {
 421       // Failed
 422       return errno;
 423     }
 424   }
 425 
 426   // Success
 427   return 0;
 428 }
 429 
 430 ZErrno ZPhysicalMemoryBacking::fallocate_fill_hole_compat(size_t offset, size_t length) {
 431   // fallocate(2) is only supported by tmpfs since Linux 3.5, and by hugetlbfs
 432   // since Linux 4.3. When fallocate(2) is not supported we emulate it using
 433   // ftruncate/pwrite (for tmpfs) or ftruncate/mmap/munmap (for hugetlbfs).
 434 
 435   const size_t end = offset + length;
 436   if (end > _size) {
 437     // Increase file size
 438     const ZErrno err = fallocate_compat_ftruncate(end);
 439     if (err) {
 440       // Failed
 441       return err;
 442     }
 443   }
 444 
 445   // Allocate backing memory
 446   const ZErrno err = is_hugetlbfs() ? fallocate_compat_mmap(offset, length, false /* touch */)
 447                                     : fallocate_compat_pwrite(offset, length);
 448   if (err) {
 449     if (end > _size) {
 450       // Restore file size
 451       fallocate_compat_ftruncate(_size);
 452     }
 453 
 454     // Failed
 455     return err;
 456   }
 457 
 458   if (end > _size) {
 459     // Record new file size
 460     _size = end;
 461   }
 462 
 463   // Success
 464   return 0;
 465 }
 466 
 467 ZErrno ZPhysicalMemoryBacking::fallocate_fill_hole_syscall(size_t offset, size_t length) {
 468   const int mode = 0; // Allocate
 469   const int res = ZSyscall::fallocate(_fd, mode, offset, length);
 470   if (res == -1) {
 471     // Failed
 472     return errno;
 473   }
 474 
 475   const size_t end = offset + length;
 476   if (end > _size) {
 477     // Record new file size
 478     _size = end;
 479   }
 480 
 481   // Success
 482   return 0;
 483 }
 484 
 485 ZErrno ZPhysicalMemoryBacking::fallocate_fill_hole(size_t offset, size_t length) {
 486   // Using compat mode is more efficient when allocating space on hugetlbfs.
 487   // Note that allocating huge pages this way will only reserve them, and not
 488   // associate them with segments of the file. We must guarantee that we at
 489   // some point touch these segments, otherwise we can not punch hole in them.
 490   if (z_fallocate_supported && !is_hugetlbfs()) {
 491      const ZErrno err = fallocate_fill_hole_syscall(offset, length);
 492      if (!err) {
 493        // Success
 494        return 0;
 495      }
 496 
 497      if (err != ENOSYS && err != EOPNOTSUPP) {
 498        // Failed
 499        return err;
 500      }
 501 
 502      // Not supported
 503      log_debug(gc)("Falling back to fallocate() compatibility mode");
 504      z_fallocate_supported = false;
 505   }
 506 
 507   return fallocate_fill_hole_compat(offset, length);
 508 }
 509 
 510 ZErrno ZPhysicalMemoryBacking::fallocate_punch_hole(size_t offset, size_t length) {
 511   if (is_hugetlbfs()) {
 512     // We can only punch hole in pages that have been touched. Non-touched
 513     // pages are only reserved, and not associated with any specific file
 514     // segment. We don't know which pages have been previously touched, so
 515     // we always touch them here to guarantee that we can punch hole.
 516     const ZErrno err = fallocate_compat_mmap(offset, length, true /* touch */);
 517     if (err) {
 518       // Failed
 519       return err;
 520     }
 521   }
 522 
 523   const int mode = FALLOC_FL_PUNCH_HOLE|FALLOC_FL_KEEP_SIZE;
 524   if (ZSyscall::fallocate(_fd, mode, offset, length) == -1) {
 525     // Failed
 526     return errno;
 527   }
 528 
 529   // Success
 530   return 0;
 531 }
 532 
 533 ZErrno ZPhysicalMemoryBacking::split_and_fallocate(bool punch_hole, size_t offset, size_t length) {
 534   // Try first half
 535   const size_t offset0 = offset;
 536   const size_t length0 = align_up(length / 2, _block_size);
 537   const ZErrno err0 = fallocate(punch_hole, offset0, length0);
 538   if (err0) {
 539     return err0;
 540   }
 541 
 542   // Try second half
 543   const size_t offset1 = offset0 + length0;
 544   const size_t length1 = length - length0;
 545   const ZErrno err1 = fallocate(punch_hole, offset1, length1);
 546   if (err1) {
 547     return err1;
 548   }
 549 
 550   // Success
 551   return 0;
 552 }
 553 
 554 ZErrno ZPhysicalMemoryBacking::fallocate(bool punch_hole, size_t offset, size_t length) {
 555   assert(is_aligned(offset, _block_size), "Invalid offset");
 556   assert(is_aligned(length, _block_size), "Invalid length");
 557 
 558   const ZErrno err = punch_hole ? fallocate_punch_hole(offset, length) : fallocate_fill_hole(offset, length);
 559   if (err == EINTR && length > _block_size) {
 560     // Calling fallocate(2) with a large length can take a long time to
 561     // complete. When running profilers, such as VTune, this syscall will
 562     // be constantly interrupted by signals. Expanding the file in smaller
 563     // steps avoids this problem.
 564     return split_and_fallocate(punch_hole, offset, length);
 565   }
 566 
 567   return err;
 568 }
 569 
 570 bool ZPhysicalMemoryBacking::commit_inner(size_t offset, size_t length) {
 571   log_trace(gc, heap)("Committing memory: " SIZE_FORMAT "M-" SIZE_FORMAT "M (" SIZE_FORMAT "M)",
 572                       offset / M, (offset + length) / M, length / M);
 573 
 574 retry:
 575   const ZErrno err = fallocate(false /* punch_hole */, offset, length);
 576   if (err) {
 577     if (err == ENOSPC && !is_init_completed() && is_hugetlbfs() && z_fallocate_hugetlbfs_attempts-- > 0) {
 578       // If we fail to allocate during initialization, due to lack of space on
 579       // the hugetlbfs filesystem, then we wait and retry a few times before
 580       // giving up. Otherwise there is a risk that running JVMs back-to-back
 581       // will fail, since there is a delay between process termination and the
 582       // huge pages owned by that process being returned to the huge page pool
 583       // and made available for new allocations.
 584       log_debug(gc, init)("Failed to commit memory (%s), retrying", err.to_string());
 585 
 586       // Wait and retry in one second, in the hope that huge pages will be
 587       // available by then.
 588       sleep(1);
 589       goto retry;
 590     }
 591 
 592     // Failed
 593     log_error(gc)("Failed to commit memory (%s)", err.to_string());
 594     return false;
 595   }
 596 
 597   // Success
 598   return true;
 599 }
 600 
 601 static int offset_to_node(size_t offset) {
 602   const GrowableArray<int>* mapping = os::Linux::numa_nindex_to_node();
 603   const size_t nindex = (offset >> ZGranuleSizeShift) % mapping->length();
 604   return mapping->at((int)nindex);
 605 }
 606 
 607 size_t ZPhysicalMemoryBacking::commit_numa_interleaved(size_t offset, size_t length) {
 608   size_t committed = 0;
 609 
 610   // Commit one granule at a time, so that each granule
 611   // can be allocated from a different preferred node.
 612   while (committed < length) {
 613     const size_t granule_offset = offset + committed;
 614 
 615     // Setup NUMA policy to allocate memory from a preferred node
 616     os::Linux::numa_set_preferred(offset_to_node(granule_offset));
 617 
 618     if (!commit_inner(granule_offset, ZGranuleSize)) {
 619       // Failed
 620       break;
 621     }
 622 
 623     committed += ZGranuleSize;
 624   }
 625 
 626   // Restore NUMA policy
 627   os::Linux::numa_set_preferred(-1);
 628 
 629   return committed;
 630 }
 631 
 632 size_t ZPhysicalMemoryBacking::commit_default(size_t offset, size_t length) {
 633   // Try to commit the whole region
 634   if (commit_inner(offset, length)) {
 635     // Success
 636     return length;
 637   }
 638 
 639   // Failed, try to commit as much as possible
 640   size_t start = offset;
 641   size_t end = offset + length;
 642 
 643   for (;;) {
 644     length = align_down((end - start) / 2, ZGranuleSize);
 645     if (length < ZGranuleSize) {
 646       // Done, don't commit more
 647       return start - offset;
 648     }
 649 
 650     if (commit_inner(start, length)) {
 651       // Success, try commit more
 652       start += length;
 653     } else {
 654       // Failed, try commit less
 655       end -= length;
 656     }
 657   }
 658 }
 659 
 660 size_t ZPhysicalMemoryBacking::commit(size_t offset, size_t length) {
 661   if (ZNUMA::is_enabled() && !ZLargePages::is_explicit()) {
 662     // To get granule-level NUMA interleaving when using non-large pages,
 663     // we must explicitly interleave the memory at commit/fallocate time.
 664     return commit_numa_interleaved(offset, length);
 665   }
 666 
 667   return commit_default(offset, length);
 668 }
 669 
 670 size_t ZPhysicalMemoryBacking::uncommit(size_t offset, size_t length) {
 671   log_trace(gc, heap)("Uncommitting memory: " SIZE_FORMAT "M-" SIZE_FORMAT "M (" SIZE_FORMAT "M)",
 672                       offset / M, (offset + length) / M, length / M);
 673 
 674   const ZErrno err = fallocate(true /* punch_hole */, offset, length);
 675   if (err) {
 676     log_error(gc)("Failed to uncommit memory (%s)", err.to_string());
 677     return 0;
 678   }
 679 
 680   return length;
 681 }
 682 
 683 void ZPhysicalMemoryBacking::map(uintptr_t addr, size_t size, uintptr_t offset) const {
 684   const void* const res = mmap((void*)addr, size, PROT_READ|PROT_WRITE, MAP_FIXED|MAP_SHARED, _fd, offset);
 685   if (res == MAP_FAILED) {
 686     ZErrno err;
 687     fatal("Failed to map memory (%s)", err.to_string());
 688   }
 689 }
 690 
 691 void ZPhysicalMemoryBacking::unmap(uintptr_t addr, size_t size) const {
 692   // Note that we must keep the address space reservation intact and just detach
 693   // the backing memory. For this reason we map a new anonymous, non-accessible
 694   // and non-reserved page over the mapping instead of actually unmapping.
 695   const void* const res = mmap((void*)addr, size, PROT_NONE, MAP_FIXED | MAP_ANONYMOUS | MAP_PRIVATE | MAP_NORESERVE, -1, 0);
 696   if (res == MAP_FAILED) {
 697     ZErrno err;
 698     fatal("Failed to map memory (%s)", err.to_string());
 699   }
 700 }
--- EOF ---