Skip to contentSkip to conversation

New: xfs: data races on ip->i_itemp->ili_fields in xfs_inode_clean()

Bug 207713 [email protected] 2 participants 5 messages May 13, 2020

lore
Share & tools
lore
#0 lore raw

https://bugzilla.kernel.org/show_bug.cgi?id=207713

            Bug ID: 207713
           Summary: xfs: data races on ip->i_itemp->ili_fields in
                    xfs_inode_clean()
           Product: File System
           Version: 2.5
    Kernel Version: 5.4
          Hardware: All
                OS: Linux
              Tree: Mainline
            Status: NEW
          Severity: normal
          Priority: P1
         Component: XFS
          Assignee: [email protected]
          Reporter: [email protected]
        Regression: No

The function xfs_inode_clean() is concurrently executed with the functions xfs_inode_item_format_data_fork(), xfs_trans_log_inode() and xfs_inode_item_format() at runtime in the following call contexts:

Thread 1:
xfsaild()
xfsaild_push()
xfsaild_push_item()
xfs_inode_item_push()
xfs_iflush()
xfs_iflush_cluster()
xfs_inode_clean()

Thread 2 (case 1):
xfs_file_write_iter()
xfs_file_buffered_aio_write()

    xfs_file_aio_write_checks()
      xfs_vn_update_time()
        xfs_trans_commit()
          __xfs_trans_commit()
            xfs_log_commit_cil()
              xlog_cil_insert_items()
                xlog_cil_insert_format_items()
                  xfs_inode_item_format()
                    xfs_inode_item_format_data_fork()

Thread 2 (case 2):
xfs_file_write_iter()
xfs_file_buffered_aio_write()
xfs_file_aio_write_checks()
xfs_vn_update_time()
xfs_trans_log_inode()

Thread 2 (case 3):
xfs_file_write_iter()
xfs_file_buffered_aio_write()

    xfs_file_aio_write_checks()
      xfs_vn_update_time()
        xfs_trans_commit()
          __xfs_trans_commit()
            xfs_log_commit_cil()
              xlog_cil_insert_items()
                xlog_cil_insert_format_items()
                  xfs_inode_item_format()

In xfs_inode_clean():
return !ip->i_itemp || !(ip->i_itemp->ili_fields & XFS_ILOG_ALL);

In xfs_inode_item_format_data_fork() (case 1):
iip->ili_fields &=

        ~(XFS_ILOG_DDATA | XFS_ILOG_DBROOT | XFS_ILOG_DEV);

In xfs_trans_log_inode() (case 2):
ip->i_itemp->ili_fields |= flags;

In xfs_inode_item_format() (case 3):
iip->ili_fields &=

        ~(XFS_ILOG_ADATA | XFS_ILOG_ABROOT | XFS_ILOG_AEXT);

The variables ip->i_itemp->ili_fields and iip->ili_fields access the same memory, and thus data races can occur.

These data race were found and actually reproduced by our concurrency fuzzer.

I am not sure whether these data races are harmful and how to fix them properly, so I want to listen to your opinions, thanks :)

Signature · 2 lines
You are receiving this mail because:
You are watching the assignee of the bug.
#1 lore raw
Quoted 23 lines · On Wed, May 13, 2020 at 03:02:10AM +0000, [email protected]…

https://bugzilla.kernel.org/show_bug.cgi?id=207713

            Bug ID: 207713
           Summary: xfs: data races on ip->i_itemp->ili_fields in
                    xfs_inode_clean()
           Product: File System
           Version: 2.5
    Kernel Version: 5.4
          Hardware: All
                OS: Linux
              Tree: Mainline
            Status: NEW
          Severity: normal
          Priority: P1
         Component: XFS
          Assignee: [email protected]
          Reporter: [email protected]
        Regression: No

The function xfs_inode_clean() is concurrently executed with the functions xfs_inode_item_format_data_fork(), xfs_trans_log_inode() and xfs_inode_item_format() at runtime in the following call contexts:

Quoted 8 lines

Thread 1:
xfsaild()
xfsaild_push()
xfsaild_push_item()
xfs_inode_item_push()
xfs_iflush()
xfs_iflush_cluster()
xfs_inode_clean()

The code explains:

                /*
                 * Do an un-protected check to see if the inode is dirty and
                 * is a candidate for flushing.  These checks will be repeated
                 * later after the appropriate locks are acquired.
                 */
                if (xfs_inode_clean(cip) && xfs_ipincount(cip) == 0)
                        continue;

Then it is repeated if the racy check indicates the inode is dirty whilst holding the correct inode locks. All three of the "thread 2" cases are modifying the fields with the correct locks held, so there isn't actually a data race bug we care about here.

IOWs, this code is simply avoiding taking locks if we don't need them - it's a pattern we use in numerous places in XFS, so you're going to get lots of false positives from your tool.

Not a bug, please close.

-Dave.

Signature · 2 lines
Dave Chinner
[email protected]
#2 lore raw

[Bug 207713] xfs: data races on ip->i_itemp->ili_fields in xfs_inode_clean()

In reply to [email protected]

https://bugzilla.kernel.org/show_bug.cgi?id=207713

--- Comment #1 from Dave Chinner ([email protected]) ---

Quoted 23 lines · On Wed, May 13, 2020 at 03:02:10AM +0000, [email protected]…

https://bugzilla.kernel.org/show_bug.cgi?id=207713

            Bug ID: 207713
           Summary: xfs: data races on ip->i_itemp->ili_fields in
                    xfs_inode_clean()
           Product: File System
           Version: 2.5
    Kernel Version: 5.4
          Hardware: All
                OS: Linux
              Tree: Mainline
            Status: NEW
          Severity: normal
          Priority: P1
         Component: XFS
          Assignee: [email protected]
          Reporter: [email protected]
        Regression: No

The function xfs_inode_clean() is concurrently executed with the functions xfs_inode_item_format_data_fork(), xfs_trans_log_inode() and xfs_inode_item_format() at runtime in the following call contexts:

Quoted 8 lines

Thread 1:
xfsaild()
xfsaild_push()
xfsaild_push_item()
xfs_inode_item_push()
xfs_iflush()
xfs_iflush_cluster()
xfs_inode_clean()

The code explains:

                /*
                 * Do an un-protected check to see if the inode is dirty and
                 * is a candidate for flushing.  These checks will be repeated
                 * later after the appropriate locks are acquired.
                 */
                if (xfs_inode_clean(cip) && xfs_ipincount(cip) == 0)
                        continue;

Then it is repeated if the racy check indicates the inode is dirty whilst holding the correct inode locks. All three of the "thread 2" cases are modifying the fields with the correct locks held, so there isn't actually a data race bug we care about here.

IOWs, this code is simply avoiding taking locks if we don't need them - it's a pattern we use in numerous places in XFS, so you're going to get lots of false positives from your tool.

Not a bug, please close.

-Dave.

Signature · 2 lines
You are receiving this mail because:
You are watching the assignee of the bug.
#3 lore raw

[Bug 207713] xfs: data races on ip->i_itemp->ili_fields in xfs_inode_clean()

In reply to [email protected]

https://bugzilla.kernel.org/show_bug.cgi?id=207713

Jia-Ju Bai ([email protected]) changed:

What |Removed |Added
----------------------------------------------------------------------------

             Status|NEW                         |RESOLVED
         Resolution|---                         |INVALID

--- Comment #2 from Jia-Ju Bai ([email protected]) --- Okay, thanks for explanation.

Signature · 2 lines
You are receiving this mail because:
You are watching the assignee of the bug.
#4 lore raw

[Bug 207713] xfs: data races on ip->i_itemp->ili_fields in xfs_inode_clean()

In reply to [email protected]

https://bugzilla.kernel.org/show_bug.cgi?id=207713

Eric Sandeen ([email protected]) changed:

What |Removed |Added
----------------------------------------------------------------------------
CC| |[email protected]

--- Comment #3 from Eric Sandeen ([email protected]) ---
A comment to indicate that this is intentional may be helpful, though. :)

Signature · 2 lines
You are receiving this mail because:
You are watching the assignee of the bug.