xfs_vfsops.c

来自「Linux Kernel 2.6.9 for OMAP1710」· C语言 代码 · 共 1,927 行 · 第 1/4 页

#define XFS_PREEMPT_MASK	0x7f

	if (bypassed)
		*bypassed = 0;
	if (XFS_MTOVFS(mp)->vfs_flag & VFS_RDONLY)
		return 0;
	error = 0;
	last_error = 0;
	preempt = 0;

	/* Allocate a reference marker */
	ipointer = (xfs_iptr_t *)kmem_zalloc(sizeof(xfs_iptr_t), KM_SLEEP);

	fflag = XFS_B_ASYNC;		/* default is don't wait */
	if (flags & SYNC_BDFLUSH)
		fflag = XFS_B_DELWRI;
	if (flags & SYNC_WAIT)
		fflag = 0;		/* synchronous overrides all */

	base_lock_flags = XFS_ILOCK_SHARED;
	if (flags & (SYNC_DELWRI | SYNC_CLOSE)) {
		/*
		 * We need the I/O lock if we're going to call any of
		 * the flush/inval routines.
		 */
		base_lock_flags |= XFS_IOLOCK_SHARED;
	}

	XFS_MOUNT_ILOCK(mp);

	ip = mp->m_inodes;

	mount_locked = B_TRUE;
	vnode_refed  = B_FALSE;

	IPOINTER_CLR;

	do {
		ASSERT(ipointer_in == B_FALSE);
		ASSERT(vnode_refed == B_FALSE);

		lock_flags = base_lock_flags;

		/*
		 * There were no inodes in the list, just break out
		 * of the loop.
		 */
		if (ip == NULL) {
			break;
		}

		/*
		 * We found another sync thread marker - skip it
		 */
		if (ip->i_mount == NULL) {
			ip = ip->i_mnext;
			continue;
		}

		vp = XFS_ITOV_NULL(ip);

		/*
		 * If the vnode is gone then this is being torn down,
		 * call reclaim if it is flushed, else let regular flush
		 * code deal with it later in the loop.
		 */

		if (vp == NULL) {
			/* Skip ones already in reclaim */
			if (ip->i_flags & XFS_IRECLAIM) {
				ip = ip->i_mnext;
				continue;
			}
			if (xfs_ilock_nowait(ip, XFS_ILOCK_EXCL) == 0) {
				ip = ip->i_mnext;
			} else if ((xfs_ipincount(ip) == 0) &&
				    xfs_iflock_nowait(ip)) {
				IPOINTER_INSERT(ip, mp);

				xfs_finish_reclaim(ip, 1,
						XFS_IFLUSH_DELWRI_ELSE_ASYNC);

				XFS_MOUNT_ILOCK(mp);
				mount_locked = B_TRUE;
				IPOINTER_REMOVE(ip, mp);
			} else {
				xfs_iunlock(ip, XFS_ILOCK_EXCL);
				ip = ip->i_mnext;
			}
			continue;
		}

		if (VN_BAD(vp)) {
			ip = ip->i_mnext;
			continue;
		}

		if (XFS_FORCED_SHUTDOWN(mp) && !(flags & SYNC_CLOSE)) {
			XFS_MOUNT_IUNLOCK(mp);
			kmem_free(ipointer, sizeof(xfs_iptr_t));
			return 0;
		}

		/*
		 * If this is just vfs_sync() or pflushd() calling
		 * then we can skip inodes for which it looks like
		 * there is nothing to do.  Since we don't have the
		 * inode locked this is racey, but these are periodic
		 * calls so it doesn't matter.  For the others we want
		 * to know for sure, so we at least try to lock them.
		 */
		if (flags & SYNC_BDFLUSH) {
			if (((ip->i_itemp == NULL) ||
			     !(ip->i_itemp->ili_format.ilf_fields &
			       XFS_ILOG_ALL)) &&
			    (ip->i_update_core == 0)) {
				ip = ip->i_mnext;
				continue;
			}
		}

		/*
		 * Try to lock without sleeping.  We're out of order with
		 * the inode list lock here, so if we fail we need to drop
		 * the mount lock and try again.  If we're called from
		 * bdflush() here, then don't bother.
		 *
		 * The inode lock here actually coordinates with the
		 * almost spurious inode lock in xfs_ireclaim() to prevent
		 * the vnode we handle here without a reference from
		 * being freed while we reference it.  If we lock the inode
		 * while it's on the mount list here, then the spurious inode
		 * lock in xfs_ireclaim() after the inode is pulled from
		 * the mount list will sleep until we release it here.
		 * This keeps the vnode from being freed while we reference
		 * it.  It is also cheaper and simpler than actually doing
		 * a vn_get() for every inode we touch here.
		 */
		if (xfs_ilock_nowait(ip, lock_flags) == 0) {

			if ((flags & SYNC_BDFLUSH) || (vp == NULL)) {
				ip = ip->i_mnext;
				continue;
			}

			/*
			 * We need to unlock the inode list lock in order
			 * to lock the inode. Insert a marker record into
			 * the inode list to remember our position, dropping
			 * the lock is now done inside the IPOINTER_INSERT
			 * macro.
			 *
			 * We also use the inode list lock to protect us
			 * in taking a snapshot of the vnode version number
			 * for use in calling vn_get().
			 */
			VMAP(vp, vmap);
			IPOINTER_INSERT(ip, mp);

			vp = vn_get(vp, &vmap);
			if (vp == NULL) {
				/*
				 * The vnode was reclaimed once we let go
				 * of the inode list lock.  Skip to the
				 * next list entry. Remove the marker.
				 */

				XFS_MOUNT_ILOCK(mp);

				mount_locked = B_TRUE;
				vnode_refed  = B_FALSE;

				IPOINTER_REMOVE(ip, mp);

				continue;
			}

			xfs_ilock(ip, lock_flags);

			ASSERT(vp == XFS_ITOV(ip));
			ASSERT(ip->i_mount == mp);

			vnode_refed = B_TRUE;
		}

		/* From here on in the loop we may have a marker record
		 * in the inode list.
		 */

		if ((flags & SYNC_CLOSE)  && (vp != NULL)) {
			/*
			 * This is the shutdown case.  We just need to
			 * flush and invalidate all the pages associated
			 * with the inode.  Drop the inode lock since
			 * we can't hold it across calls to the buffer
			 * cache.
			 *
			 * We don't set the VREMAPPING bit in the vnode
			 * here, because we don't hold the vnode lock
			 * exclusively.  It doesn't really matter, though,
			 * because we only come here when we're shutting
			 * down anyway.
			 */
			xfs_iunlock(ip, XFS_ILOCK_SHARED);

			if (XFS_FORCED_SHUTDOWN(mp)) {
				VOP_TOSS_PAGES(vp, 0, -1, FI_REMAPF);
			} else {
				VOP_FLUSHINVAL_PAGES(vp, 0, -1, FI_REMAPF);
			}

			xfs_ilock(ip, XFS_ILOCK_SHARED);

		} else if ((flags & SYNC_DELWRI) && (vp != NULL)) {
			if (VN_DIRTY(vp)) {
				/* We need to have dropped the lock here,
				 * so insert a marker if we have not already
				 * done so.
				 */
				if (mount_locked) {
					IPOINTER_INSERT(ip, mp);
				}

				/*
				 * Drop the inode lock since we can't hold it
				 * across calls to the buffer cache.
				 */
				xfs_iunlock(ip, XFS_ILOCK_SHARED);
				VOP_FLUSH_PAGES(vp, (xfs_off_t)0, -1,
							fflag, FI_NONE, error);
				xfs_ilock(ip, XFS_ILOCK_SHARED);
			}

		}

		if (flags & SYNC_BDFLUSH) {
			if ((flags & SYNC_ATTR) &&
			    ((ip->i_update_core) ||
			     ((ip->i_itemp != NULL) &&
			      (ip->i_itemp->ili_format.ilf_fields != 0)))) {

				/* Insert marker and drop lock if not already
				 * done.
				 */
				if (mount_locked) {
					IPOINTER_INSERT(ip, mp);
				}

				/*
				 * We don't want the periodic flushing of the
				 * inodes by vfs_sync() to interfere with
				 * I/O to the file, especially read I/O
				 * where it is only the access time stamp
				 * that is being flushed out.  To prevent
				 * long periods where we have both inode
				 * locks held shared here while reading the
				 * inode's buffer in from disk, we drop the
				 * inode lock while reading in the inode
				 * buffer.  We have to release the buffer
				 * and reacquire the inode lock so that they
				 * are acquired in the proper order (inode
				 * locks first).  The buffer will go at the
				 * end of the lru chain, though, so we can
				 * expect it to still be there when we go
				 * for it again in xfs_iflush().
				 */
				if ((xfs_ipincount(ip) == 0) &&
				    xfs_iflock_nowait(ip)) {

					xfs_ifunlock(ip);
					xfs_iunlock(ip, XFS_ILOCK_SHARED);

					error = xfs_itobp(mp, NULL, ip,
							  &dip, &bp, 0);
					if (!error) {
						xfs_buf_relse(bp);
					} else {
						/* Bailing out, remove the
						 * marker and free it.
						 */
						XFS_MOUNT_ILOCK(mp);

						IPOINTER_REMOVE(ip, mp);

						XFS_MOUNT_IUNLOCK(mp);

						ASSERT(!(lock_flags &
							XFS_IOLOCK_SHARED));

						kmem_free(ipointer,
							sizeof(xfs_iptr_t));
						return (0);
					}

					/*
					 * Since we dropped the inode lock,
					 * the inode may have been reclaimed.
					 * Therefore, we reacquire the mount
					 * lock and check to see if we were the
					 * inode reclaimed. If this happened
					 * then the ipointer marker will no
					 * longer point back at us. In this
					 * case, move ip along to the inode
					 * after the marker, remove the marker
					 * and continue.
					 */
					XFS_MOUNT_ILOCK(mp);
					mount_locked = B_TRUE;

					if (ip != ipointer->ip_mprev) {
						IPOINTER_REMOVE(ip, mp);

						ASSERT(!vnode_refed);
						ASSERT(!(lock_flags &
							XFS_IOLOCK_SHARED));
						continue;
					}

					ASSERT(ip->i_mount == mp);

					if (xfs_ilock_nowait(ip,
						    XFS_ILOCK_SHARED) == 0) {
						ASSERT(ip->i_mount == mp);
						/*
						 * We failed to reacquire
						 * the inode lock without
						 * sleeping, so just skip
						 * the inode for now.  We
						 * clear the ILOCK bit from
						 * the lock_flags so that we
						 * won't try to drop a lock
						 * we don't hold below.
						 */
						lock_flags &= ~XFS_ILOCK_SHARED;
						IPOINTER_REMOVE(ip_next, mp);
					} else if ((xfs_ipincount(ip) == 0) &&
						   xfs_iflock_nowait(ip)) {
						ASSERT(ip->i_mount == mp);
						/*
						 * Since this is vfs_sync()
						 * calling we only flush the
						 * inode out if we can lock
						 * it without sleeping and
						 * it is not pinned.  Drop
						 * the mount lock here so
						 * that we don't hold it for
						 * too long. We already have
						 * a marker in the list here.
						 */
						XFS_MOUNT_IUNLOCK(mp);
						mount_locked = B_FALSE;
						error = xfs_iflush(ip,
							   XFS_IFLUSH_DELWRI);
					} else {
						ASSERT(ip->i_mount == mp);
						IPOINTER_REMOVE(ip_next, mp);
					}
				}

			}

		} else {
			if ((flags & SYNC_ATTR) &&
			    ((ip->i_update_core) ||
			     ((ip->i_itemp != NULL) &&
			      (ip->i_itemp->ili_format.ilf_fields != 0)))) {
				if (mount_locked) {
					IPOINTER_INSERT(ip, mp);
				}

				if (flags & SYNC_WAIT) {
					xfs_iflock(ip);
					error = xfs_iflush(ip,
							   XFS_IFLUSH_SYNC);
				} else {
					/*
					 * If we can't acquire the flush
					 * lock, then the inode is already
					 * being flushed so don't bother
					 * waiting.  If we can lock it then
					 * do a delwri flush so we can
					 * combine multiple inode flushes
					 * in each disk write.
					 */
					if (xfs_iflock_nowait(ip)) {
						error = xfs_iflush(ip,
							   XFS_IFLUSH_DELWRI);
					}
					else if (bypassed)
						(*bypassed)++;
				}
			}
		}

		if (lock_flags != 0) {
			xfs_iunlock(ip, lock_flags);
		}

		if (vnode_refed) {
			/*
			 * If we had to take a reference on the vnode
			 * above, then wait until after we've unlocked
			 * the inode to release the reference.  This is
			 * because we can be already holding the inode
			 * lock when VN_RELE() calls xfs_inactive().
			 *
			 * Make sure to drop the mount lock before calling
			 * VN_RELE() so that we don't trip over ourselves if
			 * we have to go for the mount lock again in the
			 * inactive code.
			 */
			if (mount_locked) {
				IPOINTER_INSERT(ip, mp);
			}

			VN_RELE(vp);

			vnode_refed = B_FALSE;
		}

		if (error) {
			last_error = error;
		}

		/*
		 * bail out if the filesystem is corrupted.
		 */
		if (error == EFSCORRUPTED)  {
			if (!mount_locked) {
				XFS_MOUNT_ILOCK(mp);
				IPOINTER_REMOVE(ip, mp);
			}
			XFS_MOUNT_IUNLOCK(mp);
			ASSERT(ipointer_in == B_FALSE);
			kmem_free(ipointer, sizeof(xfs_iptr_t));
			return XFS_ERROR(error);
		}

		/* Let other threads have a chance at the mount lock
		 * if we have looped many times without dropping the
		 * lock.
		 */
		if ((++preempt & XFS_PREEMPT_MASK) == 0) {
			if (mount_locked) {
				IPOINTER_INSERT(ip, mp);
			}
		}

		if (mount_locked == B_FALSE) {
			XFS_MOUNT_ILOCK(mp);
			mount_locked = B_TRUE;
			IPOINTER_REMOVE(ip, mp);
			continue;
		}

		ASSERT(ipointer_in == B_FALSE);
		ip = ip->i_mnext;

	} while (ip != mp->m_inodes);

	XFS_MOUNT_IUNLOCK(mp);

	ASSERT(ipointer_in == B_FALSE);

	kmem_free(ipointer, sizeof(xfs_iptr_t));
	return XFS_ERROR(last_error);
}

/*
 * xfs sync routine for internal use
 *
 * This routine supports all of the flags defined for the generic VFS_SYNC
 * interface as explained above under xfs_sync.  In the interests of not
 * changing interfaces within the 6.5 family, additional internallly-
 * required functions are specified within a separate xflags parameter,
 * only available by calling this routine.
 *
 */
int
xfs_syncsub(
	xfs_mount_t	*mp,