On Tue, Jan 07, 2020 at 10:09:44AM +0800, yukuai (C) wrote: > > > On 2020/1/7 5:57, Darrick J. Wong wrote: > > So your solution is to split the delalloc reservation to constrain the > > allocation to the range that's being operated on? > Yes, I'm trying to split delalloc reservation. > > > > If so, I think a better solution (at least from the perspective of > > reducing fragmentation) would be to map the extent unwritten and force a > > post-writeback conversion[1] but I got shot down for performance reasons > > the last time I suggested that. > I'm wondering if spliting delalloc reservation have the same performance > issue. Probably not, but OTOH will doing so (a) have a noticeable effect on fragmentation, and (b) will the kernel code choke on two adjacent delalloc extents that could be merged? (I think the answer is no since we'll either complete the operation or the fs goes down...?) --D > Thanks! > Yu Kuai >
