Hello, On Wed, Oct 02, 2019 at 03:28:00PM +0900, Namhyung Kim wrote: > On Sat, Sep 21, 2019 at 6:04 AM Tejun Heo <[email protected]> wrote: > > > > On Fri, Sep 20, 2019 at 05:47:45PM +0900, Namhyung Kim wrote: > > > Thanks for the sharing information! For 32-bit, while the ino itself is > > > not > > > monotonic, gen << 32 + ino is monotonic right? I think we can use the > > > > It's not. gen gets incremented on every allocation, so it has not > > high but still realistic chance of collisions. > > In __kernfs_new_node(), gen gets increased only if idr_alloc_cyclic() > returns lower than the cursor... I'm not sure you talked about it.
Ah, I forgot that it's using cyclic idr, so yeah, it's not as bad in terms of recycling although cyclic allocation on idr is pretty inefficient. I still think it'd be better to switch to rbtree and so that 64bit can simply use monotonically increasing numbers but that definitely isn't a must and we can juse continue with the current allocation method. Thanks. -- tejun
