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  * Ensure that we do not have any outstanding pNFS layouts that can be used by
  23  * clients to directly read from or write to this inode.  This must be called
  24  * before every operation that can remove blocks from the extent map.
  25  * Additionally we call it during the write operation, where aren't concerned
  26  * about exposing unallocated blocks but just want to provide basic
  27  * synchronization between a local writer and pNFS clients.  mmap writes would
  28  * also benefit from this sort of synchronization, but due to the tricky locking
  29  * rules in the page fault path we don't bother.
  30  */
  31 int
  32 xfs_break_layouts(
  33         struct inode            *inode,
  34         uint                    *iolock)
  35 {
  36         struct xfs_inode        *ip = XFS_I(inode);
  37         int                     error;
  38
  39         ASSERT(xfs_isilocked(ip, XFS_IOLOCK_SHARED|XFS_IOLOCK_EXCL));
  40
  41         while ((error = break_layout(inode, false) == -EWOULDBLOCK)) {
  42                 xfs_iunlock(ip, *iolock);
  43                 error = break_layout(inode, true);
  44                 *iolock = XFS_IOLOCK_EXCL;
  45                 xfs_ilock(ip, *iolock);
  46         }
  47
  48         return error;
  49 }
  50
  51 /*
  52  * Get a unique ID including its location so that the client can identify
  53  * the exported device.
  54  */
  55 int
  56 xfs_fs_get_uuid(
  57         struct super_block      *sb,
  58         u8                      *buf,
  59         u32                     *len,
  60         u64                     *offset)
  61 {
  62         struct xfs_mount        *mp = XFS_M(sb);
  63
  64         printk_once(KERN_NOTICE
  65 "XFS (%s): using experimental pNFS feature, use at your own risk!\n",
  66                 mp->m_fsname);
  67
  68         if (*len < sizeof(uuid_t))
  69                 return -EINVAL;
  70
  71         memcpy(buf, &mp->m_sb.sb_uuid, sizeof(uuid_t));
  72         *len = sizeof(uuid_t);
  73         *offset = offsetof(struct xfs_dsb, sb_uuid);
  74         return 0;
  75 }
  76
  77 static void
  78 xfs_bmbt_to_iomap(
  79         struct xfs_inode        *ip,
  80         struct iomap            *iomap,
  81         struct xfs_bmbt_irec    *imap)
  82 {
  83         struct xfs_mount        *mp = ip->i_mount;
  84
  85         if (imap->br_startblock == HOLESTARTBLOCK) {
  86                 iomap->blkno = IOMAP_NULL_BLOCK;
  87                 iomap->type = IOMAP_HOLE;
  88         } else if (imap->br_startblock == DELAYSTARTBLOCK) {
  89                 iomap->blkno = IOMAP_NULL_BLOCK;
  90                 iomap->type = IOMAP_DELALLOC;
  91         } else {
  92                 iomap->blkno =
  93                         XFS_FSB_TO_DADDR(ip->i_mount, imap->br_startblock);
  94                 if (imap->br_state == XFS_EXT_UNWRITTEN)
  95                         iomap->type = IOMAP_UNWRITTEN;
  96                 else
  97                         iomap->type = IOMAP_MAPPED;
  98         }
  99         iomap->offset = XFS_FSB_TO_B(mp, imap->br_startoff);
 100         iomap->length = XFS_FSB_TO_B(mp, imap->br_blockcount);
 101 }
 102
 103 /*
 104  * Get a layout for the pNFS client.
 105  */
 106 int
 107 xfs_fs_map_blocks(
 108         struct inode            *inode,
 109         loff_t                  offset,
 110         u64                     length,
 111         struct iomap            *iomap,
 112         bool                    write,
 113         u32                     *device_generation)
 114 {
 115         struct xfs_inode        *ip = XFS_I(inode);
 116         struct xfs_mount        *mp = ip->i_mount;
 117         struct xfs_bmbt_irec    imap;
 118         xfs_fileoff_t           offset_fsb, end_fsb;
 119         loff_t                  limit;
 120         int                     bmapi_flags = XFS_BMAPI_ENTIRE;
 121         int                     nimaps = 1;
 122         uint                    lock_flags;
 123         int                     error = 0;
 124
 125         if (XFS_FORCED_SHUTDOWN(mp))
 126                 return -EIO;
 127
 128         /*
 129          * We can't export inodes residing on the realtime device.  The realtime
 130          * device doesn't have a UUID to identify it, so the client has no way
 131          * to find it.
 132          */
 133         if (XFS_IS_REALTIME_INODE(ip))
 134                 return -ENXIO;
 135
 136         /*
 137          * Lock out any other I/O before we flush and invalidate the pagecache,
 138          * and then hand out a layout to the remote system.  This is very
 139          * similar to direct I/O, except that the synchronization is much more
 140          * complicated.  See the comment near xfs_break_layouts for a detailed
 141          * explanation.
 142          */
 143         xfs_ilock(ip, XFS_IOLOCK_EXCL);
 144
 145         error = -EINVAL;
 146         limit = mp->m_super->s_maxbytes;
 147         if (!write)
 148                 limit = max(limit, round_up(i_size_read(inode),
 149                                      inode->i_sb->s_blocksize));
 150         if (offset > limit)
 151                 goto out_unlock;
 152         if (offset > limit - length)
 153                 length = limit - offset;
 154
 155         error = filemap_write_and_wait(inode->i_mapping);
 156         if (error)
 157                 goto out_unlock;
 158         error = invalidate_inode_pages2(inode->i_mapping);
 159         if (WARN_ON_ONCE(error))
 160                 return error;
 161
 162         end_fsb = XFS_B_TO_FSB(mp, (xfs_ufsize_t)offset + length);
 163         offset_fsb = XFS_B_TO_FSBT(mp, offset);
 164
 165         lock_flags = xfs_ilock_data_map_shared(ip);
 166         error = xfs_bmapi_read(ip, offset_fsb, end_fsb - offset_fsb,
 167                                 &imap, &nimaps, bmapi_flags);
 168         xfs_iunlock(ip, lock_flags);
 169
 170         if (error)
 171                 goto out_unlock;
 172
 173         if (write) {
 174                 enum xfs_prealloc_flags flags = 0;
 175
 176                 ASSERT(imap.br_startblock != DELAYSTARTBLOCK);
 177
 178                 if (!nimaps || imap.br_startblock == HOLESTARTBLOCK) {
 179                         error = xfs_iomap_write_direct(ip, offset, length,
 180                                                        &imap, nimaps);
 181                         if (error)
 182                                 goto out_unlock;
 183
 184                         /*
 185                          * Ensure the next transaction is committed
 186                          * synchronously so that the blocks allocated and
 187                          * handed out to the client are guaranteed to be
 188                          * present even after a server crash.
 189                          */
 190                         flags |= XFS_PREALLOC_SET | XFS_PREALLOC_SYNC;
 191                 }
 192
 193                 error = xfs_update_prealloc_flags(ip, flags);
 194                 if (error)
 195                         goto out_unlock;
 196         }
 197         xfs_iunlock(ip, XFS_IOLOCK_EXCL);
 198
 199         xfs_bmbt_to_iomap(ip, iomap, &imap);
 200         *device_generation = mp->m_generation;
 201         return error;
 202 out_unlock:
 203         xfs_iunlock(ip, XFS_IOLOCK_EXCL);
 204         return error;
 205 }
 206
 207 /*
 208  * Ensure the size update falls into a valid allocated block.
 209  */
 210 static int
 211 xfs_pnfs_validate_isize(
 212         struct xfs_inode        *ip,
 213         xfs_off_t               isize)
 214 {
 215         struct xfs_bmbt_irec    imap;
 216         int                     nimaps = 1;
 217         int                     error = 0;
 218
 219         xfs_ilock(ip, XFS_ILOCK_SHARED);
 220         error = xfs_bmapi_read(ip, XFS_B_TO_FSBT(ip->i_mount, isize - 1), 1,
 221                                 &imap, &nimaps, 0);
 222         xfs_iunlock(ip, XFS_ILOCK_SHARED);
 223         if (error)
 224                 return error;
 225
 226         if (imap.br_startblock == HOLESTARTBLOCK ||
 227             imap.br_startblock == DELAYSTARTBLOCK ||
 228             imap.br_state == XFS_EXT_UNWRITTEN)
 229                 return -EIO;
 230         return 0;
 231 }
 232
 233 /*
 234  * Make sure the blocks described by maps are stable on disk.  This includes
 235  * converting any unwritten extents, flushing the disk cache and updating the
 236  * time stamps.
 237  *
 238  * Note that we rely on the caller to always send us a timestamp update so that
 239  * we always commit a transaction here.  If that stops being true we will have
 240  * to manually flush the cache here similar to what the fsync code path does
 241  * for datasyncs on files that have no dirty metadata.
 242  */
 243 int
 244 xfs_fs_commit_blocks(
 245         struct inode            *inode,
 246         struct iomap            *maps,
 247         int                     nr_maps,
 248         struct iattr            *iattr)
 249 {
 250         struct xfs_inode        *ip = XFS_I(inode);
 251         struct xfs_mount        *mp = ip->i_mount;
 252         struct xfs_trans        *tp;
 253         bool                    update_isize = false;
 254         int                     error, i;
 255         loff_t                  size;
 256
 257         ASSERT(iattr->ia_valid & (ATTR_ATIME|ATTR_CTIME|ATTR_MTIME));
 258
 259         xfs_ilock(ip, XFS_IOLOCK_EXCL);
 260
 261         size = i_size_read(inode);
 262         if ((iattr->ia_valid & ATTR_SIZE) && iattr->ia_size > size) {
 263                 update_isize = true;
 264                 size = iattr->ia_size;
 265         }
 266
 267         for (i = 0; i < nr_maps; i++) {
 268                 u64 start, length, end;
 269
 270                 start = maps[i].offset;
 271                 if (start > size)
 272                         continue;
 273
 274                 end = start + maps[i].length;
 275                 if (end > size)
 276                         end = size;
 277
 278                 length = end - start;
 279                 if (!length)
 280                         continue;
 281
 282                 /*
 283                  * Make sure reads through the pagecache see the new data.
 284                  */
 285                 error = invalidate_inode_pages2_range(inode->i_mapping,
 286                                         start >> PAGE_CACHE_SHIFT,
 287                                         (end - 1) >> PAGE_CACHE_SHIFT);
 288                 WARN_ON_ONCE(error);
 289
 290                 error = xfs_iomap_write_unwritten(ip, start, length);
 291                 if (error)
 292                         goto out_drop_iolock;
 293         }
 294
 295         if (update_isize) {
 296                 error = xfs_pnfs_validate_isize(ip, size);
 297                 if (error)
 298                         goto out_drop_iolock;
 299         }
 300
 301         tp = xfs_trans_alloc(mp, XFS_TRANS_SETATTR_NOT_SIZE);
 302         error = xfs_trans_reserve(tp, &M_RES(mp)->tr_ichange, 0, 0);
 303         if (error) {
 304                 xfs_trans_cancel(tp, 0);
 305                 goto out_drop_iolock;
 306         }
 307
 308         xfs_ilock(ip, XFS_ILOCK_EXCL);
 309         xfs_trans_ijoin(tp, ip, XFS_ILOCK_EXCL);
 310         xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
 311
 312         xfs_setattr_time(ip, iattr);
 313         if (update_isize) {
 314                 i_size_write(inode, iattr->ia_size);
 315                 ip->i_d.di_size = iattr->ia_size;
 316         }
 317
 318         xfs_trans_set_sync(tp);
 319         error = xfs_trans_commit(tp, 0);
 320
 321 out_drop_iolock:
 322         xfs_iunlock(ip, XFS_IOLOCK_EXCL);
 323         return error;
 324 }