On Fri, Aug 06, 2021 at 07:44:57PM +0200, Peter Zijlstra wrote:
On Fri, Aug 06, 2021 at 05:22:08PM +0200, Mete Polat wrote:Commit e977145aeaad ("[RBTREE] Add explicit alignment to sizeof(long) for struct rb_node.") adds an explicit alignment to the struct rb_node due to some speciality of the CRIS architecture. The support for the CRIS architecture was removed with commit c690eddc2f3b ("CRIS: Drop support for the CRIS port") So, remove this now unneeded explicit alignment in struct rb_node as well. This basically reverts commit e977145aeaad ("[RBTREE] Add explicit alignment to sizeof(long) for struct rb_node."). The rbtree node color is stored in the LSB of '__rb_parent_color'. Only mask the first bit in '__rb_parent()', otherwise it modifies the node's parent address on m68k.I still don't believe for a second this will actually work. We rely on rcu_assign_pointer() and rcu_dereference() to work on the rb_{left,right} members, and I don't think any architecture can provide single copy atomic loads and stores that are not naturally aligned (eg. when they straddle a cache or page boundary).
I guess I am misunderstanding something here, but isn't that then a problem that all rcu pointers in any struct would face, independent of an 'aligned' struct attribute? As long as allocators do not place a small struct as rb_node over page boundaries and the rcu pointers itself are aligned we should be fine, aren't we? I am not sure if any of the SL*B allocators is doing that though.
