On 02/17/2014 08:43 PM, Ben Hutchings wrote:
On Sat, 2014-02-15 at 14:38 -0500, Peter Hurley wrote:Since commit a2c1c57be8d9fd5b716113c8991d3d702eeacf77, workqueue: consider work function when searching for busy work items work items whose work functions are re-assigned are no longer guaranteed non-reentrancy with the previously assigned work function. For example, PREPARE_WORK(&work, funcA) schedule_work(&work) . < funcA starts > . PREPARE_WORK(&work, funcB) schedule_work(&work) funcA() may run concurrently with funcB(). The work item non-reentrancy guarantee is a crucial design guarantee which, if violated, may not have obvious consequences. For example, the entire firewire subsystem expects the as-documented per-work item non-reentrancy guarantee, which was the behavior prior to the above commit, not the per-work function + per-work item behavior since. Document the known exceptions to this guarantee.[...] It never would have occurred to me that you could safely change the function for a work item that is already scheduled or running. Especially given that PREPARE_WORK() is just a simple assignment (i.e. no serialisation).
process_one_work() has an established order that safely allows for resetting the work function and scheduling the work, and further guaranteeing that the new work function will run. Further, existing memory barriers ensure that 1. The new work function is visible on all cpus before testing if the work is already pending. 2. The new work function is stored as the worker's current function before the work is marked as not pending. If this wasn't possible, then single-threaded workqueues could not be used for multiple functions without flushing work. I wonder if the floppy driver is broken too. Regards, Peter Hurley -- To unsubscribe from this list: send the line "unsubscribe stable" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html