Date: Mon, 26 Jan 2015 21:01:58 +0100 From: Manuel Bouyer <[email protected]>
On Mon, Jan 26, 2015 at 07:33:18PM +0000, Taylor R Campbell wrote: > We should probably do the same thing in genfs_do_io and anywhere else > that uses nestiobuf for writes, and invent a better way to express the > nestiobuf API so it's less ridiculously error-prone. Yes, I'll look at this (but I don't know why it does +2 here). genfs_do_io does +2 because it needs to count the nestiobuf master biodone and the caller-supplied (*iodone)(mbp) callback, which calls biodone or vwakeup (see uvm_aio_biodone or genfs_dio_iodone). Counting the nestiobuf master biodone can be done like we're doing with vnd. I would guess (and hope) that counting the caller-supplied iodone callback can be done after biowait and before (*iodone)(mbp). But I guess genfs_do_io() is safe as is: it's probably called with the vnode locked, as VOP_BMAP() is called without taking and releasing the lock. Maybe... I think there's a great deal too much blocking (i.e., any) while the vnode lock is held, but I suppose cleaning that up can be left for another day.
