[LU-12977] fix i_mutex for ldiskfs_truncate() in osd_execute_truncate()

• Type: Improvement
• Resolution: Fixed
• Priority: Minor
• Fix Version/s: Lustre 2.14.0
• Assignee: James A Simmons
• Reporter: Andreas Dilger

Description

The inode->i_mutex should be held when calling osd_execute_truncate->ldiskfs_truncate(). However, due to locking order in osd-ldiskfs we cannot just grab i_mutex around ldiskfs_truncate() as that can lead to deadlocks as described in LU-6446 and LU-4252.

Currently, there is an ldiskfs patch to remove the WARN_ON() in ext4_truncate() when it is called without i_mutex held. This is safe because the osd-ldiskfs code holds its own lock (oo_ext_idx_sem) on the object over the ldiskfs_truncate() call, but potentially there is a way to improve this without patching the ldiskfs code.

The truncate-warning patch was first added with:

 commit 3daba43a191a0a2c9358d39b3f48b1b759a41819
 Author:     Yang Sheng <[e-mail]>

    LU-6446 ldiskfs: remove WARN_ON from ldiskfs_orphan_add{del}
    
    RHEL7.1 kernel carefully check i_mutex whether locked in
    many places. It will be triggered while ldiskfs_truncate()
    was invoked in osd layer. A dead lock would occured if we
    just locked i_mutex around it since Lustre using a reverse
    order for start journal==>lock i_mutex. Consider Lustre has
    own ldlm lock. So just remove such messages in ldiskfs patches.

    Reviewed-on: http://review.whamcloud.com/14690

so I think moving the lock is not as easily done as what is done in the current patch, see comments in LU-6446 and LU-4252 for details. The code used to look like the following:

       if (!(inode->i_state & (I_NEW|I_FREEING)))
               mutex_lock(&inode->i_mutex);
       ldiskfs_truncate(inode);
       if (!(inode->i_state & (I_NEW|I_FREEING)))
               mutex_unlock(&inode->i_mutex);

but this was removed as part of the patch:

 commit b98792ae96ad2bb3c73ca3ebebdd8edfdfbc35df
 Author:     yangsheng <[e-mail]>

    LU-4252 osd: remove locking i_mutex in osd_punch
    
    This piece of code intent to calm down the kernel
    WARN_ON messsage show up since 3.9 upstream. So we
    can just remove it for now. We'll add a ldiskfs
    patch to resolve this issue while 3.9 server support
    come up.

    Reviewed-on: http://review.whamcloud.com/9644

In more recent developments, LU-10048 changed the truncate code so that it is called asynchronously from the main transaction, so that it doesn't bloat transaction sizes, and avoids lock ordering issues:

commit cf29a5e7bfa254ccfcea023028fe7da80503c512
Author:     Alex Zhuravlev <[e-mail]>

    LU-10048 osd: async truncate
    
    osd-ldiskfs should execute truncate outside of main transaction
    handle. This avoids restarting truncate transaction handles in
    main transaction, and allows "transaction first, locking second"
    model on OST.
    
    Reviewed-on: https://review.whamcloud.com/27488

It may now be OK to grab i_mutex around ldiskfs_truncate() in osd_execute_truncate() and remove ext4-remove-truncate-warning.patch, but this should be reviewed and tested properly (e.g. with the racer workload that triggered LU-4252 originally).

Issue Links

• is related to: LU-13234 ldiskfs/namei.c:3310 ldiskfs_orphan_add+0x11e/0x2a0 [ldiskfs]
• is related to: LU-4252 Failure on test suite racer test_1
• is related to: LU-6446 Warn-on in ldiskfs_orphan_add/del
• is related to: LU-10048 osd-ldiskfs to truncate outside of main transaction
• is related to: LU-11832 ARM servers crashing on MDS startup
• is related to: LU-13183 Linux 5.4 ldiskfs build (drop: ext4-remove-truncate-warning.patch)