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);
  74     if (fd < 0) {
  75       return;
  76     }
  77     char partition[PATH_MAX], devname[PATH_MAX];
  78     int r = get_device_by_fd(fd, partition, devname, sizeof(devname));
  79     if (r < 0) {
  80       dout(1) << "unable to get device name for " << get_basedir_path() << ": "
  81               << cpp_strerror(r) << dendl;
  82       m_rotational = true;
  83     } else {
  84       m_rotational = block_device_is_rotational(devname);
  85       dout(20) << __func__ << " devname " << devname
  86                << " rotational " << (int)m_rotational << dendl;
  87     }
  88     ::close(fd);
  89   }
  90 }
  91
  92 int GenericFileStoreBackend::detect_features()
  93 {
  94   char fn[PATH_MAX];
  95   snprintf(fn, sizeof(fn), "%s/fiemap_test", get_basedir_path().c_str());
  96
  97   int fd = ::open(fn, O_CREAT|O_RDWR|O_TRUNC, 0644);
  98   if (fd < 0) {
  99     fd = -errno;
 100     derr << "detect_features: unable to create " << fn << ": " << cpp_strerror(fd) << dendl;
 101     return fd;
 102   }
 103
 104   // ext4 has a bug in older kernels where fiemap will return an empty
 105   // result in some cases.  this is a file layout that triggers the bug
 106   // on 2.6.34-rc5.
 107   int v[] = {
 108     0x0000000000016000, 0x0000000000007000,
 109     0x000000000004a000, 0x0000000000007000,
 110     0x0000000000060000, 0x0000000000001000,
 111     0x0000000000061000, 0x0000000000008000,
 112     0x0000000000069000, 0x0000000000007000,
 113     0x00000000000a3000, 0x000000000000c000,
 114     0x000000000024e000, 0x000000000000c000,
 115     0x000000000028b000, 0x0000000000009000,
 116     0x00000000002b1000, 0x0000000000003000,
 117     0, 0
 118   };
 119   for (int i=0; v[i]; i++) {
 120     int off = v[i++];
 121     int len = v[i];
 122
 123     // write a large extent
 124     char buf[len];
 125     memset(buf, 1, sizeof(buf));
 126     int r = ::lseek(fd, off, SEEK_SET);
 127     if (r < 0) {
 128       r = -errno;
 129       derr << "detect_features: failed to lseek " << fn << ": " << cpp_strerror(r) << dendl;
 130       VOID_TEMP_FAILURE_RETRY(::close(fd));
 131       return r;
 132     }
 133     r = write(fd, buf, sizeof(buf));
 134     if (r < 0) {
 135       derr << "detect_features: failed to write to " << fn << ": " << cpp_strerror(r) << dendl;
 136       VOID_TEMP_FAILURE_RETRY(::close(fd));
 137       return r;
 138     }
 139   }
 140
 141   // fiemap an extent inside that
 142   if (!m_filestore_fiemap) {
 143     dout(0) << "detect_features: FIEMAP ioctl is disabled via 'filestore fiemap' config option" << dendl;
 144     ioctl_fiemap = false;
 145   } else {
 146     struct fiemap *fiemap;
 147     int r = do_fiemap(fd, 2430421, 59284, &fiemap);
 148     if (r < 0) {
 149       dout(0) << "detect_features: FIEMAP ioctl is NOT supported" << dendl;
 150       ioctl_fiemap = false;
 151     } else {
 152       if (fiemap->fm_mapped_extents == 0) {
 153         dout(0) << "detect_features: FIEMAP ioctl is supported, but buggy -- upgrade your kernel" << dendl;
 154         ioctl_fiemap = false;
 155       } else {
 156         dout(0) << "detect_features: FIEMAP ioctl is supported and appears to work" << dendl;
 157         ioctl_fiemap = true;
 158       }
 159       free(fiemap);
 160     }
 161   }
 162
 163   // SEEK_DATA/SEEK_HOLE detection
 164   if (!m_filestore_seek_data_hole) {
 165     dout(0) << "detect_features: SEEK_DATA/SEEK_HOLE is disabled via 'filestore seek data hole' config option" << dendl;
 166     seek_data_hole = false;
 167   } else {
 168 #if defined(__linux__) && defined(SEEK_HOLE) && defined(SEEK_DATA)
 169     // If compiled on an OS with SEEK_HOLE/SEEK_DATA support, but running
 170     // on an OS that doesn't support SEEK_HOLE/SEEK_DATA, EINVAL is returned.
 171     // Fall back to use fiemap.
 172     off_t hole_pos;
 173
 174     hole_pos = lseek(fd, 0, SEEK_HOLE);
 175     if (hole_pos < 0) {
 176       if (errno == EINVAL) {
 177         dout(0) << "detect_features: lseek SEEK_DATA/SEEK_HOLE is NOT supported" << dendl;
 178         seek_data_hole = false;
 179       } else {
 180         derr << "detect_features: failed to lseek " << fn << ": " << cpp_strerror(-errno) << dendl;
 181         VOID_TEMP_FAILURE_RETRY(::close(fd));
 182         return -errno;
 183       }
 184     } else {
 185       dout(0) << "detect_features: lseek SEEK_DATA/SEEK_HOLE is supported" << dendl;
 186       seek_data_hole = true;
 187     }
 188 #endif
 189   }
 190
 191   //splice detection
 192 #ifdef CEPH_HAVE_SPLICE
 193   if (!m_filestore_splice) {
 194     dout(0) << __func__ << ": splice() is disabled via 'filestore splice' config option" << dendl;
 195     use_splice = false;
 196   } else {
 197     int pipefd[2];
 198     loff_t off_in = 0;
 199     int r;
 200     if ((r = pipe(pipefd)) < 0)
 201       dout(0) << "detect_features: splice  pipe met error " << cpp_strerror(errno) << dendl;
 202     else {
 203       lseek(fd, 0, SEEK_SET);
 204       r = splice(fd, &off_in, pipefd[1], NULL, 10, 0);
 205       if (!(r < 0 && errno == EINVAL)) {
 206         use_splice = true;
 207         dout(0) << "detect_features: splice is supported" << dendl;
 208       } else
 209         dout(0) << "detect_features: splice is NOT supported" << dendl;
 210       close(pipefd[0]);
 211       close(pipefd[1]);
 212     }
 213   }
 214 #endif
 215   ::unlink(fn);
 216   VOID_TEMP_FAILURE_RETRY(::close(fd));
 217
 218
 219   bool have_syncfs = false;
 220 #ifdef HAVE_SYS_SYNCFS
 221   if (::syncfs(get_basedir_fd()) == 0) {
 222     dout(0) << "detect_features: syncfs(2) syscall fully supported (by glibc and kernel)" << dendl;
 223     have_syncfs = true;
 224   } else {
 225     dout(0) << "detect_features: syncfs(2) syscall supported by glibc BUT NOT the kernel" << dendl;
 226   }
 227 #elif defined(SYS_syncfs)
 228   if (syscall(SYS_syncfs, get_basedir_fd()) == 0) {
 229     dout(0) << "detect_features: syscall(SYS_syncfs, fd) fully supported" << dendl;
 230     have_syncfs = true;
 231   } else {
 232     dout(0) << "detect_features: syscall(SYS_syncfs, fd) supported by libc BUT NOT the kernel" << dendl;
 233   }
 234 #elif defined(__NR_syncfs)
 235   if (syscall(__NR_syncfs, get_basedir_fd()) == 0) {
 236     dout(0) << "detect_features: syscall(__NR_syncfs, fd) fully supported" << dendl;
 237     have_syncfs = true;
 238   } else {
 239     dout(0) << "detect_features: syscall(__NR_syncfs, fd) supported by libc BUT NOT the kernel" << dendl;
 240   }
 241 #endif
 242   if (!have_syncfs) {
 243     dout(0) << "detect_features: syncfs(2) syscall not supported" << dendl;
 244     if (m_filestore_fsync_flushes_journal_data) {
 245       dout(0) << "detect_features: no syncfs(2), but 'filestore fsync flushes journal data = true', so fsync will suffice." << dendl;
 246     } else {
 247       dout(0) << "detect_features: no syncfs(2), must use sync(2)." << dendl;
 248       dout(0) << "detect_features: WARNING: multiple ceph-osd daemons on the same host will be slow" << dendl;
 249     }
 250   }
 251
 252   return 0;
 253 }
 254
 255 int GenericFileStoreBackend::create_current()
 256 {
 257   struct stat st;
 258   int ret = ::stat(get_current_path().c_str(), &st);
 259   if (ret == 0) {
 260     // current/ exists
 261     if (!S_ISDIR(st.st_mode)) {
 262       dout(0) << "_create_current: current/ exists but is not a directory" << dendl;
 263       ret = -EINVAL;
 264     }
 265   } else {
 266     ret = ::mkdir(get_current_path().c_str(), 0755);
 267     if (ret < 0) {
 268       ret = -errno;
 269       dout(0) << "_create_current: mkdir " << get_current_path() << " failed: "<< cpp_strerror(ret) << dendl;
 270     }
 271   }
 272   return ret;
 273 }
 274
 275 int GenericFileStoreBackend::syncfs()
 276 {
 277   int ret;
 278   if (m_filestore_fsync_flushes_journal_data) {
 279     dout(15) << "syncfs: doing fsync on " << get_op_fd() << dendl;
 280     // make the file system's journal commit.
 281     //  this works with ext3, but NOT ext4
 282     ret = ::fsync(get_op_fd());
 283     if (ret < 0)
 284       ret = -errno;
 285   } else {
 286     dout(15) << "syncfs: doing a full sync (syncfs(2) if possible)" << dendl;
 287     ret = sync_filesystem(get_current_fd());
 288   }
 289   return ret;
 290 }
 291
 292 int GenericFileStoreBackend::do_fiemap(int fd, off_t start, size_t len, struct fiemap **pfiemap)
 293 {
 294   struct fiemap *fiemap = NULL;
 295   struct fiemap *_realloc_fiemap = NULL;
 296   int size;
 297   int ret;
 298
 299   fiemap = (struct fiemap*)calloc(sizeof(struct fiemap), 1);
 300   if (!fiemap)
 301     return -ENOMEM;
 302   /*
 303    * There is a bug on xfs about fiemap. Suppose(offset=3990, len=4096),
 304    * the result is (logical=4096, len=4096). It leak the [3990, 4096).
 305    * Commit:"xfs: fix rounding error of fiemap length parameter
 306    * (eedf32bfcace7d8e20cc66757d74fc68f3439ff7)" fix this bug.
 307    * Here, we make offset aligned with CEPH_PAGE_SIZE to avoid this bug.
 308    */
 309   fiemap->fm_start = start - start % CEPH_PAGE_SIZE;
 310   fiemap->fm_length = len + start % CEPH_PAGE_SIZE;
 311   fiemap->fm_flags = FIEMAP_FLAG_SYNC; /* flush extents to disk if needed */
 312
 313 #if defined(DARWIN) || defined(__FreeBSD__)
 314   ret = -ENOTSUP;
 315   goto done_err;
 316 #else
 317   if (ioctl(fd, FS_IOC_FIEMAP, fiemap) < 0) {
 318     ret = -errno;
 319     goto done_err;
 320   }
 321 #endif
 322   size = sizeof(struct fiemap_extent) * (fiemap->fm_mapped_extents);
 323
 324   _realloc_fiemap = (struct fiemap *)realloc(fiemap, sizeof(struct fiemap) + size);
 325   if (!_realloc_fiemap) {
 326     ret = -ENOMEM;
 327     goto done_err;
 328   } else {
 329     fiemap = _realloc_fiemap;
 330   }
 331
 332   memset(fiemap->fm_extents, 0, size);
 333
 334   fiemap->fm_extent_count = fiemap->fm_mapped_extents;
 335   fiemap->fm_mapped_extents = 0;
 336
 337 #if defined(DARWIN) || defined(__FreeBSD__)
 338   ret = -ENOTSUP;
 339   goto done_err;
 340 #else
 341   if (ioctl(fd, FS_IOC_FIEMAP, fiemap) < 0) {
 342     ret = -errno;
 343     goto done_err;
 344   }
 345   *pfiemap = fiemap;
 346 #endif
 347   return 0;
 348
 349 done_err:
 350   *pfiemap = NULL;
 351   free(fiemap);
 352   return ret;
 353 }
 354
 355
 356 int GenericFileStoreBackend::_crc_load_or_init(int fd, SloppyCRCMap *cm)
 357 {
 358   char buf[100];
 359   bufferptr bp;
 360   int r = 0;
 361   int l = chain_fgetxattr(fd, SLOPPY_CRC_XATTR, buf, sizeof(buf));
 362   if (l == -ENODATA) {
 363     return 0;
 364   }
 365   if (l >= 0) {
 366     bp = buffer::create(l);
 367     memcpy(bp.c_str(), buf, l);
 368   } else if (l == -ERANGE) {
 369     l = chain_fgetxattr(fd, SLOPPY_CRC_XATTR, 0, 0);
 370     if (l > 0) {
 371       bp = buffer::create(l);
 372       l = chain_fgetxattr(fd, SLOPPY_CRC_XATTR, bp.c_str(), l);
 373     }
 374   }
 375   bufferlist bl;
 376   bl.append(std::move(bp));
 377   bufferlist::iterator p = bl.begin();
 378   try {
 379     ::decode(*cm, p);
 380   }
 381   catch (buffer::error &e) {
 382     r = -EIO;
 383   }
 384   if (r < 0)
 385     derr << __func__ << " got " << cpp_strerror(r) << dendl;
 386   return r;
 387 }
 388
 389 int GenericFileStoreBackend::_crc_save(int fd, SloppyCRCMap *cm)
 390 {
 391   bufferlist bl;
 392   ::encode(*cm, bl);
 393   int r = chain_fsetxattr(fd, SLOPPY_CRC_XATTR, bl.c_str(), bl.length());
 394   if (r < 0)
 395     derr << __func__ << " got " << cpp_strerror(r) << dendl;
 396   return r;
 397 }
 398
 399 int GenericFileStoreBackend::_crc_update_write(int fd, loff_t off, size_t len, const bufferlist& bl)
 400 {
 401   SloppyCRCMap scm(get_crc_block_size());
 402   int r = _crc_load_or_init(fd, &scm);
 403   if (r < 0)
 404     return r;
 405   ostringstream ss;
 406   scm.write(off, len, bl, &ss);
 407   dout(30) << __func__ << "\n" << ss.str() << dendl;
 408   r = _crc_save(fd, &scm);
 409   return r;
 410 }
 411
 412 int GenericFileStoreBackend::_crc_update_truncate(int fd, loff_t off)
 413 {
 414   SloppyCRCMap scm(get_crc_block_size());
 415   int r = _crc_load_or_init(fd, &scm);
 416   if (r < 0)
 417     return r;
 418   scm.truncate(off);
 419   r = _crc_save(fd, &scm);
 420   return r;
 421 }
 422
 423 int GenericFileStoreBackend::_crc_update_zero(int fd, loff_t off, size_t len)
 424 {
 425   SloppyCRCMap scm(get_crc_block_size());
 426   int r = _crc_load_or_init(fd, &scm);
 427   if (r < 0)
 428     return r;
 429   scm.zero(off, len);
 430   r = _crc_save(fd, &scm);
 431   return r;
 432 }
 433
 434 int GenericFileStoreBackend::_crc_update_clone_range(int srcfd, int destfd,
 435                                                      loff_t srcoff, size_t len, loff_t dstoff)
 436 {
 437   SloppyCRCMap scm_src(get_crc_block_size());
 438   SloppyCRCMap scm_dst(get_crc_block_size());
 439   int r = _crc_load_or_init(srcfd, &scm_src);
 440   if (r < 0)
 441     return r;
 442   r = _crc_load_or_init(destfd, &scm_dst);
 443   if (r < 0)
 444     return r;
 445   ostringstream ss;
 446   scm_dst.clone_range(srcoff, len, dstoff, scm_src, &ss);
 447   dout(30) << __func__ << "\n" << ss.str() << dendl;
 448   r = _crc_save(destfd, &scm_dst);
 449   return r;
 450 }
 451
 452 int GenericFileStoreBackend::_crc_verify_read(int fd, loff_t off, size_t len, const bufferlist& bl,
 453                                               ostream *out)
 454 {
 455   SloppyCRCMap scm(get_crc_block_size());
 456   int r = _crc_load_or_init(fd, &scm);
 457   if (r < 0)
 458     return r;
 459   return scm.read(off, len, bl, out);
 460 }