[LU-3868] lockless IO does not update inode timestamps

• Type: Bug
• Resolution: Fixed
• Priority: Minor
• Fix Version/s: Lustre 2.6.0, Lustre 2.5.2
• Affects Version/s: Lustre 2.5.0
• Labels: osc patch
• Severity: 3
• Assignee: WC Triage
• Reporter: John Hammond

Description

In osc_io_read_start() and osc_io_write_start() if the lockless IO bit is set in the osc_io then the inode timestamps are not updated by the IO. Is there any reason for this?

In osc_io_write_start() we have the following unnecessary attr get before we set the m- and c-times?

                result = cl_object_attr_get(env, obj, attr);
                if (result == 0) {
			attr->cat_mtime = attr->cat_ctime =
                                LTIME_S(CFS_CURRENT_TIME);
                        result = cl_object_attr_set(env, obj, attr,
                                                    CAT_MTIME | CAT_CTIME);
                }

Issue Links

• is related to: LU-3832 osc_io_read_start() ignores O_NOATIME, updates i_atime

Comments

[Peter Jones]

Landed for 2.5.2 and 2.6

[Swapnil Pimpale (Inactive)]

Patch: http://review.whamcloud.com/8797

[Andreas Dilger]

There is no requirement that the server clocks are in sync with the client clocks, so all file timestamps should be generated by the client. In the case of lockless IO, there should be a timestamp update sent by the client with each RPC, regardless of whether it is cached under a DLM lock by the client or not.

[Vitaly Fertman]

timestamps cached on client, as well as other attributes, have no sense without locks, so no need to set it for lockless IO, client just passes attributes to server.
if there is a raced locked IO, its lock will be cancelled by server-side lock taken by lockless IO and next attempt of acquiring a lock will bring proper attributes on client.
btw, it concerns truncate as well.

[John Hammond]

Vladimir and Vitaly, I see that you wrote and inspected this code. Could you explain why we would not update timestamps for lockless IO?