Opened 5 days ago

Closed 5 days ago

#441 closed defect (fixed)

Long running IMFT requests get aborted with no ability to extend timeout

Reported by: Eric Swenson Owned by: Eric Swenson
Priority: minor Milestone: MR12.9
Component: Networking Version: MR12.5
Keywords: Cc:

Description

An IMFT output driver sends an object by running backup_dump into the transport, then sends an EOF command and waits for the remote input driver’s EOF reply, which says whether the object was loaded without errors (imft_transmit object_). The remote driver sends that reply only after backup load has loaded everything. The wait is fixed at 5 minutes: the EOF command is written with local icri.timeout = 5 * ONE MINUTE and the same value is used for the reply read.

So the sender’s 5 minutes must cover the time the receiver needs to take in and load whatever is still in transit when the sender finishes, plus backup_load ’s own time for the last entries. When that is longer, the sender reports

imft_transmit_object_: Attempt to read/write record timed out. Attempting to read from SYSTEM.

drops the connection, and defers the request, and the request fails the same way on every retry.

This was seen transferring a large subtree over dial-up (FNP) lines between two emulated systems, where the emulator let the sender run several minutes of data ahead of the receiver. (That emulator
problem has been fixed separately.) A slow line, a heavily loaded receiver or a large final segment can still need more than 5 minutes, and a site has no way to allow for it.

Change History (1)

comment:1 by Eric Swenson, 5 days ago

Resolution: → fixed
Status: new → closed
Note: See TracTickets for help on using tickets.