ceph/src/os/filestore/GenericFileStoreBackend.cc

   1 // -*- mode:C++; tab-width:8; c-basic-offset:2; indent-tabs-mode:t -*-
   2 // vim: ts=8 sw=2 smarttab
   3 /*
   4  * Ceph - scalable distributed file system
   5  *
   6  * Copyright (C) 2004-2006 Sage Weil <sage@newdream.net>
   7  *
   8  * This is free software; you can redistribute it and/or
   9  * modify it under the terms of the GNU Lesser General Public
  10  * License version 2.1, as published by the Free Software
  11  * Foundation.  See file COPYING.
  12  *
  13  */
  14
  15 #include "include/int_types.h"
  16 #include "include/types.h"
  17
  18 #include <unistd.h>
  19 #include <fcntl.h>
  20 #include <errno.h>
  21 #include <stdlib.h>
  22 #include <sys/types.h>
  23 #include <sys/stat.h>
  24 #include <sys/ioctl.h>
  25
  26 #if defined(__linux__)
  27 #include <linux/fs.h>
  28 #endif
  29
  30 #include "include/compat.h"
  31 #include "include/linux_fiemap.h"
  32
  33 #include <iostream>
  34 #include <fstream>
  35 #include <sstream>
  36
  37 #include "GenericFileStoreBackend.h"
  38
  39 #include "common/errno.h"
  40 #include "common/config.h"
  41 #include "common/sync_filesystem.h"
  42 #include "common/blkdev.h"
  43
  44 #include "common/SloppyCRCMap.h"
  45 #include "os/filestore/chain_xattr.h"
  46
  47 #define SLOPPY_CRC_XATTR "user.cephos.scrc"
  48
  49
  50 #define dout_context cct()
  51 #define dout_subsys ceph_subsys_filestore
  52 #undef dout_prefix
  53 #define dout_prefix *_dout << "genericfilestorebackend(" << get_basedir_path() << ") "
  54
  55 #define ALIGN_DOWN(x, by) ((x) - ((x) % (by)))
  56 #define ALIGNED(x, by) (!((x) % (by)))
  57 #define ALIGN_UP(x, by) (ALIGNED((x), (by)) ? (x) : (ALIGN_DOWN((x), (by)) + (by)))
  58
  59 GenericFileStoreBackend::GenericFileStoreBackend(FileStore *fs):
  60   FileStoreBackend(fs),
  61   ioctl_fiemap(false),
  62   seek_data_hole(false),
  63   use_splice(false),
  64   m_filestore_fiemap(cct()->_conf->filestore_fiemap),
  65   m_filestore_seek_data_hole(cct()->_conf->filestore_seek_data_hole),
  66   m_filestore_fsync_flushes_journal_data(cct()->_conf->filestore_fsync_flushes_journal_data),
  67   m_filestore_splice(cct()->_conf->filestore_splice)
  68 {
  69   // rotational?
  70   {
  71     // NOTE: the below won't work on btrfs; we'll assume rotational.
  72     string fn = get_basedir_path();
  73     int fd = ::open(fn.c_str(), O_RDONLY|O_CLOEXEC);
  74     if (fd < 0) {
  75       return;
  76     }
  77     BlkDev blkdev(fd);
  78     m_rotational = blkdev.is_rotational();
  79     dout(20) << __func__ << " basedir " << fn
  80              << " rotational " << (int)m_rotational << dendl;
  81     ::close(fd);
  82   }
  83   // journal rotational?
  84   {
  85     // NOTE: the below won't work on btrfs; we'll assume rotational.
  86     string fn = get_journal_path();
  87     int fd = ::open(fn.c_str(), O_RDONLY|O_CLOEXEC);
  88     if (fd < 0) {
  89       return;
  90     }
  91     BlkDev blkdev(fd);
  92     m_journal_rotational = blkdev.is_rotational();
  93     dout(20) << __func__ << " journal filename " << fn.c_str()
  94              << " journal rotational " << (int)m_journal_rotational << dendl;
  95     ::close(fd);
  96   }
  97 }
  98
  99 int GenericFileStoreBackend::detect_features()
 100 {
 101   char fn[PATH_MAX];
 102   snprintf(fn, sizeof(fn), "%s/fiemap_test", get_basedir_path().c_str());
 103
 104   int fd = ::open(fn, O_CREAT|O_RDWR|O_TRUNC|O_CLOEXEC, 0644);
 105   if (fd < 0) {
 106     fd = -errno;
 107     derr << "detect_features: unable to create " << fn << ": " << cpp_strerror(fd) << dendl;
 108     return fd;
 109   }
 110
 111   // ext4 has a bug in older kernels where fiemap will return an empty
 112   // result in some cases.  this is a file layout that triggers the bug
 113   // on 2.6.34-rc5.
 114   int v[] = {
 115     0x0000000000016000, 0x0000000000007000,
 116     0x000000000004a000, 0x0000000000007000,
 117     0x0000000000060000, 0x0000000000001000,
 118     0x0000000000061000, 0x0000000000008000,
 119     0x0000000000069000, 0x0000000000007000,
 120     0x00000000000a3000, 0x000000000000c000,
 121     0x000000000024e000, 0x000000000000c000,
 122     0x000000000028b000, 0x0000000000009000,
 123     0x00000000002b1000, 0x0000000000003000,
 124     0, 0
 125   };
 126   for (int i=0; v[i]; i++) {
 127     int off = v[i++];
 128     int len = v[i];
 129
 130     // write a large extent
 131     char buf[len];
 132     memset(buf, 1, sizeof(buf));
 133     int r = ::lseek(fd, off, SEEK_SET);
 134     if (r < 0) {
 135       r = -errno;
 136       derr << "detect_features: failed to lseek " << fn << ": " << cpp_strerror(r) << dendl;
 137       VOID_TEMP_FAILURE_RETRY(::close(fd));
 138       return r;
 139     }
 140     r = write(fd, buf, sizeof(buf));
 141     if (r < 0) {
 142       derr << "detect_features: failed to write to " << fn << ": " << cpp_strerror(r) << dendl;
 143       VOID_TEMP_FAILURE_RETRY(::close(fd));
 144       return r;
 145     }
 146   }
 147
 148   // fiemap an extent inside that
 149   if (!m_filestore_fiemap) {
 150     dout(0) << "detect_features: FIEMAP ioctl is disabled via 'filestore fiemap' config option" << dendl;
 151     ioctl_fiemap = false;
 152   } else {
 153     struct fiemap *fiemap;
 154     int r = do_fiemap(fd, 2430421, 59284, &fiemap);
 155     if (r < 0) {
 156       dout(0) << "detect_features: FIEMAP ioctl is NOT supported" << dendl;
 157       ioctl_fiemap = false;
 158     } else {
 159       if (fiemap->fm_mapped_extents == 0) {
 160         dout(0) << "detect_features: FIEMAP ioctl is supported, but buggy -- upgrade your kernel" << dendl;
 161         ioctl_fiemap = false;
 162       } else {
 163         dout(0) << "detect_features: FIEMAP ioctl is supported and appears to work" << dendl;
 164         ioctl_fiemap = true;
 165       }
 166       free(fiemap);
 167     }
 168   }
 169
 170   // SEEK_DATA/SEEK_HOLE detection
 171   if (!m_filestore_seek_data_hole) {
 172     dout(0) << "detect_features: SEEK_DATA/SEEK_HOLE is disabled via 'filestore seek data hole' config option" << dendl;
 173     seek_data_hole = false;
 174   } else {
 175 #if defined(__linux__) && defined(SEEK_HOLE) && defined(SEEK_DATA)
 176     // If compiled on an OS with SEEK_HOLE/SEEK_DATA support, but running
 177     // on an OS that doesn't support SEEK_HOLE/SEEK_DATA, EINVAL is returned.
 178     // Fall back to use fiemap.
 179     off_t hole_pos;
 180
 181     hole_pos = lseek(fd, 0, SEEK_HOLE);
 182     if (hole_pos < 0) {
 183       if (errno == EINVAL) {
 184         dout(0) << "detect_features: lseek SEEK_DATA/SEEK_HOLE is NOT supported" << dendl;
 185         seek_data_hole = false;
 186       } else {
 187         derr << "detect_features: failed to lseek " << fn << ": " << cpp_strerror(-errno) << dendl;
 188         VOID_TEMP_FAILURE_RETRY(::close(fd));
 189         return -errno;
 190       }
 191     } else {
 192       dout(0) << "detect_features: lseek SEEK_DATA/SEEK_HOLE is supported" << dendl;
 193       seek_data_hole = true;
 194     }
 195 #endif
 196   }
 197
 198   //splice detection
 199 #ifdef CEPH_HAVE_SPLICE
 200   if (!m_filestore_splice) {
 201     dout(0) << __func__ << ": splice() is disabled via 'filestore splice' config option" << dendl;
 202     use_splice = false;
 203   } else {
 204     int pipefd[2];
 205     loff_t off_in = 0;
 206     int r;
 207     if (pipe_cloexec(pipefd, 0) < 0) {
 208       int e = errno;
 209       dout(0) << "detect_features: splice pipe met error " << cpp_strerror(e) << dendl;
 210     } else {
 211       lseek(fd, 0, SEEK_SET);
 212       r = splice(fd, &off_in, pipefd[1], NULL, 10, 0);
 213       if (!(r < 0 && errno == EINVAL)) {
 214         use_splice = true;
 215         dout(0) << "detect_features: splice is supported" << dendl;
 216       } else
 217         dout(0) << "detect_features: splice is NOT supported" << dendl;
 218       close(pipefd[0]);
 219       close(pipefd[1]);
 220     }
 221   }
 222 #endif
 223   ::unlink(fn);
 224   VOID_TEMP_FAILURE_RETRY(::close(fd));
 225
 226
 227   bool have_syncfs = false;
 228 #ifdef HAVE_SYS_SYNCFS
 229   if (::syncfs(get_basedir_fd()) == 0) {
 230     dout(0) << "detect_features: syncfs(2) syscall fully supported (by glibc and kernel)" << dendl;
 231     have_syncfs = true;
 232   } else {
 233     dout(0) << "detect_features: syncfs(2) syscall supported by glibc BUT NOT the kernel" << dendl;
 234   }
 235 #elif defined(SYS_syncfs)
 236   if (syscall(SYS_syncfs, get_basedir_fd()) == 0) {
 237     dout(0) << "detect_features: syscall(SYS_syncfs, fd) fully supported" << dendl;
 238     have_syncfs = true;
 239   } else {
 240     dout(0) << "detect_features: syscall(SYS_syncfs, fd) supported by libc BUT NOT the kernel" << dendl;
 241   }
 242 #elif defined(__NR_syncfs)
 243   if (syscall(__NR_syncfs, get_basedir_fd()) == 0) {
 244     dout(0) << "detect_features: syscall(__NR_syncfs, fd) fully supported" << dendl;
 245     have_syncfs = true;
 246   } else {
 247     dout(0) << "detect_features: syscall(__NR_syncfs, fd) supported by libc BUT NOT the kernel" << dendl;
 248   }
 249 #endif
 250   if (!have_syncfs) {
 251     dout(0) << "detect_features: syncfs(2) syscall not supported" << dendl;
 252     if (m_filestore_fsync_flushes_journal_data) {
 253       dout(0) << "detect_features: no syncfs(2), but 'filestore fsync flushes journal data = true', so fsync will suffice." << dendl;
 254     } else {
 255       dout(0) << "detect_features: no syncfs(2), must use sync(2)." << dendl;
 256       dout(0) << "detect_features: WARNING: multiple ceph-osd daemons on the same host will be slow" << dendl;
 257     }
 258   }
 259
 260   return 0;
 261 }
 262
 263 int GenericFileStoreBackend::create_current()
 264 {
 265   struct stat st;
 266   int ret = ::stat(get_current_path().c_str(), &st);
 267   if (ret == 0) {
 268     // current/ exists
 269     if (!S_ISDIR(st.st_mode)) {
 270       dout(0) << "_create_current: current/ exists but is not a directory" << dendl;
 271       ret = -EINVAL;
 272     }
 273   } else {
 274     ret = ::mkdir(get_current_path().c_str(), 0755);
 275     if (ret < 0) {
 276       ret = -errno;
 277       dout(0) << "_create_current: mkdir " << get_current_path() << " failed: "<< cpp_strerror(ret) << dendl;
 278     }
 279   }
 280   return ret;
 281 }
 282
 283 int GenericFileStoreBackend::syncfs()
 284 {
 285   int ret;
 286   if (m_filestore_fsync_flushes_journal_data) {
 287     dout(15) << "syncfs: doing fsync on " << get_op_fd() << dendl;
 288     // make the file system's journal commit.
 289     //  this works with ext3, but NOT ext4
 290     ret = ::fsync(get_op_fd());
 291     if (ret < 0)
 292       ret = -errno;
 293   } else {
 294     dout(15) << "syncfs: doing a full sync (syncfs(2) if possible)" << dendl;
 295     ret = sync_filesystem(get_current_fd());
 296   }
 297   return ret;
 298 }
 299
 300 int GenericFileStoreBackend::do_fiemap(int fd, off_t start, size_t len, struct fiemap **pfiemap)
 301 {
 302   struct fiemap *fiemap = NULL;
 303   struct fiemap *_realloc_fiemap = NULL;
 304   int size;
 305   int ret;
 306
 307   fiemap = (struct fiemap*)calloc(sizeof(struct fiemap), 1);
 308   if (!fiemap)
 309     return -ENOMEM;
 310   /*
 311    * There is a bug on xfs about fiemap. Suppose(offset=3990, len=4096),
 312    * the result is (logical=4096, len=4096). It leak the [3990, 4096).
 313    * Commit:"xfs: fix rounding error of fiemap length parameter
 314    * (eedf32bfcace7d8e20cc66757d74fc68f3439ff7)" fix this bug.
 315    * Here, we make offset aligned with CEPH_PAGE_SIZE to avoid this bug.
 316    */
 317   fiemap->fm_start = start - start % CEPH_PAGE_SIZE;
 318   fiemap->fm_length = len + start % CEPH_PAGE_SIZE;
 319   fiemap->fm_flags = FIEMAP_FLAG_SYNC; /* flush extents to disk if needed */
 320
 321 #if defined(__APPLE__) || defined(__FreeBSD__)
 322   ret = -ENOTSUP;
 323   goto done_err;
 324 #else
 325   if (ioctl(fd, FS_IOC_FIEMAP, fiemap) < 0) {
 326     ret = -errno;
 327     goto done_err;
 328   }
 329 #endif
 330   size = sizeof(struct fiemap_extent) * (fiemap->fm_mapped_extents);
 331
 332   _realloc_fiemap = (struct fiemap *)realloc(fiemap, sizeof(struct fiemap) + size);
 333   if (!_realloc_fiemap) {
 334     ret = -ENOMEM;
 335     goto done_err;
 336   } else {
 337     fiemap = _realloc_fiemap;
 338   }
 339
 340   memset(fiemap->fm_extents, 0, size);
 341
 342   fiemap->fm_extent_count = fiemap->fm_mapped_extents;
 343   fiemap->fm_mapped_extents = 0;
 344
 345 #if defined(__APPLE__) || defined(__FreeBSD__)
 346   ret = -ENOTSUP;
 347   goto done_err;
 348 #else
 349   if (ioctl(fd, FS_IOC_FIEMAP, fiemap) < 0) {
 350     ret = -errno;
 351     goto done_err;
 352   }
 353   *pfiemap = fiemap;
 354 #endif
 355   return 0;
 356
 357 done_err:
 358   *pfiemap = NULL;
 359   free(fiemap);
 360   return ret;
 361 }
 362
 363
 364 int GenericFileStoreBackend::_crc_load_or_init(int fd, SloppyCRCMap *cm)
 365 {
 366   char buf[100];
 367   bufferptr bp;
 368   int r = 0;
 369   int l = chain_fgetxattr(fd, SLOPPY_CRC_XATTR, buf, sizeof(buf));
 370   if (l == -ENODATA) {
 371     return 0;
 372   }
 373   if (l >= 0) {
 374     bp = buffer::create(l);
 375     memcpy(bp.c_str(), buf, l);
 376   } else if (l == -ERANGE) {
 377     l = chain_fgetxattr(fd, SLOPPY_CRC_XATTR, 0, 0);
 378     if (l > 0) {
 379       bp = buffer::create(l);
 380       l = chain_fgetxattr(fd, SLOPPY_CRC_XATTR, bp.c_str(), l);
 381     }
 382   }
 383   bufferlist bl;
 384   bl.append(std::move(bp));
 385   auto p = bl.cbegin();
 386   try {
 387     decode(*cm, p);
 388   }
 389   catch (buffer::error &e) {
 390     r = -EIO;
 391   }
 392   if (r < 0)
 393     derr << __func__ << " got " << cpp_strerror(r) << dendl;
 394   return r;
 395 }
 396
 397 int GenericFileStoreBackend::_crc_save(int fd, SloppyCRCMap *cm)
 398 {
 399   bufferlist bl;
 400   encode(*cm, bl);
 401   int r = chain_fsetxattr(fd, SLOPPY_CRC_XATTR, bl.c_str(), bl.length());
 402   if (r < 0)
 403     derr << __func__ << " got " << cpp_strerror(r) << dendl;
 404   return r;
 405 }
 406
 407 int GenericFileStoreBackend::_crc_update_write(int fd, loff_t off, size_t len, const bufferlist& bl)
 408 {
 409   SloppyCRCMap scm(get_crc_block_size());
 410   int r = _crc_load_or_init(fd, &scm);
 411   if (r < 0)
 412     return r;
 413   ostringstream ss;
 414   scm.write(off, len, bl, &ss);
 415   dout(30) << __func__ << "\n" << ss.str() << dendl;
 416   r = _crc_save(fd, &scm);
 417   return r;
 418 }
 419
 420 int GenericFileStoreBackend::_crc_update_truncate(int fd, loff_t off)
 421 {
 422   SloppyCRCMap scm(get_crc_block_size());
 423   int r = _crc_load_or_init(fd, &scm);
 424   if (r < 0)
 425     return r;
 426   scm.truncate(off);
 427   r = _crc_save(fd, &scm);
 428   return r;
 429 }
 430
 431 int GenericFileStoreBackend::_crc_update_zero(int fd, loff_t off, size_t len)
 432 {
 433   SloppyCRCMap scm(get_crc_block_size());
 434   int r = _crc_load_or_init(fd, &scm);
 435   if (r < 0)
 436     return r;
 437   scm.zero(off, len);
 438   r = _crc_save(fd, &scm);
 439   return r;
 440 }
 441
 442 int GenericFileStoreBackend::_crc_update_clone_range(int srcfd, int destfd,
 443                                                      loff_t srcoff, size_t len, loff_t dstoff)
 444 {
 445   SloppyCRCMap scm_src(get_crc_block_size());
 446   SloppyCRCMap scm_dst(get_crc_block_size());
 447   int r = _crc_load_or_init(srcfd, &scm_src);
 448   if (r < 0)
 449     return r;
 450   r = _crc_load_or_init(destfd, &scm_dst);
 451   if (r < 0)
 452     return r;
 453   ostringstream ss;
 454   scm_dst.clone_range(srcoff, len, dstoff, scm_src, &ss);
 455   dout(30) << __func__ << "\n" << ss.str() << dendl;
 456   r = _crc_save(destfd, &scm_dst);
 457   return r;
 458 }
 459
 460 int GenericFileStoreBackend::_crc_verify_read(int fd, loff_t off, size_t len, const bufferlist& bl,
 461                                               ostream *out)
 462 {
 463   SloppyCRCMap scm(get_crc_block_size());
 464   int r = _crc_load_or_init(fd, &scm);
 465   if (r < 0)
 466     return r;
 467   return scm.read(off, len, bl, out);
 468 }