On Wed, 11 Jul 2007, Christoph Hellwig wrote: > On Wed, Jul 11, 2007 at 02:12:45PM +0400, Dmitry Monakhov wrote: > > > > Or would this simpler patch be the right one? I suspect the > > > > mspec driver only says VM_LOCKED because of a deep-seated but > > > > irrational fear that its pages might fall into reclaim.
I was perhaps too unkind: that fear was not irrational in 2.4, when reclaim scanned vmas; but in 2.6, pages have to be on one of reclaim's LRUs to fall vulnerable to it. > > > > > > Looks good. We probably should add a debug check to do_mmap_pgoff > > > so that ->mmap methods don't change flags that are not for drivers. We could indeed, though I'd rather not jump into that: something to do when we tidy up those driver mmaps (something I promised to do 18 months ago?), there's a lot of pointless flag setting. > > As result fglrx totally goes crazy, because it change vm_flags > > even from ->nopage() calback :) That must be an exciting new State of the Art version of fglrx. Looking at what I downloaded for inspection a year ago, I can't see any sign of that. But I do see lots of random stabs at setting different vm_flags in different mmap cases (VM_IO, VM_SHM, VM_RESERVED, VM_LOCKED): of which the VM_IO serves some point, VM_SHM is defunct, VM_RESERVED will be defunct, and VM_LOCKED can confuse us. But interestingly, where they set VM_LOCKED, they did increment vm_locked: so there shouldn't be that issue of wrapping negative on munmap, which had worried me. And now I look at Dmitry's patch again, I see that he was indeed assuming that the driver had done that incrementation. > Well, everyone with half a brain knows that fglrx is not just legally > problematic but an utter piece of junk. We should add more debug checks > to stop it from doing such stupid things. And yes, chaning flags from > ->nopage() does not just deserve a warning but a panic. Well, they'll get what they deserve. I'm not convinced Dmitry's right about their nopage; and I'm not going to waste any more time working out ways to protect ourselves from them. Let's stick with the mspec. Hugh - To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to [EMAIL PROTECTED] More majordomo info at http://vger.kernel.org/majordomo-info.html Please read the FAQ at http://www.tux.org/lkml/
