fs/xfs/xfs_pnfs.c

   1 /*
   2  * Copyright (c) 2014 Christoph Hellwig.
   3  */
   4 #include "xfs.h"
   5 #include "xfs_format.h"
   6 #include "xfs_log_format.h"
   7 #include "xfs_trans_resv.h"
   8 #include "xfs_sb.h"
   9 #include "xfs_mount.h"
  10 #include "xfs_inode.h"
  11 #include "xfs_trans.h"
  12 #include "xfs_log.h"
  13 #include "xfs_bmap.h"
  14 #include "xfs_bmap_util.h"
  15 #include "xfs_error.h"
  16 #include "xfs_iomap.h"
  17 #include "xfs_shared.h"
  18 #include "xfs_bit.h"
  19 #include "xfs_pnfs.h"
  20
  21 /*
  22  * Get a unique ID including its location so that the client can identify
  23  * the exported device.
  24  */
  25 int
  26 xfs_fs_get_uuid(
  27         struct super_block      *sb,
  28         u8                      *buf,
  29         u32                     *len,
  30         u64                     *offset)
  31 {
  32         struct xfs_mount        *mp = XFS_M(sb);
  33
  34         printk_once(KERN_NOTICE
  35 "XFS (%s): using experimental pNFS feature, use at your own risk!\n",
  36                 mp->m_fsname);
  37
  38         if (*len < sizeof(uuid_t))
  39                 return -EINVAL;
  40
  41         memcpy(buf, &mp->m_sb.sb_uuid, sizeof(uuid_t));
  42         *len = sizeof(uuid_t);
  43         *offset = offsetof(struct xfs_dsb, sb_uuid);
  44         return 0;
  45 }
  46
  47 static void
  48 xfs_bmbt_to_iomap(
  49         struct xfs_inode        *ip,
  50         struct iomap            *iomap,
  51         struct xfs_bmbt_irec    *imap)
  52 {
  53         struct xfs_mount        *mp = ip->i_mount;
  54
  55         if (imap->br_startblock == HOLESTARTBLOCK) {
  56                 iomap->blkno = IOMAP_NULL_BLOCK;
  57                 iomap->type = IOMAP_HOLE;
  58         } else if (imap->br_startblock == DELAYSTARTBLOCK) {
  59                 iomap->blkno = IOMAP_NULL_BLOCK;
  60                 iomap->type = IOMAP_DELALLOC;
  61         } else {
  62                 iomap->blkno =
  63                         XFS_FSB_TO_DADDR(ip->i_mount, imap->br_startblock);
  64                 if (imap->br_state == XFS_EXT_UNWRITTEN)
  65                         iomap->type = IOMAP_UNWRITTEN;
  66                 else
  67                         iomap->type = IOMAP_MAPPED;
  68         }
  69         iomap->offset = XFS_FSB_TO_B(mp, imap->br_startoff);
  70         iomap->length = XFS_FSB_TO_B(mp, imap->br_blockcount);
  71 }
  72
  73 /*
  74  * Get a layout for the pNFS client.
  75  */
  76 int
  77 xfs_fs_map_blocks(
  78         struct inode            *inode,
  79         loff_t                  offset,
  80         u64                     length,
  81         struct iomap            *iomap,
  82         bool                    write,
  83         u32                     *device_generation)
  84 {
  85         struct xfs_inode        *ip = XFS_I(inode);
  86         struct xfs_mount        *mp = ip->i_mount;
  87         struct xfs_bmbt_irec    imap;
  88         xfs_fileoff_t           offset_fsb, end_fsb;
  89         loff_t                  limit;
  90         int                     bmapi_flags = XFS_BMAPI_ENTIRE;
  91         int                     nimaps = 1;
  92         uint                    lock_flags;
  93         int                     error = 0;
  94
  95         if (XFS_FORCED_SHUTDOWN(mp))
  96                 return -EIO;
  97
  98         /*
  99          * We can't export inodes residing on the realtime device.  The realtime
 100          * device doesn't have a UUID to identify it, so the client has no way
 101          * to find it.
 102          */
 103         if (XFS_IS_REALTIME_INODE(ip))
 104                 return -ENXIO;
 105
 106         /*
 107          * Lock out any other I/O before we flush and invalidate the pagecache,
 108          * and then hand out a layout to the remote system.  This is very
 109          * similar to direct I/O, except that the synchronization is much more
 110          * complicated.  See the comment near xfs_break_layouts for a detailed
 111          * explanation.
 112          */
 113         xfs_ilock(ip, XFS_IOLOCK_EXCL);
 114
 115         error = -EINVAL;
 116         limit = mp->m_super->s_maxbytes;
 117         if (!write)
 118                 limit = max(limit, round_up(i_size_read(inode),
 119                                      inode->i_sb->s_blocksize));
 120         if (offset > limit)
 121                 goto out_unlock;
 122         if (offset > limit - length)
 123                 length = limit - offset;
 124
 125         error = filemap_write_and_wait(inode->i_mapping);
 126         if (error)
 127                 goto out_unlock;
 128         error = invalidate_inode_pages2(inode->i_mapping);
 129         if (WARN_ON_ONCE(error))
 130                 return error;
 131
 132         end_fsb = XFS_B_TO_FSB(mp, (xfs_ufsize_t)offset + length);
 133         offset_fsb = XFS_B_TO_FSBT(mp, offset);
 134
 135         lock_flags = xfs_ilock_data_map_shared(ip);
 136         error = xfs_bmapi_read(ip, offset_fsb, end_fsb - offset_fsb,
 137                                 &imap, &nimaps, bmapi_flags);
 138         xfs_iunlock(ip, lock_flags);
 139
 140         if (error)
 141                 goto out_unlock;
 142
 143         if (write) {
 144                 enum xfs_prealloc_flags flags = 0;
 145
 146                 ASSERT(imap.br_startblock != DELAYSTARTBLOCK);
 147
 148                 if (!nimaps || imap.br_startblock == HOLESTARTBLOCK) {
 149                         error = xfs_iomap_write_direct(ip, offset, length,
 150                                                        &imap, nimaps);
 151                         if (error)
 152                                 goto out_unlock;
 153
 154                         /*
 155                          * Ensure the next transaction is committed
 156                          * synchronously so that the blocks allocated and
 157                          * handed out to the client are guaranteed to be
 158                          * present even after a server crash.
 159                          */
 160                         flags |= XFS_PREALLOC_SET | XFS_PREALLOC_SYNC;
 161                 }
 162
 163                 error = xfs_update_prealloc_flags(ip, flags);
 164                 if (error)
 165                         goto out_unlock;
 166         }
 167         xfs_iunlock(ip, XFS_IOLOCK_EXCL);
 168
 169         xfs_bmbt_to_iomap(ip, iomap, &imap);
 170         *device_generation = mp->m_generation;
 171         return error;
 172 out_unlock:
 173         xfs_iunlock(ip, XFS_IOLOCK_EXCL);
 174         return error;
 175 }
 176
 177 /*
 178  * Ensure the size update falls into a valid allocated block.
 179  */
 180 static int
 181 xfs_pnfs_validate_isize(
 182         struct xfs_inode        *ip,
 183         xfs_off_t               isize)
 184 {
 185         struct xfs_bmbt_irec    imap;
 186         int                     nimaps = 1;
 187         int                     error = 0;
 188
 189         xfs_ilock(ip, XFS_ILOCK_SHARED);
 190         error = xfs_bmapi_read(ip, XFS_B_TO_FSBT(ip->i_mount, isize - 1), 1,
 191                                 &imap, &nimaps, 0);
 192         xfs_iunlock(ip, XFS_ILOCK_SHARED);
 193         if (error)
 194                 return error;
 195
 196         if (imap.br_startblock == HOLESTARTBLOCK ||
 197             imap.br_startblock == DELAYSTARTBLOCK ||
 198             imap.br_state == XFS_EXT_UNWRITTEN)
 199                 return -EIO;
 200         return 0;
 201 }
 202
 203 /*
 204  * Make sure the blocks described by maps are stable on disk.  This includes
 205  * converting any unwritten extents, flushing the disk cache and updating the
 206  * time stamps.
 207  *
 208  * Note that we rely on the caller to always send us a timestamp update so that
 209  * we always commit a transaction here.  If that stops being true we will have
 210  * to manually flush the cache here similar to what the fsync code path does
 211  * for datasyncs on files that have no dirty metadata.
 212  */
 213 int
 214 xfs_fs_commit_blocks(
 215         struct inode            *inode,
 216         struct iomap            *maps,
 217         int                     nr_maps,
 218         struct iattr            *iattr)
 219 {
 220         struct xfs_inode        *ip = XFS_I(inode);
 221         struct xfs_mount        *mp = ip->i_mount;
 222         struct xfs_trans        *tp;
 223         bool                    update_isize = false;
 224         int                     error, i;
 225         loff_t                  size;
 226
 227         ASSERT(iattr->ia_valid & (ATTR_ATIME|ATTR_CTIME|ATTR_MTIME));
 228
 229         xfs_ilock(ip, XFS_IOLOCK_EXCL);
 230
 231         size = i_size_read(inode);
 232         if ((iattr->ia_valid & ATTR_SIZE) && iattr->ia_size > size) {
 233                 update_isize = true;
 234                 size = iattr->ia_size;
 235         }
 236
 237         for (i = 0; i < nr_maps; i++) {
 238                 u64 start, length, end;
 239
 240                 start = maps[i].offset;
 241                 if (start > size)
 242                         continue;
 243
 244                 end = start + maps[i].length;
 245                 if (end > size)
 246                         end = size;
 247
 248                 length = end - start;
 249                 if (!length)
 250                         continue;
 251
 252                 /*
 253                  * Make sure reads through the pagecache see the new data.
 254                  */
 255                 error = invalidate_inode_pages2_range(inode->i_mapping,
 256                                         start >> PAGE_CACHE_SHIFT,
 257                                         (end - 1) >> PAGE_CACHE_SHIFT);
 258                 WARN_ON_ONCE(error);
 259
 260                 error = xfs_iomap_write_unwritten(ip, start, length);
 261                 if (error)
 262                         goto out_drop_iolock;
 263         }
 264
 265         if (update_isize) {
 266                 error = xfs_pnfs_validate_isize(ip, size);
 267                 if (error)
 268                         goto out_drop_iolock;
 269         }
 270
 271         tp = xfs_trans_alloc(mp, XFS_TRANS_SETATTR_NOT_SIZE);
 272         error = xfs_trans_reserve(tp, &M_RES(mp)->tr_ichange, 0, 0);
 273         if (error)
 274                 goto out_drop_iolock;
 275
 276         xfs_ilock(ip, XFS_ILOCK_EXCL);
 277         xfs_trans_ijoin(tp, ip, XFS_ILOCK_EXCL);
 278         xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
 279
 280         xfs_setattr_time(ip, iattr);
 281         if (update_isize) {
 282                 i_size_write(inode, iattr->ia_size);
 283                 ip->i_d.di_size = iattr->ia_size;
 284         }
 285
 286         xfs_trans_set_sync(tp);
 287         error = xfs_trans_commit(tp, 0);
 288
 289 out_drop_iolock:
 290         xfs_iunlock(ip, XFS_IOLOCK_EXCL);
 291         return error;
 292 }