On 04/24/2014 07:46 PM, Linus Torvalds wrote: > On Thu, Apr 24, 2014 at 7:41 PM, Benjamin Herrenschmidt > <benh@xxxxxxxxxxxxxxxxxxx> wrote: >> On Fri, 2014-04-25 at 11:37 +1000, Benjamin Herrenschmidt wrote: >>> >>> The flip side is that we do a lot more IPIs for large invalidates, >>> since we drop the PTL on every page table page. >> >> Oh I missed that your patch was smart enough to only do that in the >> presence of non-anonymous dirty pages. That should take care of the >> common case of short lived programs, those should still fit in a >> single big batch. > > Right. It only causes extra TLB shootdowns for dirty shared mappings. > > Which, let's face it, don't actually ever happen. Using mmap to write > to files is actually very rare, because it really sucks in just about > all possible ways. There are almost no loads where it's not better to > just use a "write()" system call instead. > > So the dirty shared mapping case _exists_, but it's pretty darn > unusual. The case where you have lots of mmap/munmap activity is even > less unusual. I suspect the most common case is for stress-testing the > VM, because nobody sane does it for an actual real load. > The cases where they occur the mappings tend to be highly stable, i.e. map once *specifically* to be able to do a whole bunch of things without system calls, and then unmap when done. -hpa -- To unsubscribe, send a message with 'unsubscribe linux-mm' in the body to majordomo@xxxxxxxxx. For more info on Linux MM, see: http://www.linux-mm.org/ . Don't email: <a href=mailto:"dont@xxxxxxxxx";> email@xxxxxxxxx </a>