fs/xfs/xfs_pnfs.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * Copyright (c) 2014 Christoph Hellwig.
   4  */
   5 #include "xfs.h"
   6 #include "xfs_shared.h"
   7 #include "xfs_format.h"
   8 #include "xfs_log_format.h"
   9 #include "xfs_trans_resv.h"
  10 #include "xfs_mount.h"
  11 #include "xfs_inode.h"
  12 #include "xfs_trans.h"
  13 #include "xfs_bmap.h"
  14 #include "xfs_iomap.h"
  15 #include "xfs_pnfs.h"
  16
  17 /*
  18  * Ensure that we do not have any outstanding pNFS layouts that can be used by
  19  * clients to directly read from or write to this inode.  This must be called
  20  * before every operation that can remove blocks from the extent map.
  21  * Additionally we call it during the write operation, where aren't concerned
  22  * about exposing unallocated blocks but just want to provide basic
  23  * synchronization between a local writer and pNFS clients.  mmap writes would
  24  * also benefit from this sort of synchronization, but due to the tricky locking
  25  * rules in the page fault path we don't bother.
  26  */
  27 int
  28 xfs_break_leased_layouts(
  29         struct inode            *inode,
  30         uint                    *iolock,
  31         bool                    *did_unlock)
  32 {
  33         struct xfs_inode        *ip = XFS_I(inode);
  34         int                     error;
  35
  36         while ((error = break_layout(inode, false)) == -EWOULDBLOCK) {
  37                 xfs_iunlock(ip, *iolock);
  38                 *did_unlock = true;
  39                 error = break_layout(inode, true);
  40                 *iolock &= ~XFS_IOLOCK_SHARED;
  41                 *iolock |= XFS_IOLOCK_EXCL;
  42                 xfs_ilock(ip, *iolock);
  43         }
  44
  45         return error;
  46 }
  47
  48 /*
  49  * Get a unique ID including its location so that the client can identify
  50  * the exported device.
  51  */
  52 int
  53 xfs_fs_get_uuid(
  54         struct super_block      *sb,
  55         u8                      *buf,
  56         u32                     *len,
  57         u64                     *offset)
  58 {
  59         struct xfs_mount        *mp = XFS_M(sb);
  60
  61         xfs_notice_once(mp,
  62 "Using experimental pNFS feature, use at your own risk!");
  63
  64         if (*len < sizeof(uuid_t))
  65                 return -EINVAL;
  66
  67         memcpy(buf, &mp->m_sb.sb_uuid, sizeof(uuid_t));
  68         *len = sizeof(uuid_t);
  69         *offset = offsetof(struct xfs_dsb, sb_uuid);
  70         return 0;
  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          * The pNFS block layout spec actually supports reflink like
 108          * functionality, but the Linux pNFS server doesn't implement it yet.
 109          */
 110         if (xfs_is_reflink_inode(ip))
 111                 return -ENXIO;
 112
 113         /*
 114          * Lock out any other I/O before we flush and invalidate the pagecache,
 115          * and then hand out a layout to the remote system.  This is very
 116          * similar to direct I/O, except that the synchronization is much more
 117          * complicated.  See the comment near xfs_break_leased_layouts
 118          * for a detailed explanation.
 119          */
 120         xfs_ilock(ip, XFS_IOLOCK_EXCL);
 121
 122         error = -EINVAL;
 123         limit = mp->m_super->s_maxbytes;
 124         if (!write)
 125                 limit = max(limit, round_up(i_size_read(inode),
 126                                      inode->i_sb->s_blocksize));
 127         if (offset > limit)
 128                 goto out_unlock;
 129         if (offset > limit - length)
 130                 length = limit - offset;
 131
 132         error = filemap_write_and_wait(inode->i_mapping);
 133         if (error)
 134                 goto out_unlock;
 135         error = invalidate_inode_pages2(inode->i_mapping);
 136         if (WARN_ON_ONCE(error))
 137                 goto out_unlock;
 138
 139         end_fsb = XFS_B_TO_FSB(mp, (xfs_ufsize_t)offset + length);
 140         offset_fsb = XFS_B_TO_FSBT(mp, offset);
 141
 142         lock_flags = xfs_ilock_data_map_shared(ip);
 143         error = xfs_bmapi_read(ip, offset_fsb, end_fsb - offset_fsb,
 144                                 &imap, &nimaps, bmapi_flags);
 145
 146         ASSERT(!nimaps || imap.br_startblock != DELAYSTARTBLOCK);
 147
 148         if (!error && write &&
 149             (!nimaps || imap.br_startblock == HOLESTARTBLOCK)) {
 150                 if (offset + length > XFS_ISIZE(ip))
 151                         end_fsb = xfs_iomap_eof_align_last_fsb(ip, end_fsb);
 152                 else if (nimaps && imap.br_startblock == HOLESTARTBLOCK)
 153                         end_fsb = min(end_fsb, imap.br_startoff +
 154                                                imap.br_blockcount);
 155                 xfs_iunlock(ip, lock_flags);
 156
 157                 error = xfs_iomap_write_direct(ip, offset_fsb,
 158                                 end_fsb - offset_fsb, &imap);
 159                 if (error)
 160                         goto out_unlock;
 161
 162                 /*
 163                  * Ensure the next transaction is committed synchronously so
 164                  * that the blocks allocated and handed out to the client are
 165                  * guaranteed to be present even after a server crash.
 166                  */
 167                 error = xfs_update_prealloc_flags(ip,
 168                                 XFS_PREALLOC_SET | XFS_PREALLOC_SYNC);
 169                 if (error)
 170                         goto out_unlock;
 171         } else {
 172                 xfs_iunlock(ip, lock_flags);
 173         }
 174         xfs_iunlock(ip, XFS_IOLOCK_EXCL);
 175
 176         error = xfs_bmbt_to_iomap(ip, iomap, &imap, 0);
 177         *device_generation = mp->m_generation;
 178         return error;
 179 out_unlock:
 180         xfs_iunlock(ip, XFS_IOLOCK_EXCL);
 181         return error;
 182 }
 183
 184 /*
 185  * Ensure the size update falls into a valid allocated block.
 186  */
 187 static int
 188 xfs_pnfs_validate_isize(
 189         struct xfs_inode        *ip,
 190         xfs_off_t               isize)
 191 {
 192         struct xfs_bmbt_irec    imap;
 193         int                     nimaps = 1;
 194         int                     error = 0;
 195
 196         xfs_ilock(ip, XFS_ILOCK_SHARED);
 197         error = xfs_bmapi_read(ip, XFS_B_TO_FSBT(ip->i_mount, isize - 1), 1,
 198                                 &imap, &nimaps, 0);
 199         xfs_iunlock(ip, XFS_ILOCK_SHARED);
 200         if (error)
 201                 return error;
 202
 203         if (imap.br_startblock == HOLESTARTBLOCK ||
 204             imap.br_startblock == DELAYSTARTBLOCK ||
 205             imap.br_state == XFS_EXT_UNWRITTEN)
 206                 return -EIO;
 207         return 0;
 208 }
 209
 210 /*
 211  * Make sure the blocks described by maps are stable on disk.  This includes
 212  * converting any unwritten extents, flushing the disk cache and updating the
 213  * time stamps.
 214  *
 215  * Note that we rely on the caller to always send us a timestamp update so that
 216  * we always commit a transaction here.  If that stops being true we will have
 217  * to manually flush the cache here similar to what the fsync code path does
 218  * for datasyncs on files that have no dirty metadata.
 219  */
 220 int
 221 xfs_fs_commit_blocks(
 222         struct inode            *inode,
 223         struct iomap            *maps,
 224         int                     nr_maps,
 225         struct iattr            *iattr)
 226 {
 227         struct xfs_inode        *ip = XFS_I(inode);
 228         struct xfs_mount        *mp = ip->i_mount;
 229         struct xfs_trans        *tp;
 230         bool                    update_isize = false;
 231         int                     error, i;
 232         loff_t                  size;
 233
 234         ASSERT(iattr->ia_valid & (ATTR_ATIME|ATTR_CTIME|ATTR_MTIME));
 235
 236         xfs_ilock(ip, XFS_IOLOCK_EXCL);
 237
 238         size = i_size_read(inode);
 239         if ((iattr->ia_valid & ATTR_SIZE) && iattr->ia_size > size) {
 240                 update_isize = true;
 241                 size = iattr->ia_size;
 242         }
 243
 244         for (i = 0; i < nr_maps; i++) {
 245                 u64 start, length, end;
 246
 247                 start = maps[i].offset;
 248                 if (start > size)
 249                         continue;
 250
 251                 end = start + maps[i].length;
 252                 if (end > size)
 253                         end = size;
 254
 255                 length = end - start;
 256                 if (!length)
 257                         continue;
 258
 259                 /*
 260                  * Make sure reads through the pagecache see the new data.
 261                  */
 262                 error = invalidate_inode_pages2_range(inode->i_mapping,
 263                                         start >> PAGE_SHIFT,
 264                                         (end - 1) >> PAGE_SHIFT);
 265                 WARN_ON_ONCE(error);
 266
 267                 error = xfs_iomap_write_unwritten(ip, start, length, false);
 268                 if (error)
 269                         goto out_drop_iolock;
 270         }
 271
 272         if (update_isize) {
 273                 error = xfs_pnfs_validate_isize(ip, size);
 274                 if (error)
 275                         goto out_drop_iolock;
 276         }
 277
 278         error = xfs_trans_alloc(mp, &M_RES(mp)->tr_ichange, 0, 0, 0, &tp);
 279         if (error)
 280                 goto out_drop_iolock;
 281
 282         xfs_ilock(ip, XFS_ILOCK_EXCL);
 283         xfs_trans_ijoin(tp, ip, XFS_ILOCK_EXCL);
 284         xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
 285
 286         xfs_setattr_time(ip, iattr);
 287         if (update_isize) {
 288                 i_size_write(inode, iattr->ia_size);
 289                 ip->i_disk_size = iattr->ia_size;
 290         }
 291
 292         xfs_trans_set_sync(tp);
 293         error = xfs_trans_commit(tp);
 294
 295 out_drop_iolock:
 296         xfs_iunlock(ip, XFS_IOLOCK_EXCL);
 297         return error;
 298 }