On Wed, May 11, 2022 at 05:49:49PM -0700, Paul E. McKenney wrote: > On Wed, May 11, 2022 at 05:34:52PM -0700, John Hubbard wrote: > > On 5/11/22 17:26, Minchan Kim wrote: > > > > > Let me try to say this more clearly: I don't think that the following > > > > > __READ_ONCE() statement can actually help anything, given that > > > > > get_pageblock_migratetype() is non-inlined: > > > > > > > > > > + int __mt = get_pageblock_migratetype(page); > > > > > + int mt = __READ_ONCE(__mt); > > > > > + > > > > > + if (mt & (MIGRATE_CMA | MIGRATE_ISOLATE)) > > > > > + return false; > > > > > > > > > > > > > > > Am I missing anything here? > > > > > > > > In the absence of future aggression from link-time optimizations (LTO), > > > > you are missing nothing. > > > > > > A thing I want to note is Android kernel uses LTO full mode. > > > > Thanks Paul for explaining the state of things. > > > > Minchan, how about something like very close to your original draft, > > then, but with a little note, and the "&" as well: > > > > int __mt = get_pageblock_migratetype(page); > > > > /* > > * Defend against future compiler LTO features, or code refactoring > > * that inlines the above function, by forcing a single read. Because, this > > * routine races with set_pageblock_migratetype(), and we want to avoid > > * reading zero, when actually one or the other flags was set. > > */ > > int mt = __READ_ONCE(__mt); > > > > if (mt & (MIGRATE_CMA | MIGRATE_ISOLATE)) > > return false; > > > > > > ...which should make everyone comfortable and protected from the > > future sins of the compiler and linker teams? :) > > This would work, but it would force a store to the stack and an immediate > reload. Which might be OK on this code path. > > But using READ_ONCE() in (I think?) __get_pfnblock_flags_mask() > would likely generate the same code that is produced today. > > word = READ_ONCE(bitmap[word_bitidx]); > > But I could easily have missed a turn in that cascade of functions. ;-) > > Or there might be some code path that really hates a READ_ONCE() in > that place. My worry about chaning __get_pfnblock_flags_mask is it's called multiple hot places in mm codes so I didn't want to add overhead to them.
