ceph: don't early drop Fw cap
authorYan, Zheng <zheng.z.yan@intel.com>
Fri, 1 Mar 2013 02:55:39 +0000 (10:55 +0800)
committerSage Weil <sage@inktank.com>
Thu, 2 May 2013 04:16:10 +0000 (21:16 -0700)
ceph_aio_write() has an optimization that marks CEPH_CAP_FILE_WR
cap dirty before data is copied to page cache and inode size is
updated. The optimization avoids slow cap revocation caused by
balance_dirty_pages(), but introduces inode size update race. If
ceph_check_caps() flushes the dirty cap before the inode size is
updated, MDS can miss the new inode size. So just remove the
optimization.

Signed-off-by: Yan, Zheng <zheng.z.yan@intel.com>
Reviewed-by: Greg Farnum <greg@inktank.com>
fs/ceph/file.c

index b86d2a0eb145bd0cb1504d1f94ff7d8f397d49ed..3d1aefeee948fb8bd15f28491cc0dc193e343ef9 100644 (file)
@@ -724,9 +724,12 @@ static ssize_t ceph_aio_write(struct kiocb *iocb, const struct iovec *iov,
        if (ceph_snap(inode) != CEPH_NOSNAP)
                return -EROFS;
 
+       sb_start_write(inode->i_sb);
 retry_snap:
-       if (ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_FULL))
-               return -ENOSPC;
+       if (ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_FULL)) {
+               ret = -ENOSPC;
+               goto out;
+       }
        __ceph_do_pending_vmtruncate(inode);
        dout("aio_write %p %llx.%llx %llu~%u getting caps. i_size %llu\n",
             inode, ceph_vinop(inode), pos, (unsigned)iov->iov_len,
@@ -750,29 +753,10 @@ retry_snap:
                ret = ceph_sync_write(file, iov->iov_base, iov->iov_len,
                        &iocb->ki_pos);
        } else {
-               /*
-                * buffered write; drop Fw early to avoid slow
-                * revocation if we get stuck on balance_dirty_pages
-                */
-               int dirty;
-
-               spin_lock(&ci->i_ceph_lock);
-               dirty = __ceph_mark_dirty_caps(ci, CEPH_CAP_FILE_WR);
-               spin_unlock(&ci->i_ceph_lock);
-               ceph_put_cap_refs(ci, got);
-
-               ret = generic_file_aio_write(iocb, iov, nr_segs, pos);
-               if ((ret >= 0 || ret == -EIOCBQUEUED) &&
-                   ((file->f_flags & O_SYNC) || IS_SYNC(file->f_mapping->host)
-                    || ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_NEARFULL))) {
-                       err = vfs_fsync_range(file, pos, pos + ret - 1, 1);
-                       if (err < 0)
-                               ret = err;
-               }
-
-               if (dirty)
-                       __mark_inode_dirty(inode, dirty);
-               goto out;
+               mutex_lock(&inode->i_mutex);
+               ret = __generic_file_aio_write(iocb, iov, nr_segs,
+                                              &iocb->ki_pos);
+               mutex_unlock(&inode->i_mutex);
        }
 
        if (ret >= 0) {
@@ -790,12 +774,20 @@ out_put:
             ceph_cap_string(got));
        ceph_put_cap_refs(ci, got);
 
+       if (ret >= 0 &&
+           ((file->f_flags & O_SYNC) || IS_SYNC(file->f_mapping->host) ||
+            ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_NEARFULL))) {
+               err = vfs_fsync_range(file, pos, pos + ret - 1, 1);
+               if (err < 0)
+                       ret = err;
+       }
 out:
        if (ret == -EOLDSNAPC) {
                dout("aio_write %p %llx.%llx %llu~%u got EOLDSNAPC, retrying\n",
                     inode, ceph_vinop(inode), pos, (unsigned)iov->iov_len);
                goto retry_snap;
        }
+       sb_end_write(inode->i_sb);
 
        return ret;
 }