[LU-15499] WBC: deadlock for buffered write in parallel when test with dcp/ior

• Status: Open
• Project: Lustre
• Type: Technical task
• Priority: Minor
• Reporter: Qian Yingjin
• Assignee: Qian Yingjin
• Resolution: Unresolved

Description

When benchmark with IOR, we found the following deadlock:

memfs_file_write_iter()
{
    ....
    down_read(&wbci->wbci_rw_sem); // lock 1
     if (wbc_inode_data_caching(wbci)) {
                inode_lock(inode);     // lock 2
                rc = generic_write_checks(iocb, iter);
                if (rc > 0)
                        rc = __generic_file_write_iter(iocb, iter);
                inode_unlock(inode);
                 ....
    }
    up_read(&wbci->wbci_rw_sem);
    ....
}

memfs_write_begin(...)
{
    rc = memfs_write_getpage(inode, index, pagep);
        if (rc == -ENOSPC) {
                int rc2;

                // need to call inode_unlock(inode);
                up_read(&wbci->wbci_rw_sem);
                rc2 = wbc_make_data_commit(file->f_path.dentry);
                down_read(&wbci->wbci_rw_sem);
                // need to call inode_lock(inode);
                if (rc2 < 0)
                        rc = rc2;
        }

Comments

[Gerrit Updater]

"Yingjin Qian <[e-mail]>" uploaded a new patch: https://review.whamcloud.com/46376
Subject: LU-15499 wbc: avoid deadlock for parallel buffered write
Project: fs/lustre-release
Branch: master
Current Patch Set: 1
Commit: 20692edcecfaa9244451592b73c96c78dbbd584e