David - The key aim here is to avoid having something get merged that everyone misses that sets off a landmine... So. Let me propose a few things, and if none work for you, perhaps you can suggest one? 1. We also maintain mm/mmap.c under 'MEMORY MAPPING' (see patch below). Perhaps we could place these files here, and add you too as a maintainer, if Andrew were open to it? That way you can keep an eye on it? Perhaps others who have relevant knowledge? 2. We add a new PAGE TABLE section, with us and yourself + others as maintainers which reference the files which explicitly manipulate page tables as the core 'magic'? 3. We do the same as 2, but add yourself + any others you suggest as maintainers and us memory mapping/vma guys as reviewers, if you feel that this makes more sense given our relevant areas of expertise? Or whatever you might propose? Or would you prefer things remain as they are? Key aim here is so we can explcitly avoid anything that breaks or blows up memory mapping stuff? Thanks, Lorenzo On Mon, Dec 09, 2024 at 03:56:55PM +0100, David Hildenbrand wrote: > On 09.12.24 15:28, Lorenzo Stoakes wrote: > > On Mon, Dec 09, 2024 at 03:09:26PM +0100, David Hildenbrand wrote: > > > On 09.12.24 11:06, Lorenzo Stoakes wrote: > > > > On Mon, Dec 09, 2024 at 10:16:21AM +0100, David Hildenbrand wrote: > > > > > On 06.12.24 20:16, Lorenzo Stoakes wrote: > > > > > > There are a number of means of interacting with VMA operations within mm, > > > > > > and we have on occasion not been made aware of impactful changes due to > > > > > > these sitting in different files, most recently in [0]. > > > > > > > > > > > > Correct this by bringing all VMA operations under the same section in > > > > > > MAINTAINERS. Additionally take the opportunity to combine MEMORY MAPPING > > > > > > with VMA as there needn't be two entries as they amount to the same thing. > > > > > > > > > > > > [0]:https://lore.kernel.org/linux-mm/CAG48ez0siYGB8GP5+Szgj2ovBZAkL6Zi4n6GUAjzzjFV9LTkRQ@xxxxxxxxxxxxxx/ > > > > > > > > > > > > Signed-off-by: Lorenzo Stoakes <lorenzo.stoakes@xxxxxxxxxx> > > > > > > --- > > > > > > MAINTAINERS | 19 +++++++------------ > > > > > > 1 file changed, 7 insertions(+), 12 deletions(-) > > > > > > > > > > > > diff --git a/MAINTAINERS b/MAINTAINERS > > > > > > index 1e930c7a58b1..95db20c26f5f 100644 > > > > > > --- a/MAINTAINERS > > > > > > +++ b/MAINTAINERS > > > > > > @@ -15060,18 +15060,6 @@ F: tools/mm/ > > > > > > F: tools/testing/selftests/mm/ > > > > > > N: include/linux/page[-_]* > > > > > > > > > > > > -MEMORY MAPPING > > > > > > -M: Andrew Morton <akpm@xxxxxxxxxxxxxxxxxxxx> > > > > > > -M: Liam R. Howlett <Liam.Howlett@xxxxxxxxxx> > > > > > > -M: Lorenzo Stoakes <lorenzo.stoakes@xxxxxxxxxx> > > > > > > -R: Vlastimil Babka <vbabka@xxxxxxx> > > > > > > -R: Jann Horn <jannh@xxxxxxxxxx> > > > > > > -L: linux-mm@xxxxxxxxx > > > > > > -S: Maintained > > > > > > -W: http://www.linux-mm.org > > > > > > -T: git git://git.kernel.org/pub/scm/linux/kernel/git/akpm/mm > > > > > > -F: mm/mmap.c > > > > > > - > > > > > > MEMORY TECHNOLOGY DEVICES (MTD) > > > > > > M: Miquel Raynal <miquel.raynal@xxxxxxxxxxx> > > > > > > M: Richard Weinberger <richard@xxxxxx> > > > > > > @@ -25028,6 +25016,13 @@ L: linux-mm@xxxxxxxxx > > > > > > S: Maintained > > > > > > W: https://www.linux-mm.org > > > > > > T: git git://git.kernel.org/pub/scm/linux/kernel/git/akpm/mm > > > > > > +F: mm/madvise.c > > > > > > +F: mm/mlock.c > > > > > > +F: mm/mmap.c > > > > > > +F: mm/mprotect.c > > > > > > +F: mm/mremap.c > > > > > > +F: mm/mseal.c > > > > > > +F: mm/msync.c > > > > > > > > > > Not sure about mprotect.c, mlock.c and madvise.c, though. I'd claim that the > > > > > real "magic" they perform is in page table handling and not primarily VMA > > > > > handling (yes, both do VMA changes, but they are the "easy" part ;) ). > > > > > > > > And large parts of the VMA logic interface with page tables, see - the entire > > > > 6.12 cycle - around recent changes in mmap() MAP_FIXED - which... the VMA > > > > maintainers fixed :) > > > > > > > > And then there were the issues around VMA and mm locking relating to page > > > > table work which... oh right yeah we had to fix again... :>) > > > > > > > > I mean you can make this argument about probably all of these files (mremap > > > > has -tons- of page table-specific stuff), and then we get back to not being > > > > notified about key changes that interface with memory mapping/VMA which we > > > > end up having to deal with anyway. > > > > > > > > A lot of the reason we have 'magic' in these files anyway is because we > > > > don't have a decent generic page table handler. Not sure I'd actually use > > > > the word 'magic' for that though. > > > > > > > > I am planning to make significant changes to mprotect/mlock soon, which > > > > have some terribly duplicated horrible handling logic, and both are key > > > > considerations in VMA logic as a whole. > > > > > > > > Anyway, as far as I'm concerned page table manipulation after the point of > > > > faulting is completely within the purvue of VMA manipulation and a side > > > > product of it. > > > > > > > > However, can concede mm/madvise.c if you feel strongly about that as that's > > > > a bit blurry, but of course contains a whole bunch of VMA and... page table > > > > manipulation :) I mean it still to me seems very pertinent. > > > > > > > > > > And then we have mprotect.c being heavily used by uffd-wp and NUMA hinting, > > > which don't perform any VMA modification. > > > > > > That's why I don't think the change proposed here is really the right step. > > > > > > > > > > They have much more in common with memory.c, which I wouldn't want to > > > see in > > > > > here either. Hm. > > > > > > > > No, memory.c is really dedicated to fault handling. This is really > > > > different from manipulating page tables in specific cases in my opinion. > > > > > > And fork and such stuff. And if you look into huge_memory.c, we actually > > > moved all of the THP logic for mprotect()/madvise()/... in there. > > > > > > Not sure if something similar should have been done for memory.c, or if the > > > THP stuff should actually also have gone into the respective files. > > > > > > To me it sounds wrong to have VMA maintainers maintain a lot of the code in > > > these files code because these files somehow modify VMAs, sorry. > > > > This isn't what I said, I said that de facto we (that is the MEMORY MAPPING > > maintainers as well as VMA) were dealing with a great many issues around > > page tables and page table manipulation which are rather inseparable from > > one another. > > > > I even went to the lengths of writing a detailed set of documentation on > > locking behaviour in and around page table manipulation and solved > > security-sensitive issues in relation to page table teardown over the 6.12 > > rc cycle. > > > > To me, the idea that mprotect() and mlock(), operations that are explicitly > > about manipulating VMAs (_and of course consqeuent page table > > manipulation_), are somehow separate is really bizarre to me, but I respect > > your opinion even if I disagree. > > > But unfortunately your arguments apply equally as well to mremap.c (more > > than half of which is dedicated to page table manipulation), so I will have > > to drop the whole patch then. > > > > If issues arise there in future, I guess others will have to deal with them > > if we don't notice them (luckily Jann did and pinged this time, hopefully > > will in future). > > Again, the main point I have is that basic VMA handling (the now very nice > vma.c! ) that wouldn't even require page table modifications (the very nice > testing framework you added) is different to users of that functionality. > > And I agree that many VMA modification users imply page table modifications > and more. Yes. > > > > > > > > > -- > > > Cheers, > > > > > > David / dhildenb > > > > > > > To be clear, I made this change in the interests of the community and > > contributing. It seems to me that within mm has far too little sharing of > > the maintainership burden and I only wanted to help with that and make > > explicit what I work on day-to-day. > > And I appreciate that. But putting everything that touches VMAs under VMA is > wrong to me. Well perhaps the issue here is the section then? See the bit at the top. > > > > > I am glad you at least don't object to my doing so with respect to at least > > some parts of the VMA logic. > > I really enjoy how well you separated the core VMA logic from everything > else. > Thanks! Why I admire your work too... :) > -- > Cheers, > > David / dhildenb >
