On Fri, Nov 17, 2006 at 12:40:53AM +0100, Pavel Machek wrote: > On Fri 2006-11-17 10:23:49, David Chinner wrote: > > On Sun, Nov 12, 2006 at 11:43:05PM +0100, Rafael J. Wysocki wrote: > > > On Sunday, 12 November 2006 23:30, David Chinner wrote: > > > > And how does freezing them at that point in time guarantee consistent > > > > filesystem state? > > > > > > If the work queues are frozen, there won't be any fs-related activity _after_ > > > we create the suspend image. > > > > fs-related activity before or after the suspend image is captured is > > not a problem - it's fs-related activity _during_ the suspend that > > is an issue here. If we have async I/O completing during the suspend > > image capture, we've got problems.... > > fs-related activity _after_ image is captured definitely is a problem > -- it breaks swsusp invariants. > > During image capture, any fs-related activity is not possible, as we > are running interrupts disabled, DMA disabled. Ok, so the I/o that finishes during the image capture won't be reported until after the capture completes. that means we lose the capability to even run the I/O completion on those buffers once we get to resume. They've already been issued, so will be locked and never issued again because the suspend image says they've been issued, but they'll never complete on resume because their completion status was updated after the capture was taken and will never be recreated after resume. IOWs, the fileystem will start to hang on the next reference to any of these locked buffers that the filesystem thinks is still under I/O.... Freezing the workqueues doesn't fix this..... Cheers, Dave. -- Dave Chinner Principal Engineer SGI Australian Software Group -- dm-devel mailing list dm-devel@xxxxxxxxxx https://www.redhat.com/mailman/listinfo/dm-devel
