[LU-10704] open(2) with O_CREAT takes 60s (timeout) with virus scanning application

Fri, 17 Aug 2018 22:22:39 +0000

On the client when Lustre creates a file, we create the inode and instantiate the dentry in ll_create_it(), it's still hidden by lld->lld_invalid. Next lookup could discover it, however this brings an issue for some applications. like virus scanning, for example https://www.f-secure.com/en/web/business_global/downloads/linux-security

f-secure does real time scanning using fanotify. When client creates a file using open(2) with O_CREAT, the process will wait in fanotify until f-secure scanning thread does its job:

[<ffffffff8124b785>] fanotify_handle_event+0x1e5/0x330
[<ffffffff81247e15>] fsnotify+0x285/0x510
[<ffffffff812b4936>] security_file_open+0x66/0x70
[<ffffffff81200539>] do_dentry_open+0xb9/0x2e0
[<ffffffff8120078f>] finish_open+0x2f/0x40
[<ffffffffc0a31506>] ll_atomic_open+0x1d6/0x11f0 [lustre]
[<ffffffff812121bd>] do_last+0xa4d/0x12c0
[<ffffffff81212af2>] path_openat+0xc2/0x490
[<ffffffff8121508b>] do_filp_open+0x4b/0xb0
[<ffffffff81201bc3>] do_sys_open+0xf3/0x1f0
[<ffffffff81201cde>] SyS_open+0x1e/0x20
[<ffffffff816b89fd>] system_call_fastpath+0x16/0x1b
[<ffffffffffffffff>] 0xffffffffffffffff

The scanning thread will try to do lookup, the dcache lookup will fail because lld->lld_invalid, so it tries lookup_real, which requires i_mutex holding by the thread creating the file and waiting for the scanning thread.

Eventually scanning thread will be killed by f-secure after the timeout, and file creation could finish. Any thread tries to take i_mutex on the parent will block during the timeout.

Whamcloud Community JIRA

[LU-10704] open(2) with O_CREAT takes 60s (timeout) with virus scanning application