Change Details

The drawback is that it increases struct vm_page (its md part) by 4 (actually 8) bytes. It is practically impossible to generate conjection for this stuff using buildworld. I reduced the number of slots to two and I only get ``` vm.pmap.invl_busy_next_gen: 0 vm.pmap.invl_gen_wrapped: 0 vm.pmap.invl_wait: 2 vm.pmap.invl_busy_slots: 1465 ```

The basic observation is that the delayed invalidation is page-local operation. If page has the generation assigned, we only need to wait until the owner of the generation finishes. The drawback is that it increases struct vm_page (its md part) by 4 (actually 8) bytes. It is practically impossible to generate conjection for this stuff using buildworld. I reduced the number of slots to two and I only get ``` vm.pmap.invl_busy_next_gen: 0 vm.pmap.invl_gen_wrapped: 0 vm.pmap.invl_wait: 2 vm.pmap.invl_busy_slots: 1465 ```

The basic observation is that the delayed invalidation is page-local operation. If page has the generation assigned, we only need to wait until the owner of the generation finishes. The drawback is that it increases struct vm_page (its md part) by 4 (actually 8) bytes. It is practically impossible to generate conjection for this stuff using buildworld. I reduced the number of slots to two and I only get ``` vm.pmap.invl_busy_next_gen: 0 vm.pmap.invl_gen_wrapped: 0 vm.pmap.invl_wait: 2 vm.pmap.invl_busy_slots: 1465 ```