On Mon, 19 May 2014, Jan Kara wrote: > On Mon 19-05-14 13:44:25, David Herrmann wrote: > > On Thu, May 15, 2014 at 12:35 AM, Hugh Dickins <hu...@google.com> wrote: > > > The aspect which really worries me is this: the maintenance burden. > > > This approach would add some peculiar new code, introducing a rare > > > special case: which we might get right today, but will very easily > > > forget tomorrow when making some other changes to mm. If we compile > > > a list of danger areas in mm, this would surely belong on that list. > > > > I tried doing the page-replacement in the last 4 days, but honestly, > > it's far more complex than I thought. So if no-one more experienced
To be honest, I'm quite glad to hear that: it is still a solution worth considering, but I'd rather continue the search for a better solution. > > with mm/ comes up with a simple implementation, I'll have to delay > > this for some more weeks. > > > > However, I still wonder why we try to fix this as part of this > > patchset. Using FUSE, a DIRECT-IO call can be delayed for an arbitrary > > amount of time. Same is true for network block-devices, NFS, iscsi, > > maybe loop-devices, ... This means, _any_ once mapped page can be > > written to after an arbitrary delay. This can break any feature that > > makes FS objects read-only (remounting read-only, setting S_IMMUTABLE, > > sealing, ..). We need to fix it together with your sealing patchset, because your patchset is all about introducing a new kind of guarantee: a guarantee which this async i/o issue makes impossible to give, as things stand. Exasperating for you, I understand; but that's how it is. A new feature may make new demands on the infrastructure. I can imagine existing problems, but (I may be out of touch) I have not heard of them as problems in practice. Certainly they would not be recent regressions: mm-page versus fs-file has worked in this way for as long as I've known them (pages released independently of unmapping the file, with the understanding that i/o might still be in progress, so care taken not to free the pages too soon). > > > > Shouldn't we try to fix the _cause_ of this? Nobody is against fixing the cause: we are all looking for the simplest way of doing so, > > > > Isn't there a simple way to lock/mark/.. affected vmas in > > get_user_pages(_fast)() and release them once done? We could increase > > i_mmap_writable on all affected address_space and decrease it on > > release. This would at least prevent sealing and could be check on > > other operations, too (like setting S_IMMUTABLE). > > This should be as easy as checking page_mapping(page) != NULL and then > > adjusting ->i_mmap_writable in > > get_writable_user_pages/put_writable_user_pages, right? > Doing this would be quite a bit of work. Currently references returned by > get_user_pages() are page references like any other and thus are released > by put_page() or similar. Now you would make them special and they need > special releasing and there are lots of places in kernel where > get_user_pages() is used that would need changing. Lots of places that would need changing, yes; but we have often wondered in the past whether there should be a put_user_pages(). Though I'm not sure that it would actually solve anything... > > Another aspect is that it could have performance implications - if there > are several processes using get_user_pages[_fast]() on a file, they would > start contending on modifying i_mmap_writeable. Doing extra vma work in get_user_pages() wouldn't be so bad. But doing any vma work in get_user_pages_fast() would upset almost all its users: get_user_pages_fast() is a fast-path which expressly avoids the vmas, and hates additional cachelines being added to its machinations. If sealing had appeared before get_user_pages_fast(), maybe we wouldn't have let get_user_pages_fast() in; but now it's the other way around. I would be more interested in attacking from the get_user_pages() and get_user_pages_fast() end, if I could convince myself that they do actually delimit the problem; maybe they do, but I'm not yet convinced. > > One somewhat crazy idea I have is that maybe we could delay unmapping of a > page if this was last VMA referencing it until all extra page references of > pages in there are dropped. That would make i_mmap_writeable reliable for > you and it would also close those races with remount. Hugh, do you think > this might be viable? It is definitely worth pursuing further, but I'm not very hopeful on it. In a world of free page flags and free struct page fields, maybe. (And I don't see sealing as a feature sensibly restricted to 64-bit only.) I think we would have to set a page flag, maybe bump a count, for every leftover page that raises i_mmap_writable; and lower it (potentially from interrupt context) at put_page() time. Easy to make i_mmap_writable an atomic rather than guarded by i_mmap_mutex, but we still need to synchronize on it falling to 0. And how would we recognize the relevant, decrementing, put_page()? page_count divided into "read_"count and write_count? Ugh! I also have a strong instinct against adding delays into munmap+exit; though that mainly comes from the urge to free memory, and here we are only delaying until a page becomes freeable, so maybe I should abandon that bias in this case. I did start thinking in this direction last week, but stuck somewhere and retreated, I forget on what issue. At this moment I'm not really in that zone, but anxious to complete my promised responses to David's patches, which I almost but not quite completed last night. Hugh -- To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to majord...@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html Please read the FAQ at http://www.tux.org/lkml/