[ moving to tech-kern ] hi,
> y...@mwd.biglobe.ne.jp (YAMAMOTO Takashi) wrote: >> > >> > Here is the updated patch after your changes: >> > >> > http://www.netbsd.org/~rmind/uvm_anon_freelst2.diff >> > >> > As you noted, uvm_anfree() can temporarily release the amap lock - that >> > can happen in amap_copy(). Patch closes the race by moving uvm_anfree >> > () further, and changes the semantics of the function, now called >> > uvm_anon_freelst(), to return with amap lock released (plus free anons >> > without lock held). >> >> the temporary release of the amap lock is only for O->A loan >> which you disabled, isn't it? > > Right, uvm_anon_locklaonpg() dance can happen only in O->A case. However, > having uvm_anfree() able to release the lock by its interface definition > is potentially defective. It is the main motivation why I want to slightly > rework the code into uvm_anon_freelst() which would always drop the lock > and move freeing of anons to the end point. Cleaner, less error prone. the committed change seems broken in case uvm_anon_dispose sets PG_RELEASED. in that case, uvm_anon_freelst should leave the anon as it will be freed by uvm_anon_release later. YAMAMOTO Takashi > > -- > Mindaugas
