On Thu, Jan 16, 2025 at 6:04 PM Tavian Barnes <tavianator@xxxxxxxxxxxxxx> wrote: > > On Thu, Jan 16, 2025 at 01:04:42PM +0100, Mateusz Guzik wrote: > > In this context there are next to no pages found, but there is a > > gazillion relocks as the entire VA is being walked. > > Do I understand correctly that all the relocks are to look up the VMA > associated with each address, one page at a time? That's especially > wasteful as dump_user_range() is called separately for each VMA, so it's > going to find the same VMA every time anyway. > it indeed is a loop over vmas, and then over the entire range with PAGE_SIZE'd steps > > I however vote for someone mm-savvy to point out an easy way (if any)> > to just iterate pages which are there instead. > > It seems like some of the <linux/pagewalk.h> APIs might be relevant? > Not sure which one has the right semantics. Can we just use > folio_walk_start()? > > I guess the main complexity is every time we find a page, we have to > stop the walk, unlock mmap_sem, call dump_emit_page(), and restart the > walk from the next address. Maybe an mm expert can weigh in. > I don't know the way, based on my epsilon understanding of the area I *suspect* walking the maple tree would do it. -- Mateusz Guzik <mjguzik gmail.com>
